7 edition
th
of the international
scientific conference
ForestSAT 2016
15th to 18th
of November 2016
Santiago, Chile
Universidad Mayor
Título: Proceedings Book 7th edition of the International Scientific Conference ForestSAT 2016
Editores: Claudio Muñoz Riveros,
Paulina Javiera Vidal Páez,
Waldo Antonio Pérez Martínez,
Pablo Christian Cruz Johnson,
Markus Frederick Keusch,
Jesús Torralba Pérez.
ISBN: 978-956-7459-49-0
Edita: Universidad Mayor
Printed in Chile
Imprime: CUPOLICAN Servicios Gráficos - Fono 226716467
Dieciocho Nº 786, Santiago, Chile
ventas@caupolican.cl
Reservados todos los derechos. Ni la totalidad ni parte de este libro puede reproducirse o transmitirse
por ningún procedimiento electrónico o mecánico, incluyendo fotocopia, grabacion magnética
o cualquier almacenamiento de información o sistema de reproducción, sin permiso previo y por
escrito de los titulares del Copyright.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Hosts
Sponsors
3
Universidad Mayor
Organizing Committee
Claudio Muñoz
Paulina Vidal
OTERRA, Universidad Mayor, Chile.
OTERRA, Universidad Mayor, Chile.
Pablo Cruz
Jesús Torralba
OTERRA, Universidad Mayor, Chile.
OTERRA, Universidad Mayor, Chile.
Waldo Perez
Markus Frederick Keusch
OTERRA, Universidad Mayor, Chile.
OTERRA, Universidad Mayor, Chile.
Steering Committee
David Miranda,
Pablo Cruz,
Universidad Santiago Compostela, Spain.
OTERRA, Universidad Mayor, Chile.
Gherardo Chirici,
Ronald McRobert,
Universidad de Firenze, Italy.
USDA Forest Service, USA.
Hakan Olsson,
Ross Hill,
Swedish University of Agricultural Sciences, Sweden.
Bournemouth University, UK.
Javier Cano,
Tatjana Koukal,
Corporación Nacional Forestal, Chile.
Juan Suarez,
BOKU, Austria.
Warren Cohen,
Forest Research, UK.
USDA Forest Service, USA.
Maureen Duane,
Oregon State, USA.
Scientific Committee
Alfonso Condal,
Hubert Hasenauer,
Laszlo Pancel,
Michael Koehl,
Laval University, Québec, Canada.
BOKU, Austria.
GIZ, Germany.
World Forest Institute, Germany.
Alvaro Gutiérrez,
Hugo Rivera,
Lin Cao,
Nicholas C. Coops,
Universidad de Chile, Chile.
Corporación Nacional Forestal, Chile.
Nanjing Forestry University, China.
Forestry UBC, Canada.
Angela de Santis,
Jaime Hernandez,
Lori D. Daniels,
Patricio Acevedo,
Centro Científico Regional Fundación
CEQUA, Chile.
Universidad de Chile, Chile.
Forestry UBC, Canada.
Universidad de la Frontera, Chile.
Annemarie Bastrup-Birk,
Joao Roberto dos Santos,
Lucio Castro,
Richard Fournier,
INPE, Brazil.
Embrapa, Brazil.
Université de Sherbrooke, Canada.
European Environmental Agency,
Denmark.
Juan Carlos Jiménez-Muñoz, Luis Morales,
Ross Hill,
UVEG, Spain.
Universidad de Chile, Chile.
Bournemouth University, UK.
Carlos Cárdenas,
Juan de la Riva,
Marcelo Miranda,
Ruben Valbuena,
Universidad de Zaragoza, Spain.
Universidad Católica, Chile.
University of Eastern Finland, Finland.
Juan Pablo Flores,
Marcelo Scavuzzo,
Tatjana Koukal,
CIREN, Chile.
Comisión Nacional de Actividades
Espaciales CONAE, Argentina.
BOKU, Austria.
Forest Research, UK.
Maria Augusta Doetzer,
Universidad Austral, Chile.
Lars T. Waser,
Embrapa, Brazil.
Swiss Federal Research Institute
WSL, Switzerland.
CSIC, Spain.
Universidad de Magallanes, Chile.
Cristian Mattar,
Universidad de Chile, Chile.
Damiano Gianelle,
Fondazione Edmund Mach, Italy.
Daniel McInerney,
Coillte Teoranta (State Forest Board),
Ireland.
Esteban Vojkovic,
CIREN, Chile.
4
Juan Suarez,
Maria Pilar Martin,
Victor Sandoval,
Warren Cohen,
USDA Forest Service, USA.
Yony Ormazábal,
Universidad de Talca, Chile.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Presentation
Dear all,
We are pleased to introduce you the scientific programme of the 7th edition of ForestSAT. The
international conference of Spatial Application Tools in Forestry is the continuation of an idea that
started in back in 2002. An idea that has been growing progressively into an international movement
of foresters, scientists and developers around the use of Earth Observation in forestry. This year,
the organisation of this conference is happening outside the traditional venues in Europe or North
America. Chile, one of the countries with important forest resources, some of them unique and well
protected in a long list of National Parks and Reserves of the Biosphere, is our impressive venue
this year. With just 18 million population, the 796,000 km2 has 23% forested, where 19% are native
forests supporting high endemism of flora and fauna. This conference also happens at the time the
Chilean parliament approved their National Strategy for Climate Change and Vegetation Resources
for 2017-2025. This is an important step forward in the rationalization and protection of the unique
and sensitive environmental resources of this country. The strategy, coordinated by the Chilean
Forest Service, CONAF, contemplates the use of Earth Observation and others sources of spatial
information as fundamental tools for the implementation and management of policies.
As always, the scientific team has made a selection of topics based on the global interest of our
community that focused on Forest Monitoring, REDD+ and FLEGT, the usual Forest Inventory and
Mapping, Forest Management and Development Methods. This year, the organisation wanted
to introduce specific subjects of interest for the Latin American community with some of the
presentations in Spanish to allow other scientists to break language barriers and contribute with their
unique experiences.
Three keynote speakers where selected because of the relevance of their work for the international
community.
Dr Warren Cohen, from the US Forest Service based in Oregon, is a well know scientist that has been
working for many years in the development of tools to monitor the trajectories of change in the forest.
His work has created important applications for the US Forest Service to map changes, the impact
of disturbances and the recovery of the vegetation. In a time where the EU has launched the new
Sentinel-2 sensors that join the Landsat series of satellites, his tools are one of the most important for
a continuous and complete monitoring of the forested parts of this planet.
Professor Pang Yong, from the Chinese Academy of Forestry, is presenting all the interesting work
that they have been developing over many years in partnership with the Chinese Academy of
Sciences and their own Space programme. He will be talking about the, mostly unknown outside
China, Gaofen sensor and the way is being used by the forest service to monitor their constantly
expanding forest resources. Also, he will be talking about the integration of LiDAR, Hyperspectral
and video sensors under the LiCHy programme. This is a highly portable group of sensors that aims
5
Universidad Mayor
to explore the strengths of each one for a comprehensive monitoring of the forest resources from an
airborne platform.
Dr. María Pilar Martin, Spanish National Research Council, will be talking about her experience in the
monitoring of grasslands and forest fires in the Mediterranean using spectro-radiometric sensors.
She has an extensive career as a researcher that has participated in many national and international
projects with a long list of peer-reviewed papers. She will be talking about the techniques she has
been pioneering over these years and provide an overview of what is going to be the future in the use
of spectral analysis.
Finally, the organisation of the conference will like to pay tribute to Dr. Thomas Hilker, sadly missed
a few months before. At the age of just 40 years old, Thomas has left us full of sorrow but with an
important legacy of scientific work in his prolific career that covered almost every technique from
LiDAR to Satellite imagery, hyperspectral analysis, modelling, data mining, multi-angle analysis,
data fusion and even as an editor of Remote Sensing of the Environment. He was a brilliant scientist
and our friend. Professor Nicholas Coops will present a summary of his valuable contribution to our
science in a special session.
We all hope that the list of oral presentations, posters and training courses become an enjoyable
experience for you.
Best wishes,
Juan Suárez and Claudio Muñoz, conference conveners.
6
Universidad Mayor
Index
ForestSAT 2016 Abstracts Summary
EXPOSITOR
POSTER
Development of Methods .............................................................................................. 13
A novel sampling strategy for applying ALS-based plot imputation to Australian native eucalypt
forests ...................................................................................................................................................15
Airborne Dual Band Radar for Rural Cadastre............................................................................................. 19
An approximation for non-parametric bootstrapping to assess uncertainties in imputation
mapping of forest vegetation ............................................................................................................... 22
Analysis of land cover classes separability using optical and SAR data in areas of high cloud
content in northern South America ........................................................................................................23
Analyzing some factors affecting the extraction of full-waveform LiDAR metrics and their effect
in forest structure variable estimates ................................................................................................... 24
Analyzing the Lorenz Curve of Tree Growth Dominance with Multi-temporal Airborne Lidar
in Wytham Forest (UK) ..........................................................................................................................27
Change detection techniques for development of compatible height growth and site index
models using repeated Airborne Laser Scanning Data .......................................................................... 29
Combining remote sensing and dendrochronology to assess the effect of groundwater
extraction on Prosopis tamarugo: from tree to aquifer level ..................................................................31
Comparison of forest canopy point-measurement methods from above and below....................................32
Generation of a Landsat image mosaic over the Amazon rainforest ............................................................33
Detection of Dead Standing Eucalyptus for managing biodiversity using full-waveform LiDAR data ......... 34
Detection of dead trees by machine learning techniques in temperate forests of the
Bavarian Forest National Park .............................................................................................................. 36
Development of an energy balance model for estimating canopy stomatal conductance
from airborne thermal data ...................................................................................................................37
Development of precision forestry applications for New Zealand plantations using
remotely sensed datasets ..................................................................................................................... 39
Estimating the regional resource supply of forests in south-west Germany for a future
lignocellulose-based bioeconomy using airborne LiDAR, Landsat 7 and National Forest
Inventory data ...................................................................................................................................... 40
Estimation of canopy cover in hemi-boreal broad-leaved forests in Estonia using
hemispherical photography and lidar data ............................................................................................ 41
Forest-Observation-System.net – towards global reference database for forest biomass .......................... 42
Fractal Volumetric Bouligand-Minkowski Classification of Forest Trees ...................................................... 43
Full waveform LiDAR and the new PulseWaves format ..............................................................................47
Isolation of obscured forest tree stems using TLS data .............................................................................. 48
LiDAR- and SAR-based mapping of structural attributes of deciduous savannahs and woodlands
in the southern African region................................................................................................................51
Linking remotely sensed functional diversity with phylogenetic structure of a temperate forest................ 52
Local pivotal method sampling design combined with micro stands utilizing airborne laser
scanning data in a long term forest management planning setting ........................................................55
Machine learning regression algorithms for biophysical parameter retrieval from spectral
properties to detect different levels of ash vitality in Central Europe ..................................................... 56
Mapping the 3D structure of a tropical rainforest using terrestrial laser scanning – a quality
assessment........................................................................................................................................... 58
Measuring stem diameters - a comparison of three methods ..................................................................... 60
Multi-Year Comparison of Tree Species Discrimination from Formosat-2 Satellite Image Time Series ........ 61
Multiscale forest health mapping: the potential of air- and space-borne sensors........................................ 62
Near Real-Time Detection of Forest Changes using Google Earth Engine® and Sentinel-2 Imagery:
A Case Study in Curacutin region, Chile ................................................................................................. 65
7
Universidad Mayor
Optimization of dynamic a global vegetation model at a land cover remote sensing data
for better representation of Russian forests .......................................................................................... 66
Predicting Single Tree Species Diameter Distribution by Airborne Laser Scanning Using
Different Modeling Alternatives.............................................................................................................67
Quality control and quality assessment of LiDAR data ............................................................................... 68
Shadow compensation for imaging spectroscopy data using a radiative transfer approach ........................70
Simulating the spectral response of tropical tree species with 3-D radiative transfer modeling ...................72
Single tree detection with weak canopy shape constraints .........................................................................77
Study on Removal of Atmospheric Effect on Normalized Difference Vegetation Index ................................78
Synergism of SAR and optical data for land use mapping in the Amazonia transition landscape .................79
Tandem-L Global Observation of forests with Two L-Band SAR Satellites - Tandem-L ................................ 82
Temporal Albedo Dynamics in Boreal Forest Fire Scars Using Higher Resolution Albedo
Products from Landsat and Sentinel 2A ................................................................................................ 83
Terrestrial LiDAR and 3D Reconstruction Models for Estimation of Large Individual Tree
Biomass in Tropics ................................................................................................................................ 85
The Global Ecosystems Dynamics Investigation: Current Status ................................................................ 86
The Potential of Multitemporal and Polarimetric Features Derived From Sentinel-1
Data for Forest Parameters Retrieval .....................................................................................................87
Towards an “all-in-one sensor” for forestry applications – estimating forest density, species
composition and biomass from stereo WorldView-2 data ..................................................................... 88
Tropical forest degradation monitoring: Radiometric issues of using both Landsat 8 and
Sentinel 2 in one time series ................................................................................................................. 90
Tropical forest height and structure estimation from AfriSAR campaign PolInSAR data in Gabon .............. 95
Two Phase Assessment System for the Effective Monitoring of Tropical Forests .........................................97
UAV-Borne and Airborne Remote Sensing for Tree Disease Symptom Detection ....................................... 99
Updating of lidar based forest attribute maps using digital photogrammetry combined with
the lidar data .......................................................................................................................................100
Use of Hybrid Model-Based Inference with a Sample of Lidar Measurements to Produce
Gridded Biomass Estimates .................................................................................................................101
Use of partial-coverage UAV data as a sampling tool for large scale forest inventories..............................102
Using LiDAR remote sensing for identifying suitable habitat. Case: Darwin´s Fox .....................................103
Why Your Next Mapping Project Will Probably Use Stacking .....................................................................112
3-D Model of a Mediterranean Tree-Grass Ecosystem for Remote Sensing Applications............................ 113
3D measurement of tree health using multispectral intensity data from terrestrial laser scanners ............114
3D modeling of below-canopy global irradiance using terrestrial LiDAR data and ray tracing.................... 115
Forest Mapping & Inventory ........................................................................................ 117
A web service proposal for forest inventory of fast-growing species in areas with small-size
property using free software: GRASS GIS,WPS and LiDAR ...................................................................119
Above ground biomass related to field measurement errors .....................................................................121
Above-ground biomass estimation using calibrated multispectral aerial images in grasslands.
Do calibration targets matter? .............................................................................................................122
ALS-based forest inventory data and stem data from harvester to predict timber quality
of Norway spruce structural timber ..................................................................................................... 127
Characterization of forest structures through the synergetic fusion of airborne LiDAR and
multispectral sensor-derived difference vegetation index. ...................................................................128
Characterization of forest structures through the synergic fusion of airborne LiDAR and
multispectral sensor-derived difference vegetation index ....................................................................130
Effect of flying altitude, scanning angle and scanning mode on the accuracy of ALS based
forest inventory ................................................................................................................................... 131
Forest aboveground biomass mapping in Mexico using SAR, optical and airborne LiDAR data .................132
GEDI Biomass Model Development in Tropical Forests ............................................................................. 133
Improved forest cover mapping based on MODIS time series and landscape stratification .......................134
Improving assessment of fire risk in Yunnan Province, China using remote sensing ...................................138
8
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Industrial forest mapping: a Landsat Spatial and Temporal Approach .......................................................139
Improving merchantable timber volume accuracy for balsam fir plots by analyzing the spatial
distribution of airborne LiDAR returns .................................................................................................140
Industrial forest mapping: a Landsat Spatial and Temporal Approach .......................................................141
Inventory of Small Forest Areas Using an Unmanned Aerial System ..........................................................142
Large scale timber volume prediction with digital aerial photogrammetry and national forest
inventory data .....................................................................................................................................143
Large-Scale Prediction of Aboveground Biomass in Mountain Forests Utilizing Airborne Laser
Scanning .............................................................................................................................................144
Mapping Amazonian biodiversity and geology using basin-wide fern species inventories
and Landsat imagery ...........................................................................................................................146
Mapping certified forests for sustainable management - a tool for information improvement
through citizen science ........................................................................................................................147
Mapping forest degradation caused by fires in 2010 in Mato Grosso State, Brazilian Amazon
using Landsat TM fraction images .......................................................................................................148
Mapping forest height and biomass of the Chocó region, Colombia, combining stratified
random sampling of Lidar data and spaceborne remote sensing data ..................................................152
Mapping of forest attributes across Canada using Landsat pixel composites and LiDAR plots................... 153
Mapping the efficacy of fuel reduction burns using image-based point clouds..........................................154
Prediction of height, basal-area and stem volume in boreal forest using Pléiades or WorldView-2
acquisitions .........................................................................................................................................155
Regional predictive mapping of paludification black spruce forests in the north eastern
Canada using remote sensing and statistical modeling ........................................................................159
Scale-dependent mapping of stand structural heterogeneity from airborne LiDAR data ..........................160
Stereo matched very high-resolution satellite images for predictions of forest variables ..........................161
The 2016 NASA AfriSAR campaign for tropical forest structure and biomass measurements:
design, execution and first results ........................................................................................................162
Uncertainty estimation of stand structural variables in LiDAR-based forest inventories at
different sample sizes. .........................................................................................................................163
Forest Modelling ......................................................................................................... 165
Data assimilation of InSAR-based estimated forest stand attributes ......................................................... 167
Delineation of forest structure patterns in the circumpolar taiga-tundra ecotone .....................................168
Detecting the spread of invasive tree species in central Chile with combined Landsat
and Sentinel-2 data ............................................................................................................................ 169
How much forest area should be sampled to get accurate biomass estimations
at different scales?............................................................................................................................... 170
Lessons Learned – National Individual Tree Species Extent and Parameter Modeling for Insect
and Disease Risk Mapping ................................................................................................................... 171
Modeling forest structure and aboveground biomass integrating airborne LiDAR and satellite
Radar data ........................................................................................................................................... 172
Modeling of ALS data statistics in tree-level – application to single tree detection using
Bayesian inference .............................................................................................................................. 176
Modelling residual stand volume using unmanned aerial vehicles and digital aerial
photogrammetry ................................................................................................................................. 177
Modelling the effect of environmental factors on the height increment of stands with the use
of repeated Airborne Scanning data .................................................................................................... 178
Nationwide airborne laser scanning based models for volume, biomass and dominant height
in Finland.............................................................................................................................................180
Predicting the aboveground biomass of individual trees using remote sensing data and new
allometric models: a case study in Norway ..........................................................................................181
Prediction of alien species richness in two forest watershed of South-Central Chile: a remote
sensing synergistic approach ...............................................................................................................184
9
Universidad Mayor
Reconciling MODIS satellite with terrestrial forest inventory data to assess forest productivity
in Europe .............................................................................................................................................188
Relating forest height structure to virtual ground truth data .....................................................................190
Temporal and Angular Effects of the Spectral Signal on Deciduous Forest Crown Components.................191
Use of Random Forest Modeling Techniques to Predict and Detect Shrub Locations Under
Canopy using LiDAR Structure and Topography Metric ........................................................................192
Vegetation Chlorophyll estimated from multi-angle MODIS and tower hyperspectral
observations: A tool for scaling ecosystem seasonality and leaf demography across
Amazonian evergreen forests ..............................................................................................................193
Forest Monitoring ....................................................................................................... 195
A polyalgorithm for land cover trend and change detection ......................................................................197
AFIS - Wildfire Visualisation and Multi-sensor detection capabilities .........................................................198
Application of high-resolution satellite data for monitoring forest areas in changeable
climatic conditions ............................................................................................................................. 199
Assessing the cumulative climatic effects on regional forest decline dynamics in
coniferous forests ................................................................................................................................203
Assessing the predictions of high-resolution climate surfaces: a statistical analysis in a
Southern Hemisphere country ............................................................................................................ 204
Assessment of forest productivity from MODIS NPP data in relation to forest management
and optimal leaf area index..................................................................................................................205
Automatic recognition of burned areas with the use of a support vector machine (SVM) using
VNIR spectral bands with multiple satellite sensors. ........................................................................... 206
Barren ground caribou (Rangifer tarandus groenlandicus) behaviour after recent fire events;
integrating caribou telemetry data with Landsat fire detection techniques......................................... 208
Bringing Earth Observation Services for Monitoring Dynamic Forest Disturbances to the
Users – EOMonDis .............................................................................................................................. 209
Change Detection in Multitemporal SAR Orthoimages .............................................................................216
Changing northern vegetation conditions are influencing barren ground caribou (Rangifer
tarandus groenlandicus) behavior ........................................................................................................219
Characterization of the wildland-urban interface using LiDAR data and OBIA as a tool for fire
risk prevention and management at a local scale .................................................................................220
Combining Sentinel-1, Sentinel-2 and Landsat 8 images for near-real time forest change
detection ........................................................................................................................................... 224
Developing a U.S. national land use and land cover reference data set to support inter-agency
mapping, validation and statistical estimation needs .......................................................................... 226
Development of a UAV based platform for monitoring simulated disease expression using
time-series airborne laser scanning and high resolution multi-spectral imagery ..................................227
Dominant tree species dynamics informed by 30 years of Landsat time series in mountain areas
of Northern Spain ............................................................................................................................... 229
EO-1 SWIR band detection capability and comparison with Landsat .........................................................230
Estimación y monitoreo de cobertura de malezas a través de imágenes satelitales ..................................231
Exploring remote sensing potential in LULUCF Inventories in Aragón .......................................................232
Exploring remote sensing potential in land use, land use change and forestry (LULUCF)
inventories in Aragón...........................................................................................................................233
Fire behavior simulation from global fuel and climatic information ........................................................... 237
Forest area changes and impact of forest boundary delineation on change detection in forested
landscapes in Eastern Europe ..............................................................................................................241
Forest monitoring using remote sensing time-series: The case of Colombian Andes
Protected Areas .................................................................................................................................. 242
Global fire impacts assessment from long term analysis of burned area products .....................................243
Improving forest change detection in the UK using LandTrendr and TimeSync Landsat
analysis tools ...................................................................................................................................... 244
10
Universidad Mayor
Interpreted high resolution imagery for rapid assessment of land use and land cover changes
in the United States - The USDA Forest Service’s Image-based Change Estimation
(ICE) project ....................................................................................................................................... 246
Landsat reveals the impact of disturbance on carbon storage in the United States National
Forest System .....................................................................................................................................247
Landsat time series analysis – The impact of forest ecosystem history on biodiversity ............................. 248
Monitoring of forests to determine different levels of change ...................................................................250
Patagonian forests under attack: increasing large-scale insect outbreaks detected from
MODIS images ....................................................................................................................................251
Reconstructing forest changes in a fragmented landscape of southwest France from multiple
datasources: ecological implications....................................................................................................252
Remote sensing of photosynthetic light use efficiency of tropical ecosystems ..........................................253
Satellite-based monitoring of invasive species in Central-Chile .................................................................254
Spectral manifestation and signal to noise ratio of forest disturbance and recovery .................................255
Terra-i: A Pantropical Near Real Time Monitoring System for Vegetal Cover Change .................................256
Time series of Landsat images to determine burned area in the context of the Latin American
Network of Forest Fires (RedLatIF). .....................................................................................................257
Understanding the large area disturbance history of Australian Sclerophyll forests ..................................261
Use of New Technologies in Monitoring Mountain Forests’ Condition ...................................................... 262
Using historical satellite time-series to test an hypothesis of forest susceptibility to the bark
beetle outbreak ...................................................................................................................................263
Validating a Forest Canopy Disturbance Map in North Central USA ......................................................... 264
Water yield dynamics in forested watersheds: Using Landsat annual time series for the
assessment of eco services in national forests of the Intermountain West, USA .................................. 266
Forested wetland monitoring ...................................................................................... 267
3D mapping of mangrove forests along the Pacific Coasts of Central and South America ......................... 269
Finding the DAM signal: Utilizing time series of all available landsat TM/ETM+ observations
to map and monitor beaver-related flooding events ............................................................................270
Multi-Scale Remote Sensing of Mangrove Structure and Biomass/Blue Carbon ........................................274
P-Band DInSAR time series of river bank erosions: Preliminary results and comparisons with
field measurements ............................................................................................................................. 275
Statistical correction of Lidar-Derived digital elevation models with multispectral
airborne imagery .................................................................................................................................278
Forestry & Forest Management.................................................................................... 279
Analysing the relations between landscape structural changes and hydrological response at
subcatchment scale in temperate forest basins: the case of Maule’s inner dryland ...............................281
Large area tree species mapping in mixed temperate forests from multi-temporal RapidEye
satellite images and LiDAR data. ........................................................................................................ 285
Remote Sensing Contributions to Indicators of Biological Diversity in the U.S. National Report
on Sustainable Forests—2015 ..............................................................................................................287
Potential of using data assimilation to support forest planning ................................................................ 289
Use of remotely sensed data to spatially predict optimal final stand density, value and the
economic feasibility of pruning for even age plantation forests .......................................................... 290
Latin American Forests ................................................................................................291
Comparing Generalized Linear Models and random forest to model vascular plant species
richness using LiDAR data in a natural forest in central Chile ...............................................................293
Detection of low density natural forest in the Andes region using LANDSAT 8 imagery. .......................... 294
Effect of lidar pulse density on the prediction of aboveground biomass change in Brazilian
Amazon Rainforest ............................................................................................................................ 295
Estimating of the leaf area index in a forest fragment of mixed ombrophilous forest in Brazil,
using remote sensing techniques .........................................................................................................297
11
Universidad Mayor
Evaluating the ecological vulnerability of the remaining of Araucaria Forest – Southern Brazil .................300
Improving Observations of Tropical Forests with Optimized Terrestrial Lidar Scanners .............................305
Mapping forest degradation in the Valdivian Temperate Rainforest ecorregion ........................................306
Modeling aboveground biomass from individual tree LiDAR-derived metrics in tropical forest .................307
Phenological observations from a hyperspectral camera in the Amazonian Tapajos National
Forest ..................................................................................................................................................309
Recent trends of Land Surface Temperature and Vegetation Indexes over the Temperate
Rain Forest in Chile ..............................................................................................................................310
RDD+FREL-FRL and MRV ............................................................................................ 311
Assessing the carbon and water balance of Boreal forests using a process-based model driven
by satellite images ............................................................................................................................... 313
Comparison of local EO-based dense humid and dry forest cover and change area estimates
in the southwest forest massif of Central African Republic using the UMD global dataset ...................314
Comparison of local EO-based dense humid and dry forest cover estimates with the UMD
global dataset in the Central African Republic ...................................................................................... 315
Disentangling recent patterns in litter fall of European forests with remote sensing data across
a continental scale ...............................................................................................................................320
Estimating the dynamics of carbon stocks in forests with remote sensing data.........................................322
Linking Landsat 8 and forest inventory data for local biomass mapping in open canopy
woodlands ...........................................................................................................................................323
Mapping historical canopy cover change and recovery using Landsat time series
imagery (1972-2015) ............................................................................................................................324
Resumen Ejecutivo de Nivel de Referencia de Emisiones Forestales / Nivel de Referencia
Forestal Subnacional de Chile .............................................................................................................325
Synergistic use of sar and optical datasets for forest biomass retrieval and characterization
of forests in temperate zone – a national case study Poland ................................................................329
Using leaf-on and leaf-off airborne LiDAR to model vegetation structure and above-ground
carbon storage in the critical zone .......................................................................................................330
Using satellite data to estimate gas emissions into the atmosphere by burning
biomass in Mexico ............................................................................................................................... 331
12
Development of Methods
Development of Methods
14
Universidad Mayor
Universidad Mayor
ForestSAT 2016 Abstracts Summary
A novel sampling strategy for applying ALS-based
plot imputation to Australian native eucalypt forests
Gavin Melville1, Christine Stone2 and Mike Sutton3
Trangie Agricultural Research Centre, Mitchell Hwy, Trangie 2823, Australia.
Email: gavin.melville@dpi.nsw.gov.au
2
NSW Department of Industry - Lands, Forest Research, Level 12, 10 Valentine Ave., Parramatta 2124, Australia.
Email: christine.stone@industry.nsw.gov.au
3
Forestry Corporation of NSW, PO Box 100, Beecroft NSW 2119, Australia
1
Keywords: LiDAR; plot sampling design; nearest centroid; imputation; native eucalypt forests; inventory;
New South Wales.
Abstract
A new sampling design method, termed ‘Nearest Centroid’ (NC), based on k-means clustering, is
applied to ALS data captured over a native eucalypt forest on the south coast of New South Wales,
Australia. The NC method has been specifically designed for imputation and optimizes the survey
design by using the distance properties of the sample in the space defined by the auxiliary variables.
A total of 338 ground plots, 6 inventory variables and 44 ALS predictor variables were used in the
study. Reference plots were chosen from candidate plots according to the sampling strategies: simple
random; locally balanced and NC. Specifications in the imputation prediction process were also varied
and compared. In all cases, NC sampling was more efficient.
Introduction
Significant progress has been made in the utilisation
of airborne laser scanning (ALS) data (also referred
to as LiDAR data) to predict forest inventory
variables across a range of conditions. It is now well
established that the use of imputation in Australian
softwood plantations provides better precision than
traditional inventory methods [Melville et al., 2015;
Rombouts et al. 2015, Melville and Stone, 2016].
However, when ALS data are used on naturally
generated forests, such as native eucalypt forests,
the correlation between the LiDAR metrics and
the key forest attributes is not as pronounced as
it is with softwood plantations. There are several
reasons for this, including the diversity of species,
combined with different canopy heights, ages and
stem characteristics, all of which lead to a complex
three-dimensional structure which is encapsulated
in the LiDAR information.
Imputation is commonly used in conjunction with
LiDAR-based forest inventory and assessment
[McRoberts et al., 2015]. In the context of ALSenhanced forest inventory, usually the target
observations are comprised of pixel-level ALS
metrics from a LiDAR acquisition campaign. Specific
information relating to key attributes, such as
timber volume, is assigned to non-measured plots
from the most closely related measured plots as
defined by the similarity structure. In order to be
effective, ground reference plots selected for the
LiDAR imputation process must cover the full range
of variability in the targeted attributes.
There are several sampling methods which are suited
to a modelling approach including the widely used
grid-based sampling approach and/or strategies
which employ stratification, including stratification
based on LiDAR variables [e.g. Hawbaker et al.,
2009]. In addition to these conventional approaches,
15
Universidad Mayor
Development of Methods
there are methods based on spatial segregation such
as the generalized random tessellation stratified
sampling (GRTS) method [Stevens and Olsen, 2004]
and methods based on multiple associated variables.
For example, the sensor-directed response surface
(SDRS) approach [Lesch, 2005], constructs a sample
using a principal components analysis of variables
correlated with the variable of interest, whereas
the locally balanced sampling method [Grafström
et al., 2014] selects a subset of plots which are
simultaneously balanced over multiple auxiliary
variables. The new sampling method examined
in this paper relies on a cluster analysis which is
carried out in the multi-dimensional space defined
by the auxiliary variables, and is termed the nearest
centroid (NC) approach [Melville and Stone, 2016].
Methodology
Study area: The study area of 120,000 ha comprises
several State Forests close to the coastal town
of Eden on the south coast of New South Wales
(Latitude ranging from -36o20’ to -37o30’ and
Longitude ranging from 149o13’ and 150o00’). The
forests are dominated by a suite of native eucalypt
species that include Eucalyptus sieberi, E. muellerana,
E. agglomerata and E. fastigata and managed by the
Forestry Corporation of New South Wales (FCNSW)
for the sustainable extraction of sawlogs and
pulpwood. The silvicultural practices in this region
include a combination of thinning and alternate
compartment harvesting which results in a mosaic
of multiple aged stands across the landscape.
potential predictor variables in the ALS data were
reduced to around 8 by selecting variables which
minimised the mean squared prediction error.
Nearest Centroid sampling method: The Nearest
Centroid (NC) sampling method was developed
specifically with imputation in mind. It is assumed
that a reference plot well suited to imputation
would ideally have many nearest neighbours among
the target plots, that is, it is ‘well connected’ in the
covariate space [Melville and Stone, 2016; Melville et
al. 2016].The NC method is based on the arrangement
of the virtual plots in the multi-dimensional space
defined by the standardized auxiliary variables
(i.e. the LiDAR variables which are selected for
imputation). A suitable clustering algorithm is
employed, such as k-means clustering [Hartigan
& Wong, 1979] to segregate the virtual plots into
n clusters, where n is the sample size. Within each
cluster, the virtual plot closest to the cluster centroid
is chosen as the reference plot. When the reference
plots are sufficiently close to the cluster centroids,
these plots also become the nearest neighbour plots
within their respective clusters during imputation.
The method is illustrated in Figure 1 which displays
a series of target plots arranged into eight clusters,
based on the auxiliary variables p1m (proportion of
heights greater than 1 m) and mqh (mean quadratic
height). The centroid of each cluster is displayed as
an asterisk. In each cluster the reference plot closest
(in the auxiliary space) to the centroid is selected as
the reference plot.
approach: The simulations were
performed by separating the available plots at
random into a candidate set (NC=138 plots) and
a target set (N=200 plots), a process which was
repeated for each realization. A sample of reference
plots (n=25) was chosen from the candidate set
according to one of three sampling strategies;
simple random sampling, locally balanced sampling
[Grafström et al., 2014] and NC sampling. Irrespective
of the sampling method chosen, both the candidate
plots and the target plots are part of an overall grid
sample, as described above, which was used as the
original inventory. The reference plots were used to
predict the response variables in the target set and
the predicted values of the response variables were
compared to their actual values in order to evaluate
the sampling strategies in terms of relative root
mean square error (RRMSE), relative bias (RB) and
mean absolute deviation (MAD). The sample size
Simulation
Plot inventory and ALS data: Inventory data
obtained from 338 ha plots were collected during
June to September 2014 by Forest Data PL according
to FCNSW’s standard inventory specifications.
Each plot was 0.1 ha in size and established on a
systematic grid with random start at an approximate
density of 1 plot per 250 ha. Six response variables
derived from the inventory plot data were used in
this study, specifically: merchantable volume (m3)
(MVol); merchantable volume for trees over 60cm
(m3) (MVol60), basal area (m2ha-1) (BA) , stems ha-1
(SPH), stems ha-1 for trees over 60cm (SPH60) and
quadratic mean of mean dominant height for trees
over 60cm (m) (QMMDH60).
Discrete return ALS data were acquired over the
Eden forests in July 2013 by RPS Australia East PL
using a Harrier 68i LiDAR mapping system. The 44
16
Universidad Mayor
ForestSAT 2016 Abstracts Summary
was then varied from n=10 to n=75 so as to observe
the relationship between precision and the number
of reference plots.
which is 1.36 times more efficient than random
sampling (12.0% vs 14.0% RRMSE). Relative
efficiencies across all variables range from 1.52
(SPH) to 1.20 (QMDBH60). The effect of sample size
on the prediction of MVol is shown in Figure 2, where
once again the NC sampling method gave the best
precision.
Fig 1. Illustration of nearest centroid (NC) sampling
method
Results
NC sampling was the most efficient strategy for all
the response variables (Table 1). For example the
best precision for MVol occurs with NC sampling
Figure 2. Relative RMSE (%) of MVol vs sample size
for three sampling strategies
Table 1. Precision and bias of three sampling strategies using six response variables (n=25, k=1)
Variable
MVol
MVol60
SPH
BA
SPH60
QMDBH60
Sample method
random
balanced
NC
random
balanced
NC
random
balanced
NC
random
balanced
NC
random
balanced
NC
random
balanced
NC
RRMSE (%)
14.0
13.0
12.0
16.7
15.0
13.9
12.1
11.3
9.8
7.2
6.7
6.2
22.2
20.1
19.1
4.6
4.6
4.2
RBias (%)
-0.4
-0.1
-0.9
1.4
1.2
-3.9
3.2
3.8
1.4
0.9
1.2
1.2
-1.7
-1.1
3.3
-0.2
-0.3
-0.5
RMAD (%)
11.1
10.4
9.5
13.3
11.9
11.3
9.6
9.1
7.8
5.7
5.3
5.0
17.9
16.1
15.1
3.9
3.7
3.3
17
Development of Methods
Conclusions
Both NC and balanced sampling were more efficient
than random sampling when they were used with
imputation in the Eden native eucalypt forest, with
NC sampling the most efficient. Efficiency gains of
up to 1.5 or more were achieved, depending on the
response variable and sample size. This would allow
the sample size to be reduced by 33% without any
adverse effect on the prediction errors.
References
Grafström, A., Saarela, S., Ene, L.T. (2014) - Efficient
sampling strategies for forest inventories by
spreading the sample in auxiliary space. Canadian
Journal of Forest Research, 44: 1156-1164. http://
dx.doi.org/10.1139/cjfr-2014-0202.
Hartigan, J.A., Wong, M.A. (1979) - A k-means
clustering algorithm. Applied Statistics, 28: 104108.
Hawbaker, T.J., Keuler, N.S., Lesak, A.A., Gobakken,
T., Contrucci, K., Radeloff, V.C. (2009) - Improved
estimates of forest vegetation structure and
biomass with a LiDAR-optimized sampling design.
Journal of Geophysical Research,114: GE00E04.
http://dx.doi.org/10.1029/2008JG000870.
Lesch, S.M. (2005) - Sensor-directed response
surface sampling designs for characterizing
spatial variation in soil properties. Computers and
Electronics in Agriculture, 46: 153-179. http://
dx.doi.org/10.1016/j.compag.2004.11.004.
18
Universidad Mayor
McRoberts, R.E., Næsset, E., Gobakken T. (2015)
- Optimizing the k-Nearest Neighbors technique
for estimating forest aboveground biomass using
airborne laser scanning data. Remote Sensing
of Environment, 163: 13-22. http://dx.doi.
org/10.1016/j.rse.2015.02.026.
Melville, G.J., Stone, C., Turner, R. (2015) - Application
of LiDAR data to maximise the efficiency of
inventory plots in softwood plantations. New
Zealand Journal of Forestry Science, 45. doi:
10.1186/s40490-015-0038-7.
Melville, G., Stone, C. (2016) - Optimizing nearest
neighbour information – a simple, efficient
sampling strategy for forestry plot imputation
using remotely sensed data. Australian Forestry,
79:3, 217-228. http://dx.doi.org/10.1080/0004915
8.2016.1218265.
Melville, G, Stone, C., Rombouts, J. (2016) - Survey
designs which maximize efficiency gains in ALSbased forestry plot imputation. Proceedings of
Spatial Accuracy, July 5-8, 2016, Montpellier,
France, pp. 52-59.
Rombouts, J., Melville, G., Kathuria, A., Rawley,
B., Stone, C. (2015) – Operational deployment of
LiDAR derived information into softwood resource
systems. Forest and Wood Products Australia Ltd.
Stevens, D.L. Jr., Olsen, A.R. (2004) - Spatially
balanced sampling of natural resources. Journal of
the American Statistical Association, 99: 262-278.
http://dx.doi.org/10.1198/016214504000000250.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Airborne Dual Band Radar for Rural Cadastre
Dieter Lübeck1, Rafael A. S. Rosa1, Christian Wimmer1, Karlus A. C. Macedo1, Juliano Lázaro1
Bradar Indústria S.A.
1
Keywords: rural cadastre, synthetic aperture radar, SAR, X-band, P-band.
The land ownership and land use rights are high
on the global agenda. 50% of the sustainable
development goals for 2030 are land related. A
very rough estimation for reaching these goals with
traditional approaches in cadastral data acquisition
ends up in over 500 years of cadastral survey work.
This estimation makes clear that breakthrough
techniques will be necessary to overcome this
demand.
A well-known approach for surveying big areas
in a short time is airborne or space borne remote
sensing. In Fit-For-Purpose approaches for land
administration printed imagery is used in the field
for boundary identification in a participatory way.
The question is, if radar image data can support
the mapping of cadastral boundaries in an efficient
way. A promising new approach in this area is using
airborne dual band InSAR (X- and P-Band). Due to
the known advantages, like cloud penetration, sun
light independency and foliage penetration, radar
imagery is increasingly being used for large scale
mapping and monitoring over the last decade.
Bradar´s airborne dual band Radar uses X- and
P-band because they have the most complementary
mapping characteristics. While X-band, with 3 cm
wavelength, maps exclusively the surface, P-band,
with 70 cm wavelength, penetrates the foliage and
allows mapping the topography below vegetation.
In the former days, the main application of this
technology was the interferometric phase processing
for the topographic mapping. Nowadays the high
quality images from the amplitudes are opening
new horizons for planimetric feature extraction
in scales up to 1:5.000 with the big advantage of
cloud penetration and a wide mapping swath. The
mentioned airborne radar systems combines all
these characteristics and could already map an
area of about 2.3 million km² in Central and South
America and Europe.
Detect fences
New studies, realized by Bradar in Brazil, are showing
that P-Band is not only capable to detect boundaries
demarked by roads, river or vegetation, but also
very small wire fences, which are commonly used
for example in South America for demarking land
(Figure 1). Applying this technology to supporting
land registration and cadastral survey could be the
solution for speeding up registration processes. This
brings help to the local people in protecting them
from land grabbing.
Figure 1: Wire fence constructions are very often
cadastral boundaries. Those fences can be easily
detected with radar.
19
Universidad Mayor
Development of Methods
Surprising results
Application
Tests showed that the feature extraction of these
fences can be realized with a high planimetric
accuracy of up to 20 cm in the P-Band imagery
(Figure 2), which meets the requirements for
support in land administration in many developing
countries. This surprisingly result can be achieved
because the horizontal wires of the wire fences
act like small antennas for the horizontal polarized
P-Band radiation. With the presence of the ground
we have the effect of the so called double-bounce
of the wave back to the radar. Due to this effect the
signal of all horizontal wires of the fence sum up and
the fence appears in the image as one line only with
its origin at the vertical projection of the fence on
the ground. Figure 3 shows an example of a P-Band
image with a fence network in a deforested area.
Beside the fences, other important features can
be extracted, like forest, road, construction, single
trees and animals.
The radar imagery can be applied for integral
cadastral boundary detection. Mostly those
boundaries are visible features - defined by roads,
vegetation, rivers, fences and constructions in a
very efficient way. Even rivers covered by vegetation
can be extracted. Figure 4 shows an example of
an area in Brazil, including orthoimages X, P, X/P
colour, and extracted boundaries overlaid to P-band
image (from left to right). Additionally, digital
terrain and/or surface models can be provided by
applying interferometric flights to obtain altimetric
information for contour lines generation and
drainage networks, including 3D simulations. This
information is not only useful for supporting the
land administration but also the rural development,
infrastructure projects, forest certification and
rural environmental cadastre, risk mapping, area
preservation, land us mapping, and topographic
mapping, among many other applications.
Change detection
Figure 2: Planimetric accuracy for fence detection
in P band imagery – the fence can be projected to
the surface of the earth.
Figure 3: P-Band Image of
rural area. Many of the line
shaped features are cadastral
boundaries.
20
For change detection the radar imagery has another
big advantage, which is the constant illumination
between different flights. This allows, on the
contrary to optical imagery, systematic comparisons
of the images supported by image processing
software. Changes in the land use can be detected
automatically, as well as in any type of construction.
This provides the option to reconstruct “replaced”
cadastral boundaries.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Figure 4: Example of boundary extracting from radar imagery.
Experience in Brazil
Fit-For-Purpose
Two technological viability projects, one close to
the hydropower station of Belo Monte and one
in the district of São José dos Campos already
demonstrated the applicability of this approach.
Agencies in Brazil and Latin America never
considered remote sensing for the development
of land administration. Actual governmental
regulations do even not allow it without performing
a complete ground survey. Radar survey accelerates
such ground survey and serves as a control method.
In Fit-For-Purpose approaches the Radar imagery
is considered to be as sufficient in order to collect
cadastral boundary data fast, quick and cheap. It
means that the printed imagery can be taken to
the field to compare the automatically extracted
features with the real cadastral boundaries. Those
boundaries can be marked with pen on top of the
image. Back in office the polygons for spatial units
(parcels) can be easily identified and included with
expected accuracy.
The planimetric accuracy was measured with
differential GPS at both sites. Wire fences presented
an RMSE error of around 20 cm and water boundaries
of around 1 m. The radar survey will not completely
substitute the ground survey, but it allows the
general definition of the ground boundaries and
their vectorization before going to the field.
In Brazil, for example, the interview with the ground
owner, the review of the ground boundaries and the
demarking must be carried out by the ground survey.
The radar could reduce from 3 to 5 times the effort
and execution time of the conventional field work.
21
Universidad Mayor
Development of Methods
An approximation for non-parametric bootstrapping to assess
uncertainties in imputation mapping of forest vegetation
David M. Bell, USDA Forest Service, Pacific Northwest Research Station
Matthew Gregory, Oregon State University, College of Forestry
Keywords: bootstrapping; k-nearest neighbor imputation; Landsat; precision; uncertainty
Abstract
Satellite-based vegetation maps are an increasingly common source of information for natural resource
planning, but the utility of these products is unclear without an understanding of their limitations.
Nearest neighbor imputation methods for mapping forest attributes across broad geographic areas
have been employed extensively, motivating research into uncertainties associated with imputed maps.
While uncertainties can be quantified in a variety of ways, non-parametric bootstrapping has emerged
as a potential alternative to parametric methodologies, especially when plots are directly imputed to
pixels (i.e., k-nearest neighbor with k = 1). However, bootstrapping methods for examining uncertainties
in imputations for large regions requires substantial computational power and data storage, limiting
the usefulness of bootstrapping for assessing uncertainties over the large areas required for regional
planning and monitoring activities. In this research, we describe a new approximation for a nonparametric bootstrapping approach for kNN with k = 1 methods (hereafter, bootstrap approximation).
Our objectives were to (1) assess the capacity of the bootstrap approximation method to reproduce
results based on non-parametric bootstrapping and (2) examine regional variation in imputation
uncertainties in forest attributes in landscapes of California, Oregon, and Washington, USA. We
performed k-NN with k = 1 imputation mapping with 4000 bootstrap samples of field plot data and
generated several metrics of prediction uncertainty and central tendency. In contrast, our bootstrap
approximation was based on a weighted sample of the seven nearest neighbors in environmental
and Landsat spectral space. Our results indicate that the bootstrap approximation are essentially
equivalent with results produced by non-parametric bootstrapping for measures of central tendency
(means and medians), presence of structural elements (e.g., attribute greater than zero), and measures
of precision (width of confidence intervals), with coefficients of determination (r2) ranging from 0.996
to 1.000. Using our bootstrap approximation, we found that spatial variation in climate, ownership,
and disturbance history explained geospatial patterns in predicted forest attributes as well as the
variability in the imputation process. The bootstrap approximation presented in this research provides
a useful new tool for examining the limitations of kNN with k = 1 imputation, informing both scientists
and managers of the limitations of consequent vegetation mapping products.
22
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Analysis of land cover classes separability using optical and
SAR data in areas of high cloud content in northern South
America
Sebastián Palomino-Ángel
Catedrático, Facultad de Ingeniería, Universidad de Medellín, Cra 87 Nº 30 – 65, Medellín, Colombia,
spalomino@udem.edu.co
Jesús A. Anaya-Acevedo
Profesor Titular, Facultad de Ingeniería, Universidad de Medellín, Cra 87 Nº 30 – 65, Medellín, Colombia,
janaya@udem.edu.co
Keywords:Land cover, high cloud content, optical data, radar data, Jeffries-Matusita
Abstract
Land cover classification in tropical regions is challenging because of the dynamics of land cover
change, persistent cloud cover, and presence of small patches. Due to its location and topography,
Colombia has different physical conditions which favors the formation of cloud cover almost over the
full extent of the country. This complexity is more pronounced to the west because of the interaction
of the humid currents of the Pacific Ocean and the Andes mountains. Northwest Colombia, is an
interesting region due to its forest and biodiversity, but those natural resources have been affected by
drivers of forest loss, mainly cattle grazing and agricultural activities. It is estimated that 70% of native
forest has been subject to land cover change in the last 40 years.
The main objective of this work was to assess the ability of radar data in separating land cover classes,
and the potential of optical and radar combination for improving classification accuracy in a region of
high cloud content in northwest Colombia. Landsat 5 and Phase Array L-Band SAR (PALSAR) were
used as input data. Jeffries-Matusita (JM) distance was used as a separability measure for the analysis.
In order to perform the classification of satellite images, we used maximum likelihood classification
algorithm for optical data and Wishart Supervised Classification for radar data. High resolution images,
aerial photographs and field information were used as reference data for training and validation.
A total of 22 classes were evaluated using JM analysis: 3 crops classes, 15 vegetation cover classes
(including different forests types and typical wetland vegetation), 3 pastures classes and 1 urban
class. The JM distance ranges from 0 to 2, being 0 the value for low separable classes and 2 the value
for high separable classes. It was found that higher JM values were obtained for optical and radar
combination, with a mean value of 1.94 JM distance for all classes. Optical data presented a mean value
of 1.90 JM distance and radar data presented a mean value of 1.16. Additionally, accuracy assessment
was performed using error matrix method. Optical, radar, and the combination of optical and radar
resulted in different overall accuracies: 75%, 50% and 76% respectively. The combination of optical
and radar not only has the best accuracy, it also reduces optical invalid data due to cloud content. The
classification problems were evaluated, focused on determining the causes of error, and identifying
techniques to improve results in areas of high cloud content.
23
Universidad Mayor
Development of Methods
Analyzing some factors affecting the extraction
of full-waveform LiDAR metrics and their effect
in forest structure variable estimates
Ruiz, Luis A.1, Crespo-Peremarch, Pablo 1, Estornell, Javier 1, Balaguer-Beser, Angel 1
1
Geo-Environmental Cartography and Remote Sensing Group, Universitat Politècnica de València,
Camino de Vera, s/n, 46022 Valencia, Spain.
email: laruiz@cgf.upv.es
Keywords: full-waveform; forest structure; fuel parameters; LiDAR metrics.
Abstract
Introduction
Full-waveform LiDAR is a promising technique to analyze vertical structure of forest stands., and
derived metrics provide better forest structure and fuel variables estimates than using discrete
LiDAR (Hermosilla et al., 2014; Cao et al., 2014). However, full-waveform derived metrics are more
sensitive to side-lap effects during data acquisition than discrete LiDAR and data are usually subject to
voxelization processes in which some methodological parameters, such as voxel size and voxel value
assigning criteria, affect the final forest structure models (Crespo-Peremarch et al., 2016).
The objectives of this work are: (i) to assess the effect of two full-waveform parameters, voxel size
and voxel value assigning method, on the estimation of forest structure and fuel variables; (ii) to
compare these methodological parameters in two different types of forest areas, Mediterranean and
temperate US Pacific Northwest; and (iii) to explore the full-waveform LiDAR data side-lap effect on
the derived metrics.
Data and methods
Full-waveform LiDAR data were acquired in a
temperate US Pacific Northwest forest in Panther
Creek, Oregon, and in a Mediterranean forest
located in Sierra de Espadán (Spain), using a
Leica ALS60 and a LiteMapper 6800, resulting in
pulse densities of >8m-2 and >11m-2, respectively.
Inventory field measurements were collected in
84 circular plots (16m radius) in Panther Creek,
and in 55 circular plots (15m radius) in Espadán. In
the first area, the dominant species is Douglas-fir
(Pseudotsuga menziesii), occasionally mixed with
other conifers, such as western hemlock (Tsuga
heterophylla), western red cedar (Thuja plicata) and
24
grand fir (Abies grandis), or deciduous species such
as red alder (Alnus rubra) and bigleaf maple (Acer
macrophyllum). In the second area, the dominant
forest species is Pinus halepensis, occasionally mixed
with Quercus suber, with the common presence of
understory shrub vegetation. LiDAR metrics were
extracted as described by Duong (2010) and Cao et
al. (2014) after a voxelization process (Hermosilla
et al., 2014). They were used to create regression
models for six forest structure and fuel variables
in both areas (aboveground biomass, basal area,
quadratic mean diameter, canopy height, canopy
base height and canopy fuel load). Four voxel sizes
(0.25m, 0.5m, 1m and 2m), and six different types of
voxel assigning methods (maximum, mean, median,
Universidad Mayor
ForestSAT 2016 Abstracts Summary
mode, percentiles 90 and 95) were tested, computing
the adjusted coefficient of determination R2 and
root-mean-square error (RMSE) and using leaveone-out cross-validation for evaluation. The effect
of LiDAR density variations within and around sidelap areas was also analyzed using pairwise samples
in the Panther Creek study site.
Results
When small voxel sizes are used (0.25m and even
0.5m), there are very few waves contained in each
individual voxel, and the different voxel assignation
methods yield very similar R2 results. Voxel sizes
near 0.5 m are, in general, the most efficient in
generating accurate models. Using voxels sizes
over 1m increases the influence of the voxel value
assigning method. Maximum and high percentiles
(95%) are usually the most efficient assigning criteria.
Table 1 shows, as an example, the R2 values of the
aboveground biomass when these methodological
parameters are varied.
The R2 of forest structure variables related to height,
such as canopy height and canopy base height, are
not significantly affected by the voxel size and voxel
value assigning criteria tested. In these cases, the
influence of the values of internal voxels is negligible.
This suggests that the use of full-waveform LiDAR
does not increase the accuracy of these height
variables models in comparison to the use of discrete
LiDAR. Results obtained in both areas are consistent,
some differences are found probably due to the
presence of different species and canopy types.
Regarding the side-lap effect, metrics such as
number of peaks of the wave (NP), roughness of
outermost canopy (ROUGH) and front slope angle
(FS) are the most affected by density differences.
Using the mean value as voxel assigning method
reduces this effect. Metrics height of median
energy (HOME), height to median ratio (HTMR) and
waveform distance (WD) are the most efficient and
robust in modeling forest structure attributes.
Main conclusions
When using full-waveform LiDAR data for modeling
forest structure variables following a voxelization
process, voxel size should be chosen based on the
density of LiDAR data by ensuring the presence of
waves within the voxels but avoiding an overload of
values. For medium pulse densities (8-12 m-2), voxel
sizes around 0.5 m are recommended. Assigning the
maximum wave value to each voxel, the negative
effect of increasing the pixel size is reduced. Some
full-waveform derived metrics are more efficient
and less affected by density differences than others,
especially when side-lap effect exists. This fact
should be considered when creating the models, by
employing those metrics less affected. Voxel value
assigning criteria should be selected attending to
the voxel size and the heterogeneity of the LiDAR
point density in the study area. Further work will be
focused in density homogeneization techniques to
avoid side-lap effect.
Table 1. R2 values of the forest stand variable Aboveground biomass when modifying the full-waveform
methodological parameters voxel size (0.25, 0.5, 1 and 2m) and voxel assigning methods
(maximum, mean, median, mode, p90 and p95) in the two study areas.
Maximum
Mean
Median
Mode
p90
p95
0.25m
0,84
0,84
0,84
0,84
0,84
0,84
Aboveground biomass
Panther Creek
Espadan
0.5m
1m
2m
0.25m 0.5m
1m
0,85
0,85
0,84
0,79
0,79
0,79
0,84
0,84
0,83
0,79
0,79
0,78
0,84
0,83
0,82
0.78
0,79
0,77
0,82
0,82
0,82
0,79
0,79
0,75
0,85
0,84
0,84
0,79
0,79
0,78
0,85
0,85
0,84
0,78
0,79
0,79
2m
0,80
0,76
0,72
0,69
0,78
0,78
25
Development of Methods
References
Cao, L., Coops, N.C., Hermosilla, T., Innes, J., Dai,
J., She, G., 2014. Using small-footprint discrete
and full-waveform airborne LiDAR metrics to
estimate total biomass and biomass components
in subtropical forests. Remote Sensing, 6, 71107135.
Crespo-Peremarch, P., Ruiz, L.A., Balaguer-Beser,
A., Estornell, J., 2016. Analysis of the side-lap
effect on full-waveform LiDAR data acquisition
for the estimation of forest structure variables.
Int. Arch. Photogramm. Remote Sens. Spatial
Inf. Sci., 12-19 July, Prague, XLI-B8, pp. 603-610.
Duong, H.V., 2010. Processing and application
of ICESat large footprint full waveform laser
range data. Ph.D. Thesis, Delft University of
Technology, Netherlands.
Hermosilla, T., Ruiz, L.A., Kazakova, A.N., Coops,
N.C, Moskal, L.M., 2014. Estimation of forest
structure and canopy fuel parameters from smallfootprint full-waveform LiDAR data. International
Journal of Wildland Fire, 23(2), 224-233.
26
Universidad Mayor
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Analyzing the Lorenz Curve of Tree Growth Dominance with
Multi-temporal Airborne Lidar in Wytham Forest (UK)
Rubén Valbuena1, 2, Matti Maltamo1, Lauri Mehtätalo1, David Coomes2
1
University of Eastern Finland. Faculty of Forest Sciences. Yliopistokatu 7. 80100 Joensuu, Finland
University of Cambridge. Department of Plant Sciences. Downing St. CB2 3EA Cambridge, UK.
2
Keywords: Airborne laser scanning; Tree competition; Metabolic scaling theory
Abstract
Growth dominance curves are employed to study the relative shares of the total growth accounted
for each tree in the forest community (Binkley, 2004). These relations can be used to define patterns
of tree competition: symmetric or asymmetric (Pretzsch & Biber, 2010). Growth dominance curves
are one specific type of Lorenz curve, and therefore their associated growth dominance coefficient is
also a particular type of Gini coefficient (GC). Relations of tree size inequality are determined by these
patterns of competition, and hence the GC can be valuable to study asymmetric competition (Cordonnier & Kunstler, 2015). After a concentrating our research on static studies applying the Lorenz curve
and derived indicators and determining the means for predicting them with airborne Lidar (Valbuena
et al., 2013), our research objectives are now on the study of growth dominance using multi-temporal
datasets. We studied the growth dominance curves that should correspond to predictions by the metabolic scaling theory (MST) of tree growth (Enquist et al. 1999), which provides grounds for predicting structure and dynamics of an average idealized natural forest. Empirical divergences from MST
predictions have been, however, considered to be caused by forest disturbance, asymmetric competition among trees in the forest community, and the allocation of different types of resources (Coomes,
2012). Lidar is a valuable tool in order to investigate the relations of competitive dominance among
trees in the forest community, since it allows to model the light environment, and investigate forest
structure and disturbance.
Our research was structured in two phases. The first phase consisted in a deductive approach to determine mathematically the growth dominance curve that corresponds to MST predictions, and therefore determine a scenario of pure symmetric competition. Empirical divergences from that ideal
situation were analysed in a second inductive phase, which consisted in quantifying these divergences
and predicting them using standard area-based Lidar approaches. This second phase was carried out
at the Wytham Woods, Oxfordshire (UK), a woodland comprised mainly of ash, oak and sycamore
trees, for which were obtained detailed data measured in 2008 and 2010. We calculated wall-to-wall
growth dominance coefficients over an 18-hectare plot, and estimated them using airborne Lidar. The
underlying envisaged goal is to enable the use of Lidar for determining whether symmetric or asymmetric competition processes apply to a given area of forest.
27
Development of Methods
References
Binkley, D. (2004) A hypothesis about the interaction
of tree dominance and stand production through
stand development. Forest Ecology and Management 190(2):265–271
Cordonnier, T., & Kunstler, G. (2015). The Gini index
brings asymmetric competition to light. Perspectives in Plant Ecology, Evolution and Systematics,
17 (2), 107-115.
Coomes, D. A., Holdaway, R. J., Kobe, R. K., Lines,
E. R. and Allen, R. B. (2012), A general integrative framework for modelling woody biomass production and carbon sequestration rates in forests.
Journal of Ecology, 100: 42–64.
Enquist, B., West, G., Charnov, E. & Brown, J. (1999)
Allometric scaling of production and life-history
variation in vascular plants. Nature, 401, 907–911.
Pretzsch, H., Biber, P. (2010) Size-symmetric versus
size-asymmetric competition and growth partitioning among trees in forest stands along an
ecological gradient in central Europe. Canadian
Journal of Forest Research 40:370–384
Valbuena, R., Packalen, P., Mehtätalo, L., García-Abril, A., Maltamo, M., 2013a. Characterizing
forest structural types and shelterwood dynamics from Lorenz-based indicators predicted by
airborne laser scanning. Canadian Journal of Forest Research 43, 1063–1074.
28
Universidad Mayor
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Change detection techniques for development of compatible
height growth and site index models using repeated Airborne
Laser Scanning Data
1
Jarosław Socha*1, Marcin Pierzchalski 1 , Radomir Bałazy2, Mariusz Ciesielski2
University of Agriculture in Krakow, Faculty of Forestry, Al. 29 Listopada, Krakow, 32-425 Poland
2
Forest Research Institute, Braci Leśnej 3, Sękocin Stary, 05 – 090 Poland
*
Speaker, e-mail: rlsocha@cyf-kr.edu.pl
Keywords: Change detection, height growth, Norway spruce, LiDAR, dynamic site index model,
site productivity
Abstract
Forest site productivity, which is a quantitative estimate of the potential of a site to produce
plant biomass remains a fundamental variable in forestry (Bontemps and Bouriaud, 2013;
Skovsgaard and Vanclay, 2008). The most commonly used and widely accepted method of
evaluating site productivity is site index (Bontemps and Bouriaud, 2013; Coops et al., 2011;
Hägglund and Lundmark, 1977; Sharma et al., 2012; Skovsgaard and Vanclay, 2008). Therefore
the construction of site index models remains a fundamental task for site productivity
differentiation (Cieszewski and Strub, 2007; Mathiasen et al., 2006; Sharma et al., 2011).
To date three main data sources were used for site index model development: 1) PSP repeated measurements on permanent sample plots (PSP); 2) Temporary Sample Plots
data (TSP) from periodic inventories that usually cover most of the forested areas of a given
territory (Raulier et al., 2003); 3) Stem analysis (SA) data. SA deduces past height growth from
growth ring observations made on dissected sample trees.
Recently, a number of remote sensing technologies have emerged that offer alternative
approaches for forest inventories, with potential of improving the accuracy and precision of
stand-based measurements while reducing data acquisition cost (Hilker et al., 2008). Laser
scanning of forest stands provides the mean canopy height of dominant and codominant
trees. Hyyppä et al. (2003) demonstrated that multi-temporal ALS data could be used to
measure forest growth and the standard error when estimating height growth at stand
level was less than 5 cm. St-Onge and Vepakomma (2004) used ALS with the density data
acquired in 1998 and 2003 to detect tree growth within a 6.8 ha study site in Québec, Canada.
Hopkinson et al. (2008) evaluated the estimation of plot level mean tree height growth
using four ALS acquisitions of a pine plantation in Toronto and found that the 100th height
percentile provided the most robust overall direct estimate of field measured forest growth,
and demonstrated that there was no statistically significant difference between all plot-level
field and Lmax ALS mean growth rate estimates. Yu et al. (2004) provided an assessment of
Norway spruce (Picea abies (L.) Karst.) and Scots pine (Pinus sylvestris L.) canopy growth
from two lidar acquisitions 21 months apart. Precision of plot-level growth estimated by
differences in raster canopy height models (CHMs) was of 10 to 15 cm. Repeated LiDAR
observations enable measurement of tree and stand height growth over time ((Hopkinson
et al., 2008; Yu et al., 2006)Yu et al. 2006, 2008; Næsset and Nelson 2007). However, to date,
29
Development of Methods
Universidad Mayor
repeated LIDAR height measurements data have not been used in modelling height growth
and site index models development, and as it was stated by Coops in recent review concerning
Characterizing of Forest Growth and Productivity using Remotely Sensed Data (Coops, 2015)
considerable research is required to determine the appropriate design of repeat LIDAR surveys
for measurement of tree height growth. In his detailed review Coops (Coops, 2015) indicated
also need to better understand the amount of time needed for sufficient growth to excess
noise and other uncertainties within LIDAR systems. Presented research is the example of
successful attempt of answer for both above mentioned questions.
The objective of this research is to describe the new application for change detection using
airborne laser scanners and present the procedure for the development of compatible height
growth - site index models using repeated ALS data. We showed how wall-to-wall light detection
and ranging (LiDAR) data obtained for large forest area exceeding 70 thousand ha could be
used in development site index model, which appropriately reflects unique region and sitespecific growth trajectories. LIDAR measurements may be biased as a result of several factors,
including instrument specifications, flying height, species architecture, and the measurement
method used (Wulder et al., 2008. We hypothesized that in spite of possible errors observed
in case of single area units (raster plots), that can range from a few centimetres up to a few
meters (e.g., Aldred and Bonnor 1985, Chen and Ni 1993, Ritchie 1995, Latypov 2002, (Yu et
al., 2006)) repeated complete LIDAR coverage on large areas should give unbiased mean
height growth trend and increment estimates and therefore may be recognised as new fully
valuable data source for stand height development modelling. Developed using repeated ALS
data compatible height growth and site index models were validated using independent data
sets consisting 156 height growth trajectories obtained by stem analysis of dominant trees
collected on the research area.
To the best of our knowledge, out study constitutes the first practical application of change
detection by airborne laser scanners for the development of compatible height growth - site
index models using repeated ALS data.
30
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Combining remote sensing and dendrochronology to assess
the effect of groundwater extraction on Prosopis tamarugo:
from tree to aquifer level
Mathieu Decuyper1,2*, Roberto O. Chávez3, Jan Clevers1, Martin Herold1
Laboratory of Geo-Information Science and Remote Sensing, Wageningen University, Droevendaalsesteeg 3, 6708 PB
Wageningen, The Netherlands
2
Forest Ecology and Forest Management Group, Wageningen University, Droevendaalsesteeg 3, 6708 PB Wageningen,
The Netherlands
3
Institute of Geography, Pontificia Universidad Catolica de Valparaíso, Valparaiso, Chile
1
Abstract
Groundwater dependent ecosystems are sensitive to changes in groundwater availability, and
therefore, threatened by desertification as a consequence of global warming. In this study, we
combined remote sensing and dendrochronology to assess drought stress on Prosopis tamarugo
Phil, endemic to the Atacama Desert in Northern Chile. This unique ecosystem is threatened by
groundwater (GW) overexploitation due to mining and urban consumption. We first performed a
spatio-temporal assessment of Tamarugos’ water condition for the whole Pampa del Tamarugal
aquifer (comprising the Tamarugo ecosystem) using two NDVI (Normalized Difference Vegetation
Index) -derived parameters: NDVI in winter (NDVIw) and the difference between NDVI in summer
and winter (NDVIws). These were calculated using Landsat satellite imagery. Based on this aquifer
level assessment, we selected two representative sites (control versus high-depletion site) to perform
more detailed remote sensing based estimations of canopy growth and water condition at the stand
level, and tree-ring based analysis of stem growth on both the stand and tree level. Also at the tree
level the Green Canopy Fraction (GCF), obtained from a Worldview2 image was used as an indicator
of the canopy greenness.
Radial stem growth was determined for the same period after cross-correlating the tree rings per site.
Time-series analysis was done over a period of 26 years and both NDVI-derived parameters showed
significant negative trends in the high-depletion site, indicating drought stress. Radial stem growth
of viable P. tamarugo trees was 48% lower in the high-depletion site. At the tree level, the GCF also
indicated drought stress since a larger percentage of trees fell within lower GCF classes.
On the aquifer level, NDVIw and NDVIws of the P. tamarugo forest declined on average 19% and
51%, respectively, while GW depleted on average 3 m over the period 1988-2013. About 730,000
trees were identified in the study area, from which 5.2% showed a GCF < 0.25 associated with severe
drought stress. A GW level > 12 m increasingly limited the paraheliotropic leaf movement, leading
to dehydration and foliage loss. This is in line with the indications at plot level for both remote
sensing as well as radial stem growth. Overall we can conclude that P. tamarugo trees at 12-16 m GW
level suffered moderate drought stress while a GW level of 16-20 m implied severe drought stress.
Combining dendrochronology and remote sensing allows us to understand the effects of drought
stress on two different carbon pools (crown and stem, respectively), providing more insights on the
physiological response of the species to drought and on potential management actions to minimize
the impacts of water extraction.
31
Universidad Mayor
Development of Methods
Comparison of forest canopy point-measurement methods
from above and below
Jaan Liira, Marta Mõistus, Kertu Lõhmus
In remote-sensing, mayor effort has been paid to the estimation of LAI and canopy height of the stand
overstory to quantify the timber volume and carbon storage. Many forest ecosystems in the World,
however, consist of multiple layers in understory, which play as important or even more important role
in provision of ecosystem services and human welfare than just the first layer of trees. Understory is
widely ignored in remote sensing because of various reasons, but in conditions of increasing quantity
of remote sensing sources and methods, the methodological differences should be estimated. We
compared different areal LiDAR and ground-based methods to evaluate the estimation bias of canopy
density between them. We capitalized on understory management gradient from mature forest to
old parks, that to avoid stand growth related side effects. We observed that all methods suffer from
the hiding effect of distant layers in multi-layered stands. Therefore, more elaborated or complex
measured should be implemented to reduce the stand characterization bias, including overall carbon
storage or biodiversity.
32
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Generation of a Landsat image mosaic over the
Amazon rainforest
Jasper Van doninck, Hanna Tuomisto
Amazon Research Team, Department of Biology, University of Turku, Finland
Keywords: Amazonia, ETM+, TM, preprocessing, BRDF, per-pixel compositing
Although Amazonian forests may at first sight seem one homogenous mass, field surveys have revealed
high floristic heterogeneity, even over short distances. The combination of this high local-scale variability
with the enormous extent of the Amazon forest makes satellite remote sensing an indispensable tool
in Amazonian biodiversity monitoring. Landsat imagery offers the desired combination between
high spatial resolution and basin-wide coverage, but the use of this dataset over Amazonian forests is
hampered by some practical issues. A first obvious problem is the persistent cloud cover over tropical
forests. Consequently, only an extremely small fraction of the available Landsat images over Amazonian
forests are entirely cloud-free, and per-pixel multi-temporal image compositing is a necessity in order
to obtain a basin-wide mosaic. In case multiple cloud-free observations are available for a pixel, the
choice of the compositing may influence the radiometric consistency of the final mosaic. A second issue
is the large spectral similarity between floristically distinct forest types. Because of this, even small
radiometric artefacts can impede image interpretation or classification. One of these artefacts is an
across-track reflectance gradient, caused by the bidirectional reflectance distribution function (BRDF).
Failure to remove directional effects will result in clearly discernible edges between adjacent Landsat
scenes in a larger mosaic.
We here present a preprocessing chain for generating a radiometrically consistent Landsat TM/
ETM+ image mosaic. First, atmospherically corrected surface reflectance products are corrected for
directional effects using the Ross-Li BRDF model calibrated with multi-angular Landsat observations.
The directional normalization is validated using the overlap area of Landsat images in adjacent paths.
The validation shows that directional effects over Amazonian forest can cause surface reflectance
differences up to 4% in the near-infrared band over the Landsat swath, but that these are eliminated
after normalization. In a second step, multi-temporal Landsat images are combined in a per-pixel image
compositing. It is shown that, depending on the used compositing method, radiometric consistency of
the composite image may improve drastically with an increasing availability of clear observations. The
resulting image mosaic reveals clear floristic and geological patterns within Amazonia at both the local
and the basin-wide scale.
33
Development of Methods
Universidad Mayor
Detection of Dead Standing Eucalyptus for managing
biodiversity using full-waveform LiDAR data
Milto Miltiadou1,2,3, Neill D.F. Campbell1, Matthew Brown1, Susana Gonzalez
Aracil3, Tony Brown4, Darren Cosker1 and Michael Grant2
1
The Centre for Digital Entertainment, University of Bath, Bath, UK
2
Remote Sensing Group, Plymouth Marine Laboratory, Plymouth, UK
3
Interpine Group Ltd, Rotorua, NZ
4
Forestry Corporation of NSW, Wauchope, Australia
Keywords: full-waveform LiDAR data, voxelisation, biodiversity, dead trees
The value of dead standing Eucalyptus Camaldulensis (Red Gum) from a biodiversity management
perspective is large. In Australia, many arboreal mammals and birds, of which some of them are close
to extinct, inhabit hollow trees (Lindenmayer, 2002). Studies have shown in some forests there is likely
to be a shortage of hollows available for colonisation. (Goldingay, 2009) (Gibbons and Lindenmayer,
2010). Dead standing Eucalyptuses are more likely to be aged and have hollows, therefore automated
detection of them plays a significant role in protecting those animals.
The full-waveform (FW) LiDAR data used for this project are supplied by RPS Australia East Pty Ltd and
they were collected in March 2015 using the Trimble AX60 Airborne LiDAR sensor. In addition, the field
plots used for the classifications are provided by Forestry Corporation of NSW and contain around 1000
Eucalyptuses of which 10% are dead.
The standard method of using the higher resolution of full-waveform LIDAR would be to generate and
extract a denser point cloud (Neuenschwander, 2009) (Reitberger, 2006). While this identifies significant
features as points, the full waveform data also contains information in single shots that may be below
the significance threshold. These data can be accumulated from multiple shots into a voxelised volume,
e.g. Persson et al, 2005, who used voxelisation to visualise the waveforms. We have implemented and
improved on the voxelisation approach in DASOS, an open source software for managing FW LiDAR
data (Miltiadou et al, 2015). DASOS further normalises the intensities so that equal pulse length exists
inside each voxel, making intensities more meaningful and it exports forestry metrics (i.e. intensity
distribution and height differences within a radius relevant to canopy heights) into a single vector for
fast interpretation in advanced statistical tools like R and Matlab.
Addressing the dead Eucalyptus case, we have generated 3D priors characterizing dead standing
Eucalyptuses. These 3D dead standing tree priors are run over the voxelised FW LiDAR data using the
Random Forest and the Nearest Neighbour algorithms in order to detect the locations of candidate
trees. A comparison between the LiDAR point cloud and the voxelised FW LiDAR data is further
performed to demonstrate the increased survey accuracy obtained with the voxelisation.
34
Universidad Mayor
ForestSAT 2016 Abstracts Summary
References:
Gibbons, P. & Lindenmayer, D. (2002), Tree Hollows
and Wildlife Conservation in Australia, CSIRO
Publishing.
Goldingay, R. L. (2009), ‘Characteristics of tree
hollows used by australian birds and bats’,
Wildlife Research 36(5), 394–409.
Lindenmayer, D. B. & Wood, J. T. (2010), ‘Longterm patterns in the decay, collapse, and
abundance of trees with hollows in the mountain
ash (eucalyptus regnans) forests of victoria,
southeastern australia’, Canadian Journal of
Forest Research 40(1), 48–54.
Miltiadou, M., Warren, M. A., Grant, M. & Brown,
M. (2015), ‘Alignment of hyperspectral imagery
and full-waveform lidar data for visualisation
and classification purposes’, The International
Archives of Photogrammetry, Remote Sensing
and Spatial Information Sciences 40(7), 1257.
Neuenschwander, A., Magruder, L. & Tyler, M. (n.d.),
‘Landcover classification of small-footprint fullwaveform lidar data’, Jounal of Applied Remote
Sensing 3(1), 033544–033544.
Persson, A., Soderman, U., Topel, J. & Ahlberg,
S. (2005), Visualisation and Analysis of fullwaveform airborne laser scanner data, V/3
Workshop, Laser scanning 2005.
Reitberger, J., Krzystek, P. & Stilla, U. (2008),
‘Analysis of full waveform LiDAR data for tree
species classification’, International Journal of
Remote Sensing 29(5), 1407–1431.
35
Universidad Mayor
Development of Methods
Detection of dead trees by machine learning techniques in
temperate forests of the Bavarian Forest National Park
Peter Krzystek, Przemyslaw Polewski
Dept. Of Geoinformatics, Munich University of Applied Sciences, 80333 Munich, Germany Keywords:
Dead tree, machine lerrning, segmentation, classification The study highlights experiments conducted
in the Bavarian Forest National Park for the areawide extraction of standing and fallen trees from ALS
LiDAR data and aerial imagery. Full waveform LiDAR data were captured in leaf-on situation with
the Riegl LMS Q-860i scanner at a point density of 30 points/m2. Multispectral airborne data were
acquired with the DMC camera at a GSD of 20 cm. ALS 3D points clouds were obtained by a waveform
decomposition using a sum of Gaussian functions. Clusters for single standing trees are found by a
wall-towall technique for single 3D segmentation in advance. Single dead tree trunks without a crown
known as snags are captured just using the segmented tree point cloud. A MSAC-based line fitting
preselects the best samples which are subsequently classified by kernelized logistic regression using
free shape contexts (FSC) as features. FSCs are 3D shape descriptors which quantify the distribution of
points around the principal axis. Standing dead trees with existing crowns are identified from images
with green, blue and infrared channels as spectral bands. The 3D clusters from the tree segmentation
are used at object level to find the corresponding enclosing polygon in the image. The means of the 3
channels and 6 independent elements of the 3x3 channel covariance matrix are used in the kernelized
logistic regression to classify the dead trees. For both types of dead trees a training set of 100 samples
is selected. The two methods are combined with a novel active and supervised learning technique to
achieve best classification results at a minimum of training costs. The active learning using a Renyi
entropy regularized expected error reduction (EER) achieves 90 % of the final classification performance
using around 60 % of the number of queries required by standard EER. If the full training set is used the
overall accuracy of the dead tree detection is 87 % and 88 %, resp. Finally, a new approach for fallen
dead trees detection is based on the idea to first segment the fallen stems in parts and to subsequently
combine them into the entire stems. First, potential stem points are identified using point descriptors
known from computer vision and a logistic regression. Then, features are calculated from the stem
points by shape descriptors and are utilized in a logistic regression to estimate potential stem
segments. To this end, the segments are merged to entire stems by a normalized cut segmentation
whose control parameters and classifier-based stopping criterion are learned from a simulated virtual
fallen tree scenario. In case of beeches, the correctness and completeness of the method amounts at
average to 85 % and 79 %, resp. For spruces we end up at 86 % and 57 %.
The paper demonstrates that techniques from machine learning and computer vision are the key to
an object-based segmentation and classification of forest parameters. The presented methods have
been successfully applied to map dead trees on a large scale forest area (ca. 300 km2).
36
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Development of an energy balance model for estimating
canopy stomatal conductance from airborne thermal data
Juan Suárez, Georgios Xenakis and Roberto Antolín.
Centre for Sustainable Forestry and Climate Change, Forest Research, Northern Research Station, Roslin,
Midlothian EH25 9SY, UK.
Keywords: Thermal, LiDAR, Tree physiology, Windflow modelling, Thermolidar, QGIS
Declining tree health is a concern due to the impact
of climate change and tree diseases that have
profound impacts on commercial activities and the
ecological quality of the ecosystems. In this context,
the early detection of signs of physiological stress
in trees becomes increasingly important in order to
plan management interventions. Forest monitoring
over large areas is labour intensive in the field and
earth observation is frequently restricted to the
detection of late-stage visible symptoms. However,
sensors have the potential to early detect some
subtle changes in leaf biochemistry and the other
processes linked to the physiological activity of the
foliage. One of those processes is the alterations in
stomata activity.
For many years, scores of researchers have used
thermal sensors to measure temperature differences
used as a proxy of constrained stomatal conductance.
Stomatal closure is influenced by the water
condition of the plant, modifying energy dissipation
as heat. Therefore, a reduce rate of gas exchange
of the foliage can be detected as increasing canopy
temperatures. Nevertheless, this methodology
is constrained by to the lack of sensors at higher
spatial resolutions (<1-2 m), the aggregation of
combined thermal responses (shadows, ground
features and other plant constituent elements) and
the air boundary layer above the foliage. All these
elements combined increase the noise to signal ratio
for an effective detection of thermal differences that
can be associated to stomata activity.
This work has reformulated the energy balance
model primarily suggested by Blonquist et al (2009)
that now estimates canopy stomatal conductance
by using a series of standard meteorological and
airborne thermal data. The model calculates canopy
stomatal conductance (gC , mol m−2 ground area
s−1) from canopy temperature, air temperature,
barometric pressure, relative humidity, net
radiation, wind speed and plant canopy height.
Canopy temperature has been measured using a
thermal sensor on a plane platform in a study area
in the Trossachs and Ben Lomond National Park of
Scotland. Measurements took place at the time
of maximum stomata activity at 10 a.m. in May,
July and September and calibrated with field data.
Wind speed has been calculated using reference
anemometers and the airflow model WAsP.
The meteorological variables (air temperature,
barometric pressure and relative humidity) have
been measured in two stations in the field closed to
the monitoring plots used as validation. Vegetation
height was extracted from airborne LiDAR. Incoming
shortwave radiation at the top of the canopy has
been calculated as a function of latitude, elevation,
slope and aspect, monthly air temperature, relative
humidity and total precipitation to calculate the
solar radiation first at the top of the atmosphere,
then corrected it for atmospheric conditions and
finally corrected it for topographic exposure and
inclination. Net assimilation has been estimated
from incoming solar radiation and a constant value
obtained from the literature.
The model has been implemented in QGIS using a
combination of C++ and Phython. This development
has been part of the EU-funded project Thermolidar.
The outputs of the model were validated against field
measurements by measuring stomatal conductance
on a number of trees at the time of thermal data
collection. The canopy of each sample tree was
divided in several layers (approximately three levels)
and at least four samples (two sunlit and two shaded
37
Development of Methods
branches). In each, leaf stomatal conductance was
measured using an infrared gas analyser (IRGA,
LiCOR 6400). The results showed the potential of
this methodology for creating thematic cartography
showing the distribution of stomatal conductance
across a forested area. More work is being underway
at the time of the submission of this abstract that
will be presented during the conference.
38
Universidad Mayor
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Development of precision forestry applications for New
Zealand plantations using remotely sensed datasets
Michael S Watt1*, Jonathan P Dash2, Grant Pearse2, David Pont2, Heidi Dungey2, Mark Kimberley2, Marie Heaphy2
Scion, PO Box 29237, Fendalton, Christchurch, New Zealand
2
Scion, PO Box 3020, Rotorua, New Zealand
1
Keywords: ALS; light detection and ranging; Pinus radiata; radiata pine; stocking
Precision forestry is the deployment of high resolution remotely sensed data to support decision making. Remotely sensed information is now at a cost
and availability that allows precision forestry to become a reality within New Zealand plantations. Implementation of precision forestry will provide the
necessary information for managers to make decisions that optimise plantation growth and value at
a fine (sub-stand to stand) scale across broad spatial
extent. The objective of this research is to describe
current and potential precision forestry applications
within New Zealand plantations.
Using LiDAR and satellite imagery as input data the
following precision forestry applications are likely
to be implemented within New Zealand plantation
forestry operations over the next decade:
● Spatial optimisation of final crop stand density
using a model based on the productivity indices
Site Index and 300 Index, which in turn are
derived from remotely sensed information (GIS
layers, satellite imagery; LiDAR)
● Characterisation of weed competition in young
plantation stands, using multispectral data
(acquired from satellite imagery or UAV’s) to
guide decision making around weed control
using aerially applied herbicides
● Regional monitoring of key needle cast diseases
at a fine scale using satellite imagery allowing
targeted disease control measures
● Area based phenotyping where LiDAR data can
be used spatially quantify genetic gain. This
information can be used to precisely identity
the optimal location of individual genotypes
throughout a forestry estate to optimise growth,
disease resistance and wood quality
● Determination of plantation areas where fertiliser
application is likely to result in the highest growth
gains, based on estimates of predicted leaf area
index derived from LiDAR
At the time of writing a model describing spatial
optimisation of final crop density has been developed
and is currently being operationalised using remotely
sensed data as input. Remotely sensed multispectral
imagery has been successfully used to characterise
weed competition and the disease red needle cast
(second and third listed applications). Research
assessing the utility of area based phenotyping across
a major plantation estate is currently underway
and expected to be completed within a year. LiDAR
data has been successfully used to describe leaf area
index and the utility of this layer for predicting areas
where fertilisation will be most effective is currently
being assessed across a major plantation estate. This
research will be completed within a year.
Datasets that support precision forestry are
becoming increasingly affordable within New
Zealand. Free regional LiDAR datasets are available
within New Zealand and the cost of airborne LiDAR
has over recent years declined to a level that makes
acquisition cost effective for most major plantation
growers. Recent satellite launches (e.g. Sentinel 2)
that provide freely available imagery on a regular
basis are also providing a useful source of information
for precision forestry applications.
39
Development of Methods
Universidad Mayor
Estimating the regional resource supply of forests in southwest Germany for a future lignocellulose-based bioeconomy
using airborne LiDAR, Landsat 7 and National Forest
Inventory data
Joachim Maack1, Marcus Lingenfelder2, Thomas Smaltschinski2, Dirk Jaeger2, Barbara Koch1
University of Freiburg, Chair of Remote Sensing and Landscape Information Systems (FeLis), Germany
2
University of Freiburg, Chair of Forest Operations, Germany
1
Keywords: LiDAR, Landsat 7, National Forest Inventory, Timber Volume Modelling, Bioeconomy
Abstract
In the endeavour to move towards sustainable development, one essential goal is to replace fossil-fuelbased materials with bio-based products derived from algae, crops and wood. There is a strong interest
in estimating biomass or timber volume because to assess the prospects of a future bioeconomy it
is important to calculate the potential supply of resources. Remote sensing-based timber volume
estimation is essential to modelling the regional potential, accessibility and price of lignocellulosic
raw material for an emerging bioeconomy. We used a unique wall-to-wall airborne LiDAR dataset and
Landsat 7 satellite images in combination with terrestrial inventory data derived from the German
National Forest Inventory (NFI), and applied generalized additive models (GAM) to estimate spatially
explicit timber distribution and volume in the federal state of Baden-Württemberg (~35000 km2).
The NFI data showed an underlying structure regarding size and ownership of the different wooded
areas that reflect various management impacts in terms of cultivation, timber stocks and logging. We
further correlated the estimated timber volume with the annual regional timber harvest for different
forest sizes and ownership classes to analyze the regional economic potential. In the final step we
will propose optimized locations for possible biomass conversion plants with varying catchment areas
using routing-enabled infrastructure data. As it is still largely unknown which size, throughput and
resource demand a conversion-plant will feature, we developed an adaptive workflow that allows
for future adjustments of the approach. The results demonstrate the usefulness of remote sensing
techniques for mapping timber volume towards a future lignocellulose-based bioeconomy.
40
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Estimation of canopy cover in hemi-boreal broad-leaved
forests in Estonia using hemispherical photography
and lidar data
Tauri Arumäe 1,2*, Mait Lang 1,3, Ave Kodar 3
Institute of Forestry and Rural Engineering, Estonian University of Life Sciences, Kreutzwaldi 5,
51014, Tartu, Estonia; *e-mail: tauri.arumae@rmk.ee;
2
State Forest Management Centre, 10149, Tallinn, Estonia;
3
Tartu Observatory, 61602, Tõravere, Tartumaa, Estonia;
1
Keywords: airborne laser data, canopy cover, living crown-base height, leaf-on and leaf-off canopy cover
Introduction & aim
Estonia has a country-wide regular airborne laser
scanning cycle and with two years the whole country
is covered with lidar data. By the end of 2015 the
whole country had a repetitive layer of airborne lidar
coverage. As approximately 50% of land in Estonia
is covered with forests, then the understanding
and application of lidar data in practical forest
management is essential.
leaved forests (Appendix 1), mainly dominated by
Silver birch (Betula pendula Roth), Norway spruce
(Picea Abies L.), Trembling aspen (Populus tremula
L.) and Black Alder (Alnus Glutinosa L.). The forests
are mostly multi-layered with Norway spruce in the
second layer.
The aim of the study was to estimate canopy cover
(CC) in dense broadleaved hemi-boreal forests in
Estonia using airborne lidar data. The estimates
were validated by using hemispherical photos for
validation data. CC is a key parameter for estimating
standing volume, biomass and leaf area index. We
tested different height breaks up to living crown
base height and first or all reflections for calculating
CC using ALS.
Materials and methods
The test site is located in south-eastern part of
Estonia, near Laeva (Figure 1). The test site was
15x15 km and mostly covered (>50%) with broad-
Figure 1. Test site location.
41
Development of Methods
Universidad Mayor
Forest-Observation-System.net – towards global reference
database for forest biomass
Dmitry Schepaschenko (IIASA, Laxenburg, Austria), Jerome Chave (CNRS, Toulouse, France), Oliver L. Phillips (University
of Leeds, UK), Stuart J. Davies (STRI, Washington, USA), Steffen Fritz (IIASA, Laxenburg, Austria), Christoph Perger
(IIASA, Laxenburg, Austria), Christopher Dresel (IIASA, Laxenburg, Austria), Simon Lewis (University of Leeds, UK), Klaus
Scipal (ESA, Netherlands)
Keywords: Forest biomass, remote sensing, permanent plots, calibration and validation, BIOMASS mission
The Forest Observation System (FOS) is an
international cooperation implemented at the
initiative of the ESA BIOMASS mission, to establish
and maintain a global in-situ forest biomass
database to support earth observation and to
encourage investment in relevant field-based
science. It is designed to be able eventually fulfil the
ground data requirements for algorithm training and
validation of spaceborne forest related missions, i.e.
ESA Sentinel-1, BIOMASS and SAOCOM-CS; NASA
GEDI and NISAR; JAXA PALSAR, etc.
The FOS serves as an interface between the remote
sensing and ecological communities. Data sharing
nowadays is one of the biggest problem despite
of the fact that everyone can benefit. Ecologists
sometimes do not realize how important their data
are for calibration/validation of remote sensing
products. Remote sensing community can provide
additional rationality and arguments for investments
in measurements on sample plots.
The implementation of the FOS is guided by four
principles. First, the ground data should be of high
quality and collected on permanent plots from
42
0.25 ha upwards by size. Second, the selection of
the sites should be realistic, i.e. proposed at sites
where previous expertise and capacity have been
built. Third, sites should cover a broad range of
geographical and environmental conditions, so as
to maximize the robustness of models. Fourth, the
procedures for ground data acquisition and database
compilation should be transparent and proofed
extensively.
Project web portal (http://forest-observationsystem.net/) presents besides several base maps,
two types of data: (1) metadata: where and what
were measured on permanent sample plots; (2)
sample plot data for subset of plots where authors
agreed to share the data: aggregated to 50x50 m
live biomass, tree height, wood density and tree
composition. In the proof-of-concept phrase, FOS
includes the Center for Tropical Forest Science
(CTFS-ForestGEO), RAINFOR, AfriTRON and IIASA
networks. FOS is an open initiative and we expect
more networks to join for common benefits, incl.
joint publications and application for funding of
fieldwork.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Fractal Volumetric Bouligand-Minkowski Classification of
Forest Trees
João Paulo Herrera - EMBRAPA Instrumentação
João do Espírito Santo Batista Neto – Universidade de São Paulo
Lúcio André de Castro Jorge – EMBRAPA Instrumentação
Keywords: Texture, descriptor, fractal, Bolingand-Minkowski, tropical, forests, fourier, classification
Abstract:
The fractal theory is a well-known technique for representing complex structures of nature. In this
work, a recent fractal texture descriptor is evaluated in aerial images of tropical forests obtained by an
unmanned aerial vehicle. The first results show that the method is able to efficiently characterize some
areas. However, additional studies are needed to enhance its performance.
Introduction
Texture is an important physical and visual attribute
and is intrinsically involved with properties that
define a surface [1]. In Computer Vision, this attribute
is defined by the distribution of image reflectance
indices and is widely used to characterize a particular
region of interest.
There are several models of texture extraction.
Recently, researchers of University of São Paulo
developed a new texture descriptor based on the
Fractal model [2] which is widely known by its good
representation of high complex structures and
is generally associated with objects of nature. In
this context, others authors [3, 4] investigated the
performance of this method in different vegetation
databases. However, these studies focus only on
the analysis of the micro texture on the leaf surface
of the plants. In addition, the authors used a very
high statistical accuracy for the selection of the
samples. In this work, the investigation is based
on the analysis of the performance of volumetric
Bouligand-Minkowski in aerial images of Brazilian
tropical forests. The automatic recognition will
enable the mapping of preservation zones and
provide new tools for monitoring practices.
Volumetric Bouligand-Minkowski
Theoretically, a fractal is made by infinite selfsimilar fragments and its characterization is given
by fractal dimension measure [5]. In practice, the
physical limits of the fractal objects make difficult
to compute its exact value of fractal dimension, so it
must be estimated.
The volumetric Bouligand-Minkowski method is
an estimator of the fractal dimension of textures.
Basically, a grey level image is transformed into a
three dimensional surface so that the coordinates
and are preserved and refers to the grey intensities
values. Then each voxel of the surface is successively
dilated by a sphere until its radius size reaches a predetermined value, as shown in Figure 1. At each step
the total volume of the surface is computed. The
fractal dimension value is given by:
such that:
43
Universidad Mayor
Development of Methods
Figure 1 – Dilatation process of the Volumetric Bouligand-Minkowski method. (a) Refers to a grey-level
image.(b) Dilatation of with . (c) Dilatation of with. (d) Dilatation of with. Source: Florindo et al. (2014)
where (vx, vy, vz) ∈ S a point of interest on the surface
and the limit is the angular coefficient of the line
fitted to the log-log curve.
Methodology
The experiment consisted in take aerial images of
tropical forests located in São Carlos city, Brazil.
The images were captured by a SONY ILCE-5100
camera docked on a fixed wing UAV model ECHAR
20B, which was managed to work at approximately
300 meters in height. The characteristics extraction
was performed locally once the image has different
textured regions. Thus, each side of the image was
divided into 32 equal parts, resulting in a grid of 1024
textured regions.
The volumetric Bouligand-Minkowski extraction was
based on the methodology suggested by [2]. They
propose to create a three-dimensional Euclidean
distance map of the surface in order to facilitate
the dilation process of the spheres. Thus, it is not
necessary to centralize them in each voxel of, only
to count the number of values smaller or equal to its
radius. In addition, [3] showed that only the value
of the fractal dimension is not enough to describe a
texture. This is because different textures can have
the same fractal dimension [6]. Therefore, a vector
of characteristics was used as a texture signature.
This vector is composed by the values of for each
increment of the distance value to the limit which in
this case was defined as 30:
44
In addition to the proposed approach, four other
methods were used: Grey-Level Co-occurrence
Matrices (GLCM), Local Binary Patterns (LBP), Local
Binary Patterns Rotation Invariation Uniform () and
Fourier Descriptors.
At the training stage, 510 tree samples and 490
samples from other areas, such as grass, road, soil
and rocks were selected and their textures were
extracted by the five descriptors.
In the classification stage, 4 images with 1024
textured regions were classified using the Linear
Discriminant Analysis (LDA) classifier.
Preliminary Results
The first results obtained from the 4 images are
presented in Table 1. The best performance was the
Fourier Descriptors, with global accuracy of 95.7%
and kappa = 0.94 and the worst was the BouligandMinkowski volumetric, with overall accuracy of 87.1
% and kappa = 0.84. The LDA and SVM classifiers
were used to calculate the kappa coefficient.
The preliminary results of the volumetric BouligandMinkowski technique suggest further studies to
leverage its discriminative power in textures of
tropical forests distributed over aerial images.
Finding the optimal value of is a fundamental key
to a better performance. Furthermore, the time
and computational resources consumed were high.
Figures 2 and 3 illustrate some regions classified by
the method.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Table 1 – Classification accuracies
Descriptor
Bouligand-Minkowski
GLCM
LBP
LBPriu2
Fourier
Overall Accuracy
87,1%
94,7%
93,1%
89,0%
95,7%
Kappa
0.84
0.93
0.85
0.93
0.94
Figure 2 – Examples of good segmentation using the proposed method
Conclusões
In this work, the texture of tropical forests in aerial
images was extracted through the recent volumetric
Bouligand-Minkowski fractal descriptor. Despite the
good representation that a fractal has on structures
of nature, the preliminary results did not show good
performance against the other texture descriptors
already diffused in the literature. Further studies are
needed in order to find the maximum radius of the
structuring element ideal for this type of texture.
45
Universidad Mayor
Development of Methods
Figure 3 – Examples of poor segmentation using the proposed method
References:
1. Petrou, M; Sevilla, P.G - “Image Processing,
dealing with texture”, John Wiley and Sons, 2006.
2. Backes, A. R; Casanova, D; Bruno, O. M - “Plant
leaf identification based on volumetric fractal
dimension”, 2009.
3. Florindo, J. B; Da Silva, N. R; Romualdo, L. M; Da
Silva, F. F; Luz, P. H. de C; Herling, V. R; Bruno,
O. M – “Brachiaria species identification using
imaging techniques based on fractal descriptors”,
Computers and Electronics in Agriculture, 2014.
4. Florindo, J. B; De Castro, M; Bruno, O. M “Enhancing Volumetric Bouligand-Minkowski
Fractal Descriptors by using Functional Data
Analysis”, International Journal of Modern
Physics C, 22:9, 2012.
46
5. Mandelbrot, B. - “How long is the coast of
Britain? Statistical self-similarity and fractional
dimension”, Science, New series, vol. 156, No
3775, 1967, pp. 636-638.
6. Florindo, J. B; Bruno, O. M – “Descritores fractais
aplicados à análise de texturas”, Universidade de
São Paulo, 2013
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Full waveform LiDAR and the new PulseWaves format
Abstract:
The earliest airborne Light Detection And Ranging (LiDAR) scanners delivered just a single elevation
return per laser shot. Later systems produced a return for the first and the last interaction between
the laser and the landscape below. Today, discrete return LiDAR commonly produces up to 4 or more
returns. This allows picking up hits of electricity or telephone wires and captures more information
about the vegetation structure.
Recently, waveform digitizers have become popular that capture the reflection of the emitted laser
pulse with much more detail: The waveform returning to the plane is digitized up to one billion times
per second so that the intensity of the laser pulse’s reflection is recorded every nanosecond giving
a vertical resolution of one sample of digitized amplitude each 15 centimeters for the returning
waveform.
Since 2012 rapidlasso GmbH has been working with the LiDAR community and hardware vendors to
create and support the PulseWaves format - a new data exchange format that is similar to the popular
LAS format but aimed at storing the entire digitized waveform instead of discrete LiDAR returns in a
fully georeferenced manner. In particular it allows storing the outgoing waveform that was “shot” at
the aircraft in addition to one or multiple samplings of the returning waveform.
In this talk we will provide a quick intuitive look at how full waveform LiDAR is different from discrete
LiDAR, introduce the key features of the PulseWaves format, how to obtain data in this format, and
how to use the existing API to read and process it. We look at the current set of PulseTools that are
already available for processing full waveform data in PulseWaves (and other full waveform formats),
summarize scenarios where operating directly on the full waveform already lead to superior results,
and conclude with an outlook on what other full wabeform exploits may lie ahead.
47
Universidad Mayor
Development of Methods
Isolation of obscured forest tree stems using TLS data
Johannes Heinzel
Swiss Federal Institute for Forest, Snow and Landscape Research WSL, Zürcherstraße 111, 8903 Birmensdorf, Switzerland
Keywords: Tree stems, Obstruction, Terrestrial laser scanning, 3D images, 3D mathematical morphology
Abstract
The presented study aims at the detection of obscured tree stems, which often appear hidden by
leaves and branches of the regeneration layer. We make use of the high information density in the
lower forest layers achieved by terrestrial laser scanning (TLS). The method builds on 3D volumetric
images and requires the original point vector data to be transferred into a regular volumetric grid. The
actual object detection workflow consists of two major steps. The first one applies a generalized threedimensional template of a tree stem to probe the complete dataset for vertically elongated structures.
The second step comprises a logical processing chain, which combines split stem segments and closes
gaps between segments belonging to the same tree stem.
Introduction
The extraction of forest tree stems from laser
scanning data provides an important basis for
deriving forest stand characteristics. Most existing
approaches are limited to the extraction of distinct
stems from mature trees, which are not obscured
by other vegetation [Eysn et al., 2013; Lovell et al.,
2011; Reitberger et al., 2007]. Others reconstruct
stems from completely isolated trees, which are
separated from any forest environment [Hosoi et
al., 2013; Raumonen et al., 2013]. The latter studies
concentrate on stem details and less on the detection
of stems by number and location within a forest
environment. Contrary to those approaches, in many
naturally regenerated forests the understory forms
a dense and unordered structure. It is characterized
by leaves and branches obscuring the tree stems of
both mature trees and those of the regeneration
layer. Only few studies attend to these difficulties,
but still show unfavourable restrictions [Brolly et
al., 2013; Xia et al., 2015]. From the perspectives of
the scanner positions, stems are often not visible in
their total extend and only parts or small fragments
can be captured. The challenge is to separate those
48
fragments from the surrounding objects and to
combine all segments belonging to the same tree
stem.
Methodology
Detecting the tree stems in dense and obscured
forest environments is subject to the present
method. A special characteristic of the presented
method is that it aims at the stem extraction from
both mature and regeneration trees within a single
dataset. While with airborne laser scanning (ALS)
data it is difficult to obtain enough reflections from
the lower forest layers, terrestrial laser scanners
(TLS) produce a much higher information density due
to the different perspective. This results in a better
data coverage of the tree stems. However, TLS data
is not suitable for capturing continuous wide area
information as ALS does. Instead, TLS is often used
on a sample plot basis for forestry applications and
possibly can be combined with area wide data.
Making use of the perspective advantages of TLS,
data has been captured on nine spatially independent
plots, located in the canton of Grisons, Switzerland.
The plots resemble diverse and often dense forest
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Figure 1. The illustration shows the original vegetation volume and the extracted tree stems. Part a)
depicts the partly transparent volume grid of the complete vegetation dataset with obscured stems in the
background. In b) green vegetation and branches are removed and only isolated stems remain.
conditions, including different numbers of structural
layers. The quadratic plots are centred on grid nodes
used by the forest inventory and have a dimension
of 25 m edge length. On each plot five scans from
different positions have been captured.
Processing of the data covers pre-processing and
subsequently two major steps of image analysis. The
pre-processing includes the registration of the single
scans, the three-dimensional (3D) thinning of the
point cloud and the normalization by terrain elevation.
The most important part of the pre-processing is the
transfer of the point cloud data into a volume grid. All
further processing builds on the advantages of the
ordered structure of 3D grids. The first major step for
stem detection uses techniques from mathematical
morphology to separate non-stem vegetation parts
from tree stem objects. We apply a series of structuring
elements for spatially variant and directional
openings of the original volume data. The results are
combined into a binary volume, including only stem
candidate segments. The second step combines
labelling techniques with a logical processing chain
in order to aggregate all segments, which belong to a
single stem. Finally we receive multiple 3D connected
components for all tree stems within a plot. Figure 1
shows the original vegetation data in comparison to
the isolated tree stems.
Results
position of the stems against the reference. For
mature trees we verified all stems as measured by
the forest inventory, while for regeneration trees
only randomly selected individuals were available.
Mature trees with a diameter at breast height
(DBH) of more than 12 cm show an almost complete
detection rate of 97.4 %. Stems from regeneration
trees with a minimum crown top height of more
than 130 cm and a DBH of less than 12 cm show an
overall accuracy of 84.6 %.
Conclusions
In conclusion, the presented method allows practical
and reliable extraction of tree stems from mixed
mature and regeneration trees. The method is
transferable to diverse and dense forest structures
and could support applications, such like the
management of protective forest against natural
hazards, forest inventory or regeneration mapping.
Acknowledgments: This study was conducted
within the framework of the Swiss National Forest
Inventory. The author would like to thank Markus
Huber for project management, Natalia Rehush
and Björn Dreier for support during fieldwork,
Christian Ginzler for expertise in GNSS technology
and Susette Haegi for scan registration. The forest
administration of the canton Grisons, Switzerland,
gave important assistance to this project.
Automatically derived tree stem objects are verified
for all nine plots by comparing the number and
49
Development of Methods
References
Brolly Gá., Király Gé., Czimber K. (2013) - Mapping
Forest Regeneration from Terrestrial Laser Scans.
Acta Silv. Lign. Hung., 9: 135-146. DOI: 10.2478/
aslh-2013-0011.
Eysn L., Pfeifer N., Ressl C., Hollaus M., Grafl A.,
Morsdorf F. (2013) - A Practical Approach for
Extracting Tree Models in Forest Environments
Based on Equirectangular Projections of
Terrestrial Laser Scans. Remote Sens., 5: 54245448. DOI: 10.3390/rs5115424.
Hosoi F., Nakai Y., Omasa K. (2013) - 3-D voxel-based
solid modeling of a broad-leaved tree for accurate
volume estimation using portable scanning lidar.
ISPRS J. Photogramm. Remote Sens., 82: 41-48.
DOI: 10.1016/j.isprsjprs.2013.04.011.
Lovell J., Jupp D., Newnham G., Culvenor D. (2011)
- Measuring tree stem diameters using intensity
profiles from ground-based scanning lidar
from a fixed viewpoint. ISPRS J. Photogramm.
Remote Sens., 66: 46-55. DOI: 10.1016/j.
isprsjprs.2010.08.006.
Raumonen P., Kaasalainen M., Åkerblom M.,
50
Universidad Mayor
Kaasalainen S., Kaartinen H., Vastaranta M.,
Holopainen M., Disney M., Lewis P. (2013) - Fast
Automatic Precision Tree Models from Terrestrial
Laser Scanner Data. Remote Sens., 5: 491-520.
DOI: 10.3390/rs5020491.
Reitberger J., Krzystek P., Stilla U. (2007) - Combined
tree segmentation and stem detection using full
waveform lidar data. ISPRS Workshop on Laser
Scanning 2007 and SilviLaser 2007, XXXVI-3/W52.
Espoo, Finland. pp. 332-337.
Xia S., Wang C., Pan F., Xi X., Zeng H., Liu H. (2015)
- Detecting stems in dense and homogeneous
forest using single-scan TLS. Forests, 6: 39233945. DOI: 10.3390/f6103923.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
LiDAR- and SAR-based mapping of structural attributes
of deciduous savannahs and woodlands in the southern
African region
Renaud Mathieu1,2, Russell Main1,2, Laven Naidoo1,2, Konrad Wessels3,2, Greg Asner4
1
Ecosystem Earth Observation, Natural Resources and the Environment, CSIR, Pretoria, South Africa
Department of Geography, Geoinformatics, and Meteorology, University of Pretoria, Pretoria, South Africa
3
Remote Sensing Research Unit, Meraka Institute, CSIR, Pretoria, South Africa
4
Department of Global Ecology, Stanford University, Stanford, CA, USA
Corresponding author contact details: RMathieu@csir.co.za; (+27)12 841 4089
2
Keyword: multi-frequency SAR, optical, woody cover, above ground biomass
Abstract
Savannahs and woodlands are the largest forested biome in southern Africa, accounting for 35-45%
of the land. These are open forests, with a continuous grass layer and a discontinuous low biomass
woody layer (< 60T/ha). Excessive harvesting of woody plants and land use changes can threaten
the sustainability of the provision of grass, timber, fuelwood, edible fruits and roots to poorest rural
communities. On the other hand, bush encroachment (or thickening) is increasingly seen as a major
regional threat for food security via a reduction of grassland for meat production. In South Africa
tree cover in savannahs is believed to have increased at a rate of 5-6% per decade and to encroach in
grasslands; bush encroachment affects 10-20 million ha. Despite these drastic changes and the legal
requirement to report on the forest status every three years at national scale (Forest Act 1998, SA),
there is yet limited information on spatial patterns of woody vegetation. We assessed the use of multifrequency Synthetic Aperture Radar (SAR) and optical datasets acquired at various seasons to map
structural attributes (woody cover, above ground biomass) in southern African deciduous woodlands
and savannahs. Imagery included X-band TerraSAR, C-band RADARSAT-2 and ENVISAT ASAR, L-band
ALOS PALSAR, and Landsat. Methods were based on the integration of field data, airborne LiDAR,
and satellite imagery. Field plots were used to calibrate and validate extensive LiDAR-based maps
of structural metrics, which were then used to upscale the metrics at satellite level using random
forest. Results demonstrated that L-band data acquired in winter performs much better than any
other SAR frequency or Landsat optical data, acquired at any seasons. Structural metric retrievals can
be can improved by ca. 10% using the combination of L-band with C-band or Landsat reflectance,
but L-band remains the most important datasets to monitor woody resource in the region, including
for woody cover. Dense time series of winter C-band imagery were shown to be a suitable substitute
when L-band data are not available. In addition, the research led to the development of a LiDAR/
SAR processing platform, and the production of unique 1-ha maps of woody cover for the entire land
mass of South Africa and Namibia, totalling 2 million km2. The maps were produced using extensive
regional LiDAR datasets for local calibration/validation and the JAXA ALOS PALSAR mosaic. They are
a significant improvement on the global products which were until recently the only available datasets
in the region.
51
Universidad Mayor
Development of Methods
Linking remotely sensed functional diversity with
phylogenetic structure of a temperate forest
Carla Guillén Escribà1, Fabian D. Schneider1, Felix Morsdorf1, Andy Tedder2, Eri Yamasaki2, Kentaro K. Shimizu2, Bernhard
Schmid2, Michael E. Schaepman1
2.
1.
Department of Geography, Remote Sensing Laboratories, University of Zurich, Zurich, Switzerland.
Institute of Evolutionary Biology and Environmental Studies, Evolutionary and Ecological Genomics, University of
Zurich, Zurich, Switzerland.
Keywords: functional diversity, genetic diversity, imaging spectroscopy, laser scanning, intraspecific
variation.
Understanding the forces driving forest biodiversity
under global environmental change conditions is
an important goal for plant ecologists (Pereira et
al. 2012). One of the metrics that is commonly used
for the assessment of forest diversity is functional
diversity (FD), that is the diversity of functional
traits in a given community (Díaz and Cabido 2001;
Cadotte et al. 2011, Cardinale et al. 2012). Some
studies have found evidence of the strong link
between the FD and the phylogenetic composition
(PD) of a community (Petchey and Gaston, 2002a;
Cadotte et al. 2009). Other studies such as Flynn et
al. (2011) and Wang et al. (2015) have delved into
the understanding of the mechanistic link between
diversity, community functioning and provision of
services, and the mechanisms explaining patterns
of local FD and PD. To understand these interactions
detailed observations are needed. Traditional insitu approaches have provided very useful data, but
are usually spatially constrained (Duro et al. 2007),
hence exhibit limited possibilities to be extrapolated
to larger scales (Pereira et al. 2013). With the advent
of emerging remote sensing methods for functional
traits mapping at regional to global scales, the gap
of missing trait distribution at larger scales is to be
filled (Jetz et al., 2016). We propose to contribute
to filling this gap at regional scale by using remote
sensing data in combination with in-situ sampling at
different spatial and temporal scales.
Here, we investigate the relevance of remotely
sensed local interspecific functional variation of
a temperate forest (9-hectare plot located on
the south-facing slope of the Lägern mountain,
52
Switzerland) to differentiate vegetation types and
species at distant and close phylogenetic distances.
Phylogeny of the species was constructed using DNA
barcode data. Biochemical and architectural plant
traits were retrieved by using canopy spectra and
point clouds from airborne imaging spectroscopy
and laser scanning data respectively. Additionally,
trait responses along environmental gradients were
detected at intraspecific level to attribute sources of
functional diversity within species.
Preliminary results focus on the retrieval of different
functional traits. Figure 1 shows the potential of trait
clustering for vegetation types classification. Three
different biochemical plant traits have been used for
this: Chlorophyll, Carotenoid and Leaf water relative
abundance. Conifers such as Abies alba present
higher leaf water content values while deciduous
trees have higher pigment content. Figure 2 shows
inter- and intraspecific variation of 3 different
architectural traits. Preliminary results seem to
indicate thatfraction of single echoes could be one of
the architectural traits performing better vegetation
type classification and maximum height could be a
good predictor for species.
References
Cadotte, M. W., Carscadden, K. & Mirotchnick, N.
(2011). Beyond species: Functional diversity and
the maintenance of ecological processes and
services. J.Appl.Ecol., 48, 1079-1087.
Cadotte, M.W., Cavender-Bares, J., Tilman, D.
& Oakley, T.H. (2009). Using phylogenetic,
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Figure 1. (a) Individual tree crowns at the test site coloured by species and (b) by abundance in
relative units [0-1%] of three biochemical functional traits (red=carotenoid, green= chlorophyll
and blue= leaf water).
Figure 2. Boxplots of intra- and
interspecific architectural trait variation
for each of the species at the test site.
Traits are: maximum height, median
intensity and fraction of single echoes.
functional and trait diversity to understand
patterns of plant community productivity. PLoS
ONE, 4, e5695
Cardinale BJ., Duffy JE., Gonzalez A., Hooper DU.,
Perrings C., Venail P., Narwani A., Mace GM.,
Tilman D., Wardle DA. & Kinzig AP. (2012).
Biodiversity loss and its impact on humanity.
Nature. Jun 7, 486, no. 7401: 59-67.
Diaz, S. and Cabido, M. (2001). Vive la difference:
plant functional diversity matters to ecosystem
processes. Trends Ecol. Evol. 18: 646- 655
Duro, D., Coops, N., Wulder, M., and Han, Tian.,
(2007). Development of a large area biodiversity
monitoring system driven by remote sensing.
Progress in Physical Geography, 31: 2
Flynn, D.F.B., Mirotchnick, N., Jain, M., Palmer, M.I.
& Naeem, S. (2011). Functional and phylogenetic
diversity as predictors of biodiversity-ecosystemfunction relationships. Ecology 92, 1573 – 1581.
Jetz, W., Cavender-Bares, J., Schimel, D., Pavlik, R.,
Davis, F., Asner, G.P., Guralnick, R., Kattge, J.,
Latimer, A.M., Moorcroft, P., Schaepman, M.E.,
Schildhauer, M.P., Schneider, F.D., Schrodt, F.,
Ustin S.L., & Turner W. (2016). A global remote
sensing mission to detect and predict plant
functional biodiversity change. Nature Plants
Pereira, H.M., Ferrier, S., Walters, M., Geller, G.N.,
Jongman, R.H.G., Scholes, R.J., Bruford, M.W.,
Brummitt, N., Butchart, S.H.M., Cardoso, A.C.,
Coops, N.C., Dulloo, E., Faith, D.P., Freyhof,
J., Gregory, R.D., Heip, C., Höft, R., Hurtt, G.,
53
Development of Methods
Jetz, W., Karp, D., McGeoch, M.A., Obura, D.,
Onoda, Y., Pettorelli, N., Reyers, B., Sayre, R.,
Scharlemann, J.P.W., Stuart, S.N., Turak, E.,
Walpole, M. and Wegmann, M. (2013). Essential
Biodiversity Variables. Science, 339:277-278
Pereira, H.M., Navarro, L.M., Martins, I.S. (2012).
Global biodiversity change: the bad, the good,
and the unknown. Annu. Rev. Environ. Resour. 37,
25–50
Petchey, O. L. and Gaston, K. J. (2002). Functional
diversity (FD), species richness and community
composition. Ecology Letters, 5: 402–411.
Wang X., Wiegand T., Swenson N. G., Wolf A. T., Howe
R. W., Hao Z., et al. (2015). Mechanisms underlying
local functional and phylogenetic beta diversity
in two temperate forests. Ecology96:1062–1073.
54
Universidad Mayor
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Local pivotal method sampling design combined with micro
stands utilizing airborne laser scanning data in a long term
forest management planning setting
Rami Saad, Jörgen Wallerman, Johan Holmgren and Tomas Lämås
Keywords: local pivotal method (LPM); segmentation; most similar neighbor (MSN) imputation; forest
management planning; suboptimal loss; Lidar; Heureka; decision support system
Abstract
A new sampling design, the local pivotal method (LPM), was combined with the micro stand approach
and compared with the traditional systematic sampling design for estimation of forest stand variables.
The LPM uses the distance between units in an auxiliary space – in this case airborne laser scanning (ALS)
data – to obtain a well-spread sample. Two sets of reference plots were acquired by the two sampling
designs and used for imputing data to evaluation plots. The first set of reference plots, acquired by
LPM, made up four imputation alternatives (varying number of reference plots) and the second set
of reference plots, acquired by systematic sampling design, made up two alternatives (varying plot
radius). The forest variables in these alternatives were estimated using the nonparametric method of
most similar neighbor imputation, with the ALS data used as auxiliary data. The relative root mean
square error (RelRMSE), stem diameter distribution error index and suboptimal loss were calculated
for each alternative, but the results showed that neither sampling design, i.e. LPM vs. systematic,
offered clear advantages over the other. It is likely that the obtained results were a consequence of the
small evaluation dataset used in the study (n = 30). Nevertheless, the LPM sampling design combined
with the micro stand approach showed potential for improvement and might be a competitive method
when considering the cost efficiency.
55
Development of Methods
Universidad Mayor
Machine learning regression algorithms for biophysical
parameter retrieval from spectral properties to detect
different levels of ash vitality in Central Europe
Michael Foerster1, Nele Steinmetz1, Stephan Pflugmacher3, Kyle Pipkins1 and Anne Clasen1, 2
Geoinformation in Environmental Planning Lab, Technische Universität Berlin,
Straße des 17. Juni 145, D-10623 Berlin, Germany
2
Helmholz Centre Potsdam-GFZ German Research Centre for Geosciences, Telegrafenberg, 14473 Potsdam, Germany
3
Department of Ecological Impact Research and Ecotoxicology, Technische Universität Berlin, Ernst-Reuter-Platz 1,
D-10587 Berlin, Germany
1
Keywords: Support Vector Regression, Gaussian Process Regression, Ash dieback, tree mortality
Abstract
The north-East of Europe is affected by the ash (Fraxinus excelsior) dieback caused by the fungal
pathogen Hymenoscyphus pseudoalbidus. The presented study will show the relation between the
biophysical parameters of different levels of ash mortality and their spectral properties, taken from
ASD spectral measurements and a CIR aerial image. Just in this way, the biophysical relation of this
disease can be related to remote sensing properties.
The study was performed for three ash damage levels, evaluated by forest experts of the federal
forestry institution. For each of the three damage levels 40 samples were taken. Close range
spectral measurements of the canopy and single leaves were taken at the at the 17th and 18th of July
2014 with an ASD FieldSpec instrument mounted to a crane measurement platform situated in a
temperate deciduous forest in North-East Germany. Simultaneously, in the 2014 field campaign,
data was collected by CIR airborne campaign. As biophysical variables for the three mortality levels,
measurements of Chlorophyll a/b, leaf water content, carotenoids were taken as well as SPAD values.
Machine learning regression algorithms (MLRA) have been introduced recently for biophysical
parameter retrieval with remote sensing. This study will utilize the Support Vector Regression (SVR)
and Gaussian Process Regression (GPR) to predict the biophysical parameters and relate the outcome
to the different mortality rates.
Both machine learning algorithms show clear relations to the spectral measurements as well as to the
CIR aerial imagery. For both tests of the data-sets, the r² of all parameters varies between 0.34 and
0.48 with the exception of leaf water contents, which provides limited predictive power. The relative
RMSE is especially low for Carotenoid (4.6 %) as well as the SPAD measurements (4.7 %).
A clear distribution of the different levels of damage can be seen in the scatterplots (see Fig. 1) and can
be transferred to map the biophysical parameters to larger areas. The MLRA algorithms were applied
to simulated EnMap and Sentinel-2 spectral values, derived by the ASD measurements. The transfer
functions could be used to show the level of damage applied on a real Sentinel-2 scene of the summer
2016.
56
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Fig. 1: Scatterplot of predicted versus
measured SPAD values for CIR data
yielded by SVR (1 = good status of ash
health; 2 = medium; 3 = bad)
57
Universidad Mayor
Development of Methods
POSTER
Mapping the 3D structure of a tropical rainforest using
terrestrial laser scanning – a quality assessment
Fabian D. Schneider, Daniel Kükenbrink, Michael E. Schaepman and Felix Morsdorf
Remote Sensing Laboratories, University of Zurich, Winterthurerstrasse 190, CH-8057, Zurich, Switzerland
Keywords: terrestrial laser scanning, lidar, forest structure, occlusion mapping, ray tracing
Terrestrial laser scanning (TLS) has emerged as a
state-of-the-art measurement technology of forest
canopy structure and the three-dimensional (3D)
distribution of plant material. The detailed and
accurate measurements of scatterer location and
abundance have been used for forest reconstruction,
forest monitoring, and modeling of radiative transfer
or biomass (e.g., Calders et al., 2015; Newnham et al.,
2015)we develop an approach to estimate AGB from
TLS data, which does not need any prior information
about allometry. We compare these estimates
against destructively harvested AGB estimates and
AGB derived from allometric equations. We also
evaluate tree parameters, diameter at breast height
(DBH. When measuring a forest plot with TLS, the
quality and completeness of the data is mainly
determined by the applied measurement setup. The
goal of reducing occlusion and reaching a complete
coverage among all vertical layers of the canopy
has to be traded against number of scan locations
and hence costly operation time. Occlusion has
been identified as a major source of uncertainty in
forest reconstruction (Béland, Baldocchi, Widlowski,
Fournier, & Verstraete, 2014)we investigate the
optimal voxel dimensions for estimating the spatial
distribution of within crown leaf area density. We
analyzed LiDAR measurements from two field sites,
located in Mali and in California, with trees having
different leaf sizes during periods with and without
leaves.We found that there is a range of voxel sizes,
which satisfy three important conditions. The first
condition is related to clumping and requires voxels
small enough to exclude large gaps between crowns
and branches. The second condition requires a voxel
size large enough for the conditions postulated by
the Poisson law to be valid, i.e., a turbid medium with
randomly positioned leaves. And, the third condition
relates to the appropriate voxel size to pinpoint the
location of those volumes within the canopy which
58
were insufficiently sampled by the LiDAR instrument
to derive reliable statistics (occlusion effects, but
very few studies have specifically investigated the
effects of occlusion on TLS data quality.
In this study, we applied a ray tracing based method
developed by Kükenbrink, Leiterer, Schneider,
Schaepman, & Morsdorf (under review) to map
occlusion in a 1 ha forest plot. The method uses
a voxel-traversal algorithm and traces each laser
pulse to determine sampled and occluded, empty or
filled voxels. We scanned 1 ha of tropical rainforest
in the Lambir Hills National Park (Sarawak,
Malaysia) from 93 positions on the ground, with an
average horizontal spacing of 10 m. Additionally,
we performed 32 TLS scans from four platforms
of a canopy crane at 24, 39, 59 and 76 m above
ground. Our main research questions are: (1) How
is occlusion distributed in the vertical canopy layers
in (a) scans from the ground and (b) scans from the
canopy crane? (2) What is the benefit of combining
ground based and above-ground TLS measurements
for forest reconstruction? (3) How is occlusion
influenced by the number of scan positions?
Preliminary results are presented in Figure 1. The
voxel grid shows parts of the canopy crane on the
left and a 1 m deep transect through the surrounding
trees. It shows that most of the vegetated voxels
have only been covered by one or two scans.
Therefore, it is crucial to have many different scan
positions to cover trees in a tropical forest with their
full extent. Furthermore, it shows that parts of the
upper canopy are occluded in more than four out
of seven scans. This suggests that the scans from
the canopy crane are the main contribution to the
coverage and quality of the data in the uppermost
canopy layers. By extending the analysis to the
whole forest plot, we want to assess the quality of
Universidad Mayor
ground and above-ground TLS measurements in a
tropical rainforest and provide some guidelines to
establish a best performing measurement setup.
Figure 1. Vertical profile of a 10 cm voxel grid showing
voxels with laser returns and occluded voxels. The
color scale indicates the number of different scans
producing a laser return or occlusion. For this
example, seven scans have been co-registered,
three of them are located on the lower three crane
platforms.
ForestSAT 2016 Abstracts Summary
References
Béland, M., Baldocchi, D. D., Widlowski, J. L.,
Fournier, R. a., & Verstraete, M. M. (2014). On
seeing the wood from the leaves and the role of
voxel size in determining leaf area distribution
of forests with terrestrial LiDAR. Agricultural
and Forest Meteorology, 184, 82–97. http://doi.
org/10.1016/j.agrformet.2013.09.005
Calders, K., Newnham, G., Burt, A., Murphy, S.,
Raumonen, P., Herold, M., … Kaasalainen, M.
(2015). Nondestructive estimates of aboveground biomass using terrestrial laser scanning.
Methods in Ecology and Evolution, 6(2), 198–208.
http://doi.org/10.1111/2041-210X.12301
Kükenbrink, D., Leiterer, R., Schneider, F. D.,
Schaepman, M. E., & Morsdorf, F. (under
review). Quantification of hidden canopy
volume of airborne laser scanning data using
a voxel traversal algorithm. Remote Sensing of
Environment, (Special Issue of the SilviLaser 2015
Conference).
Newnham, G. J., Armston, J. D., Calders, K., Disney,
M. I., Lovell, J. L., Schaaf, C. B., … Danson, F. M.
(2015). Terrestrial Laser Scanning for Plot-Scale
Forest Measurement. Current Forestry Reports,
239–251.
http://doi.org/10.1007/s40725-0150025-5
59
Universidad Mayor
Development of Methods
POSTER
Measuring stem diameters - a comparison of three methods
Susann Klatt, Thünen Institute of Forest Ecosystems
Johannes Breidenbach, Norwegian Institute of Bioeconomy Research
Rasmus Astrup, Norwegian Institute of Bioeconomy Research
Keywords: Close-range photogrammetry, Criterion, Gator Eyes
Abstract
As descriptors of stem taper and allometry, breast height and upper diameters (dbh, du) are input
parameters to estimte tree volume and biomass. Therefore, the measurement of diameters is essential
in forest inventories. Here, we compared three devices to measure dbh and an upper diameter in
approx. 30% of tree height for 50 trees in south eastern Norway. The devices were: (1) images taken
with two reflex cameras mounted on a stereo rig (RCS), (2) a Criterion dendrometer (CRD) mounted on
a tripod and a handheld Gator Eye caliper (GEC). RCS images were taken simultaneously from seven
positions on a hemisphere with five meters distance to the tree.
Close-range photogrammetry data were processed using Python scripting for wokflow automation.
The dbh and du were measured in 3D dense point clouds using open source software. For dbh, the
devices were compared to reference measurements with a caliper.
Preliminary analysis for dbh resulted in root mean squared deviances (RMSDs) of 9.8 mm and 8.2 mm
for CRD and GEC, respectively. Systematic deviances were less than 2 mm for both methods. First
analysis of the close-range photogrammetry data were promising, as the image-matching point cloud
resembles data generated by terrestrial laser scanners. The photogrammetric method allows for the
documentation of the trees’ status and future re-processing with more advanced methods. However,
the number of pictures taken would need to be reduced to save time and storage capacity in most
forest inventory settings. Random and systematic deviances of the measurements based on the three
methods will be compared and discussed.
60
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Multi-Year Comparison of Tree Species Discrimination from
Formosat-2 Satellite Image Time Series
M. Fauvel1, D. Sheeren1,*, J.-F. Dejoux2, J. Willm1
DYNAFOR Lab., INP Toulouse/INRA, University of Toulouse, France
CESBIO Lab., CNES/UPS/IRD/CNRS, University of Toulouse, France
1
2
Keywords: Tree species, forest, time series, classication, uncertainty.
Abstract
Mapping forest composition is essential for forest management, biodiversity conservation and
predicting potential shifts of tree species under the context of climate change. However, operational
production of accurate species maps is still challenging in remote sensing. In this study, the ability
of dense high spatial resolution multispectral Formosat-2 satellite image time-series (SITS) to
discriminate tree species in temperate forests is investigated. Using height years of satellite data, we
compared the performance of single-year classications to map thirteen dominant tree species in a
study area of southwest France. Formosat time-series were composed of eleven to fourty-three dates
acquired across one year from 2006 to 2014. Canopy species were identied using parametric (GMM)
and nonparametric (k-NN, RF, SVM) machine learning algorithms at three class hierarchy levels, after
ltering noisy pixels using B-Splines. The classiers were trained and cross-validated from 1235 eldcollected plots by applying a random data-splitting procedure repreated 25 times. The results showed
a very high suitability of the SITS to identify the forest tree species based on phenological dierences.
Accuracy of single-year classications varied from 0.85 (kmean in 2008 for GMM) to 0.95 (kmean in 2014 for
SVM) with a decrease in performance as the classication became more specic from level 1 to level 3.
SVM outperformed systematically the other classiers with a higher stability in accuracy between the
years. However, a spatial analysis of the interannual variability revealed disagreements between the
single-year classications in complex mixed forests, suggesting a higher uncertainty in these areas. By
contrast, a good stability was observed within monospecic plantations (e.g. for aspen or eucalyptus).
Our ndings suggest that time-series data is a promising approach for mapping forest types. It also
demonstrates the potential contribution of the Sentinel-2 satellites for large scale forest monitoring.
Corresponding author
Email address: david.sheeren@ensat.fr (D. Sheeren)
Preprint submitted to ForestSAT’2016 May 9, 2016
61
Universidad Mayor
Development of Methods
POSTER
Multiscale forest health mapping: the potential of air- and
space-borne sensors
Iurii Shendryk1,*, Mark Broich1, Mirela G. Tulbure1, Andrew McGrath2, David Keith3, Sergey V. Alexandrov4
School of Biological, Earth and Environmental Sciences, University of New South Wales, Kensington NSW 2052,
Australia.
2
Airborne Research Australia, Flinders University, Salisbury South, 5106, Australia.
3
Centre for Ecosystem Science, School of Biological, Earth and Environmental Sciences, University of New South Wales,
Kensington NSW 2052, Australia and New South Wales Office of Environment and Heritage, Hurstville NSW 2220,
Australia.
4
Automation and Control Institute, Vienna University of Technology, Vienna, Austria
* Corresponding author: iurii.shendryk@unsw.edu.au
1
Keywords: tree health, LiDAR, imaging spectroscopy, random forest, object-oriented classification,
Australia, flooding frequency, dieback, eucalypt, individual trees, WorldView-2, SAR
Abstract
Full-waveform airborne LiDAR scanning and imaging spectroscopy data were collected in May-June
2014 to cover 100 sq. km of the largest eucalypt, river red gum forest in the world, located in the
south-east of Australia. The objective was to delineate individual trees, characterize their health using
airborne remotes sensing data, and investigate spatial relation of forest health to flooding frequency.
In addition we tested the ability of optical and SAR satellite imagery as well as low density discretereturn airborne LiDAR scans in upscaling forest health metrics in terms of basal area. Here we present
the processing methodology and results of our analysis with the main focus on upscaling forest health
metrics.
Introduction
During the past 60 years sizeable areas of forest in
Australia have experienced dieback, mostly caused
by drought and high temperatures. In order to restore the extent and distribution of healthy forest
there is a need in developing a high-tech accounting
system of trees. In this respect airborne LiDAR and
imaging spectroscopy are two potentially complementary remote sensing technologies that provide
comprehensive structural and spectral characteristics of forests over large areas. Focusing on the largest eucalypt, river red gum (RRG) forest in the world
(Barmah-Millewa forest) in this study we aimed to
(1) develop algorithms for individual tree delineation
using full-waveform LiDAR scans and (2) characterize health of delineated trees utilizing LiDAR scans
and imaging spectroscopy in a structurally complex
floodplain eucalypt forest, (3) characterize the relationship between tree health and flooding frequency
62
derived from Landsat time-series, and (4) extrapolate spatially non-contiguous tree health map derived
from LiDAR and imaging spectroscopy datasets (covering ~14% of the study area) to cover the whole
extent of this forest using optical and SAR imagery
as well as low density discrete-return airborne LiDAR scans. Once developed, this methodology can
be a starting point in the development of a nationwide forest health monitoring framework in Australia
and applied to other forests worldwide.
Methodology
We conducted experiments in the largest (737 sq.
km) eucalypt, river red gum forest in the world, located in the south-east of Australia that experienced
severe dieback over the past six decades. For detection of individual trees from full-waveform LiDAR
scans we developed a novel bottom-up approach
based on Euclidean distance clustering to detect
Universidad Mayor
tree trunks and random walks segmentation to further delineate tree crowns. The proposed algorithm
presents a stepwise procedure, where firstly tree
trunks are identified based on the spatial arrangement of points in the lowest part of the LiDAR point
cloud. Secondly, the points representing identified
tree trunks are labelled within the point cloud. Thirdly, a 3D graph is built by connecting all points within
a certain radius. Finally based on the spatial connectivity of points to the labelled ones, using so-called
random walks algorithm, the point cloud is segmented into individual trees (Fig. 1). The detailed description of the individual delineation algorithm can
be found in Shendryk, Broich et al. (2016).
The accurate delineation of trees allowed us to classify the health of this forest using machine learning
and field-measured tree crown dieback ratios, which
was a good predictor of tree health and crown density, respectively, in this forest. Although tree health is a subjective term, in Barmah-Millewa it is best
approximated by dieback levels (i.e. the proportion
of dead branches to the total number of branches),
which were visually assessed in the field and grouped
into three classes. The LiDAR indices were calculated
for segmented tree crowns exploiting the full range
of full-waveform LiDAR attributes and tree geometry, and used as predictor variables in object-oriented random forest classification. Random forest is a
supervised non-parametric machine learning technique, and was particularly suitable for this study as it
is able to achieve superior classification performance
as compared to other ensemble learning algorithms
with small training samples. The detailed description of the individual tree health classification can be
found in Shendryk, Broich et al. (2016)”B.
The decrease in flooding has been frequently identified as the main cause of tree health decline in Australia’s floodplains. In order to tease out the causes of
ForestSAT 2016 Abstracts Summary
unhealthy forest we overlaid the flooding frequency
map derived from time series of Landsat imagery
(1986-2011) (Tulbure, Broich et al. 2016) with our individual tree health maps.
Finally, we utilized very high resolution satellite
imagery (i.e. WorldView-2) in combination with
SAR imagery (i.e. Sentinel-1 and ALOS-2/PALSAR-2
imagery) and low density discrete-return airborne
LiDAR scans as a substrate for extrapolating LiDAR
and imaging spectroscopy derived individual tree
health. Similarly to individual tree health classification, we developed object-oriented random forest
regressions to quantify live and dead basal area (BA)
in 60m cells using zonal statistics as predictor variables.
Results
Overall, our individual tree delineation algorithm
was able to detect 67% of tree trunks with diameter larger than 13 cm. We assessed the accuracy of
tree delineations in terms of crown height and width, with correct delineation of 68% of tree crowns.
The increase in LiDAR point density from ~12 to ~24
points/sq. m. resulted in tree trunk detection and
crown delineation increase of 11% and 13%, respectively (Shendryk, Broich et al. 2016). Trees with incorrectly delineated crowns were generally attributed
to areas with high tree density.
Returned pulse width, intensity and density related
LiDAR indices were the most important predictors
in the individual tree health classifications. At the
forest level in terms of tree crown dieback, 77% of
trees were classified as healthy, 14% as declining
and 9% as dying or dead with 81% mapping accuracy. Landsat derived flooding frequency map for over
26 years showed that trees in areas that were flooded less than 5% of the time a pixel was observed
Fig 1: Individual tree segmentation procedure
63
Universidad Mayor
Development of Methods
Fig. 1: Preliminary LiDAR,
SAR- and WorldView2-derived
tree mortality map of the Barmah-Millewa forest, Australia.
Areas with canopy cover <10%
are masked out. Map pixels are
aggregated to 60m spatial resolution and WorldView-2 imagery
is shown as a background.
were most susceptible to dieback (Shendryk, Broich
et al. 2016).
Our preliminary results also suggest that very high
resolution optical and SAR satellite imagery in combination with low density discrete-return LiDAR
scans are effective in upscaling individual tree health of this mono-specific forest to a stand level (i.e.
60m resolution). The object-oriented random forest
regressions allowed us to quantify total, live, dead
and %dead basal area (BA) with R2 of 0.74, 0.73, 0.41
and 0.46, respectively. The most important variables
in predicting total and live BA were zonal statistics
extracted from LiDAR derived canopy height model (CHM) and canopy cover (CC) rasters as well as
ALOS-2/PALSAR-2 HV- and HH- polarized imagery,
and WorldView-2 derived near-infrared (NIR2) to red
edge (RE) derived NDVI. While the most important
variables in predicting dead and %dead BA were zonal statistics extracted from LiDAR derived CHM and
CC as well as Sentinel-1 VH-polarized imagery and
WorldView-2 derived NIR2/RE NDVI. Fig. 1 shows
the preliminary LiDAR-, SAR- and WorldView2-derived tree mortality map of the Barmah-Millewa forest in terms of %dead BA.
Conclusions
Our results provide algorithms that accurately delineate and classify the health of trees in a structu-
64
rally complex forest, enabling us to prioritize areas
for forest health promotion and biodiversity conservation. We suggest using optical and SAR satellite
imagery as well as low-density discrete LiDAR as a
possible substrate for extrapolating full-waveform
LiDAR and imaging spectroscopy derived individual
tree health in order to reduce the cost of larger scale
studies.
References
Shendryk, I., M. Broich, M. G. Tulbure and S. V.
Alexandrov (2016). “Bottom-up delineation of individual trees from full-waveform airborne laser
scans in a structurally complex eucalypt forest.”
Remote Sensing of Environment 173: 69-83.
Shendryk, I., M. Broich, M. G. Tulbure, A. McGrath,
D. Keith and S. V. Alexandrov (2016). “Mapping
individual tree health using full-waveform airborne laser scans and imaging spectroscopy: A case
study for a floodplain eucalypt forest.” Remote
Sensing of Environment 187: 202-217.
Tulbure, M. G., M. Broich, S. V. Stehman and A. Kommareddy (2016). “Surface water extent dynamics
from three decades of seasonally continuous
Landsat time series at subcontinental scale in a
semi-arid region.” Remote Sensing of Environment 178: 142-157.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Near Real-Time Detection of Forest Changes using Google
Earth Engine® and Sentinel-2 Imagery: A Case Study in
Curacutin region, Chile
Nils Nölke1, Hans Fuchs1 , Victor Sandoval2, Patricio Acevedo3, Guillermo Trincado2, Manuel Castro3, Nolwenn Boucher2,
Hugo Zerda4, Christoph Kleinn1
Chair of Forest Inventory and Remote Sensing, Georg-August-Universität Göttingen, Germany
2
Instituto de Bosques y Sociedad, Universidad Austral de Chile, Valdivia, Chile
3
Departamento de Ciencias Físicas, Universidad de la Frontera, Temuco, Chile 4Facultad de Ciencias Forestales,
Universidad Nacional de Santiago del Estero, Argentina
1
Keywords: forest change, Google Earth Engine, Sentinel-2
Human induced changes in forest area
(deforestation, reforestation) and forest condition
(forest degradation) are of great significance in
national and international policy processes - both in
the context of regularly planned forest management
and illegal or accidental interventions.
The magnitude and spatial distribution of these
activities vary within a country and are difficult
to monitor in particular for larger areas. Remote
sensing is the only source of information that can be
used to identify these activities in larger areas and
shorter time intervals. The use of high-resolution
imagery acquired on airborne or UAV platforms is
time-consuming when it comes to processing, and
requires very high-performance computers. Similarly,
the processing of satellite imagery acquired at short
time intervals on national and subnational scale is
very resource-intensive. Nevertheless, the latter
may be the only practical approach if one seeks to
identify forest changes in the context if implementing
efficient “early warning systems”.
The Google Earth Engine® provides a powerful
means for identifying changes in forests by
combining high-performance cloud computing with
the automated retrieval of satellite imagery from a
vast depository of images.
Our case study is conducted in Curacautín, Chile,
and this presentation aims to present results from
the initial phase of the study and to evaluate the
use of the Google Earth Engine® for identifying
areas of forest change at an early point in time.
We distinghuish between forest conversion
(deforestation) and changes in forest conditions
(forest degradation), and match and compare the
changes with the corresponding forest management
plans in order to provide a reliable interpretation.
An Earth Engine® JavaScript was written to access
the Sentinel-2 imagery archive of the Engine and
serves to detect changes within two consecutive
months with a minimum mapping unit of 0.1 ha. The
Sentinel-2 satellite imagery was then analyzed using
the built-in image processing functions; threshold
criteria were identified and used to reliably indicate
the occurrence of forest changes.
To illustrate the findings of our study, the
presentation compares the results of our automated
approach (with the Google Earth Engine® cloud
processing) to the results of independent analyses
of aerial photographs and satellite imagery. Finally,
the potential of high temporal revisits of the only
recently available Sentinel-2 imagery for the
detection of human induced changes in forests is
explored.
65
Universidad Mayor
Development of Methods
POSTER
Optimization of dynamic a global vegetation model at a
land cover remote sensing data for better representation
of Russian forests
Sergey Venevsky Tsinghua University, Beijing, China
Sergey Khvostikov, Space Research Institute, Moscow, Russia
Sergey Bartalev, Space Research Institute, Moscow, Russia
Dynamic global vegetation model SEVER has
been regionally adapted using a remote sensing
data derived land cover map in order to improve
reconstruction conformity of vegetation functional
types distribution over Russia, particularly forests.
The SEVER model was modified to address
noticeable divergences between modelling results
and the land cover map. The model modification
included a light competition method elaboration
and introduction into the model a tundra class. The
rigorous optimisation of key model parameters was
done using two-steps procedure. First approximate
global optimum was found with Efficient Global
Optimisation algorithm, afterwards local search
in vicinity of approximate optimum was done with
quasi-Newton algorithm BFGS. The regionally
adapted model shows significant improvement
of the vegetation distribution reconstruction over
Russia with better matching with the satellite derived
TerraNorte land cover map, which was confirmed by
66
both a visual comparison and a formal conformity
criterion.
The closest fit of the model to the remote sensing
data derived plant functional types distribution is
seen in the Asian part of Russia, which got human
activities mainly in the most southern part near
the border. Especially important is geographically
correct representation of areas for boreal needleleaved deciduous forests (Larix forests) and tundra
in the Asian part of Russia. This confirms that Larix
forests and tundra are determined mainly by climatic
variables. Differences in vegetation distribution
between optimised version of SEVER DGVM and the
TerraNorte land cover map seen in the European part
of Russia also provide us with important insight. We
conclude that areas, climatically suitable for boreal
and temperate broad-leaved forests, are converted
to agricultural lands (described as grassland PFTs in
SEVER DGVM) by human activities.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Predicting Single Tree Species Diameter Distribution by
Airborne Laser Scanning Using Different Modeling Alternatives
Maltamo, M.1*, Mehtätalo, L.2, Valbuena, R. 1, Vauhkonen, J. 1 and Packalen, P. 1
1
University of Eastern Finland, Faculty of Science and Forestry, School of Forest Sciences, P.O. BOX 111,
FI-80101 Joensuu, Finland.
2
University of Eastern Finland, Faculty of Science and Forestry, School of Computing, P.O. BOX 111,
FI-80101 Joensuu, Finland.
* Corresponding author. E-mail address: matti.maltamo@uef.fi
Keywords: LiDAR, tree size distribution, Weibull, k-nn, plantation, parameter prediction,
parameter recovery
Abstract
Diameter distribution is a stand-level indicator of the structure of the tree stock and, in
forest inventory applications, the information on diameter distribution allows the calculation
of timber assortments. However, diameter distribution is usually not included in the field
measurements of stand-level management inventory systems. In such cases, the diameter
distribution can be predicted using field assessed stand attributes as predictors in theoretical
distribution functions, such as Weibull (e.g. Bailey and Dell 1973), or k-nearest neighbor (k-nn)
imputation (e.g. Maltamo and Kangas 1998).
The emergence of airborne laser scanning (ALS) techniques has had major impact on stand
level management inventories (e.g. White et al. 2013). However, there are still primary
information needs concerning diameter distribution but it cannot be directly produced by the
inventory, and thus, it must be predicted (Maltamo and Gobakken 2014). So far, ALS data
have been utilized to predict aggregated stand-level diameter distribution applying Weibull
distribution and k-nn imputation in some earlier works (Gobakken and Næsset 2004, Maltamo
et al. 2009) but the existence of several tree species and irregular distribution shapes usually
hampers the prediction (Packalen and Maltamo 2008).
This study aims to examine the ability of metrics of area based approach of ALS to estimate
single tree species diameter distribution in an even-aged monoculture plantation, which
provides disturbance free conditions to analyze the relationship between ALS and diameter
distribution due to the missing minor tree species or undergrowth. The test area is located in
Bahia, Brazil and the considered tree species is Eucalyptus urograndis. We compare different
methods to predict diameter distribution, namely parameter prediction, parameter recovery
and distribution matching based on the Weibull function form, in addition to a non-parametric
k-nn imputation. We also include such a parameter prediction alternative, where Weibull
distribution is predicted using field attributes which are typically available in plantations’ stand
register data. The criteria of the goodness-of-fit are based on the shape of the distribution, i.e.
error indices of diameter classes and third powers of diameters of proportional frequencies
between 0 and 1.
67
Development of Methods
Universidad Mayor
In general, the results showed that ALS information can predict diameter distribution with an
error of slightly more than 10% of relative root mean square error of third power of diameter
class frequencies and error index values between 50-60%. The reliability figures of the most
accurate ALS based alternatives are also close to the field information based estimates. The
results showed strong relationship between ALS information and diameter distribution but
the results concerning the comparison between different prediction methods varied regarding
the reliability criteria used.
References
Bailey, R.L. and Dell, T.R. 1973. Quantifying diameter
distributions with the Weibull function. For. Sci.
19: 97–104.
Gobakken, T. and Næsset, E. 2004. Estimation of
diameter and basal area distributions in coniferous
forest by means of airborne laser scanner data.
Scand. J. For. Res. 19: 529–542.
Maltamo, M. and Gobakken, T. 2014. Predicting tree
diameter distributions. In Forestry Applications
of Airborne Laser Scanning –concepts and
case studies by Maltamo, M., Næsset, E. and
Vauhkonen, J. Managing Forest Ecosystems vol
27. Springer, pp.177-191
Maltamo, M. and Kangas, A. 1998. Methods based on
k-nearest neighbor regression in the estimation
of basal area diameter distribution. Can. J. For.
Res. 28: 1107-1115.
68
Maltamo, M., Næsset, E., Bollandsås, O.-M.,
Gobakken, T. and Packalén, P. 2009. Nonparametric estimation of diameter distributions
by using ALS data. Scand. J. For. Res. 24: 541-553.
Packalén, P. and Maltamo, M. 2008. The estimation
of species-specific diameter distributions using
airborne laser scanning and aerial photographs.
Can. J. For. Res. 38: 1750-1760.
White, J.C., Wulder, M.A., Varhola, A., Vastaranta,
M., Coops, N.C., Cook, B.D., Pitt, D., and Woods,
M. 2013. A best practices guide for generating
forest inventory attributes from airborne laser
scanning data using an area-based approach.
Forest Chron. 89(6): 722-723.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Quality control and quality assessment of LiDAR data
Assis, M.1; Cantinho, R. Z.1; Oliveira, P. V. C.1; dos-Santos, M. N.2; Gorgens, E. B.3; Ometto, J. P.1
Earth System Science Center, National Institute for Space Research. Av. dos Astronautas, 1758, Jardim da Granja, 3o
andar. São José dos Campos, São Paulo, CEP 12227-010, Brazil.
2
Empresa Brasileira de Pesquisa Agropecuária (EMBRAPA-CNPTIA) Av. Dr. André Tosello, 209 - Cidade Universitária,
Campinas - SP, Brazil.
3
Universidade Federal dos Vales do Jequitinhonha e Mucuri. Rodovia MGT 367 - Km 583, nº 5000, Alto da Jacuba CEP
39100-000, Brazil.
1
Keywords: ALS, LiDAR, data validation, MDT, point cloud, QAQC
A consistent and reliable biomass map for the
whole Brazilian Amazon forest is one of the primary
objectives of the research project Estimation
of Biomass in the Amazon (EBA). A total of 625
randomly distributed lidar flight transects covering
an area of 300mx12500m each is being collected
since 2016 February. To establish a systematic
and consistent data collection methodology, the
vendor was required to meet a strict pre-defined
set of requirements and specifications. Although
primary quality assessment and quality control of
the data are typically a responsibility of the vendor,
confirming the integrality of the data received
became necessary. We developed a protocol
to assess return density, return homogeneity,
footprint, completeness of the data, the range of
values, spatial coverage, scan angle, datum (vertical
and horizontal), projection, linear unit and overall
adherence to accuracy requirements.
It is indispensable to predetermine which products
are considered deliverables: point cloud and its file
format (usually .las or a compacted version, laz);
digital models (commonly terrain and surface) and
file format (generally .asc). For algorithms, it is
important to define which ones will be used for the
deliverables production and the processing degree
(data interference) performed by the contractor
(i.e. outliers removal, ground filtering and digital
terrain modeling). Furthermore, it is important to
define how to proceed in case an unexpected event
occurs, especially for large campaigns as ours. Our
study presents Python scripts for data validation to
be applied by those who are contracting ALS data.
Phyton is a free software development platform, and
its use is increasing fast in Data Science processing.
It also includes laspy, a library that enables lidar data
direct accessing and processing. Finally, aiming at
raising awareness of the lidar data users community,
we present the problems commonly found during
the validation process.
69
Development of Methods
Universidad Mayor
Shadow compensation for imaging spectroscopy
data using a radiative transfer approach
Daniel Kükenbrink, Fabian D. Schneider, Andreas Hueni, Felix Morsdorf, Michael E. Schaepman
Remote Sensing Laboratories, University of Zurich, Winterthurerstrasse 190, CH-8057, Zurich, Switzerland
Keywords: imaging spectroscopy, airborne laser scanning, radiative transfer, shadow compensation,
target reflectance, forestry, raytracing
The retrieval of accurate target reflectance values
from imaging spectroscopy data in forest ecosystems
is a challenging task due to their structural
complexity, largely influencing the measured
radiance at sensor level. Topographic and tree crown
shadowing effects highly limit accurate biophysical
and biochemical parameter retrievals from imaging
spectroscopy data. Current approaches trying to
overcome these limitations often exclude shadowed
areas, significantly reducing the study area, or rely on
simple, but often inaccurate shadow compensation
approaches based on histogram thresholding, band
ratios or matched filters.
In this contribution, we introduce a novel approach
to compensate shadowed areas in imaging
spectroscopy reflectance data using a radiative
transfer approach based on the radiative transfer
model DART (Gastellu-Etchegorry et al., 2015).
DART is able to estimate the three-dimensional
radiative budget, from which we can extract the
surface irradiance per image pixel. We parameterised
the DART model using airborne laser scanning and
in-situ data for a Swiss temperate forest following
the approach proposed by Schneider et al. (2014).
A shadow compensation factor was retrieved per
spectral band and pixel based on the surface and
the top of scene irradiance unaffected by shadowing
effects, both extracted from the three-dimensional
radiative budget output of DART.
70
The proposed shadow compensation approach was
tested on simulated (using DART) and real imaging
spectroscopy data derived from the Airborne Prism
Experiment (APEX) sensor (Schaepman et al.,
2015)its calibration and subsequent radiometric
measurements as well as Earth science applications
derived from this data. APEX is a dispersive
pushbroom imaging spectrometer covering the solar
reflected wavelength range between 372 and 2540nm
with nominal 312 (max. 532. The approach showed
promising results in successfully compensating
the reflectance values of ground pixels affected by
shadows, which were cast by trees. Figure 1 shows the
preliminary results derived from a DART simulation of
a simple scene composed by a single tree located in
the middle of a meadow with a homogenous known
reflectance of 20% in the simulated band (560±1 nm).
These preliminary results show that the proposed
compensation approach is able to successfully
extract target reflectance inside the core shadow.
However, the complex light scattering mechanisms
inside the forest canopy remain difficult to model,
rendering successful target reflectance retrieval as
well as its validation challenging. Nevertheless, we
argue that the proposed shadow compensation
approach significantly improves target reflectance
estimation, as well as the retrieval of biochemical and
biophysical parameters from high resolution imaging
spectroscopy data.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Figure 1: Preliminary results of the shadow compensation approach applied to a synthetic scene composed
by a single tree located in the middle of a meadow with known homogenous reflectance of 20% in the simulated band. The scene was parameterised and the image was simulated using DART at a single spectral
band in the green spectrum (560±1 nm). The sun azimuth and zenith angles were defined as 180º and 48º
respectively. Top left shows the uncompensated original reflectance image at the given band. The bottom
left shows the top of scene irradiance and the surface irradiance for the profile denoted by the red line. Top
right shows the compensated image and the bottom right the reflectance profile along the red line for the
uncompensated and compensated image.
References:
Gastellu-Etchegorry, J.-P., Yin, T., Lauret, N.,
Cajgfinger, T., Gregoire, T., Grau, E., …
Ristorcelli, T. (2015). Discrete Anisotropic
Radiative Transfer (DART 5) for Modeling
Airborne and Satellite Spectroradiometer
and LIDAR Acquisitions of Natural and Urban
Landscapes. Remote Sensing, 7(2), 1667–1701.
doi:10.3390/rs70201667
Schaepman, M. E., Jehle, M., Hueni, A., D’Odorico, P.,
Damm, A., Weyermann, J., … Itten, K. I. (2015).
Advanced radiometry measurements and
Earth science applications with the Airborne
Prism Experiment (APEX). Remote Sensing
of Environment, 158, 207–219. doi:10.1016/j.
rse.2014.11.014
Schneider, F. D., Leiterer, R., Morsdorf, F., GastelluEtchegorry, J.-P., Lauret, N., Pfeifer, N., &
Schaepman, M. E. (2014). Simulating imaging
spectrometer data: 3D forest modeling based
on LiDAR and in situ data. Remote Sensing
of Environment, 152, 235–250. doi:10.1016/j.
rse.2014.06.015
71
Development of Methods
Universidad Mayor
Simulating the spectral response of tropical tree species with
3-D radiative transfer modeling
Matheus Pinheiro Ferreira1, Jean-Baptiste Féret2, Eloi Grau2, Yosio Edemir Shimabukuro1, Carlos Roberto de Souza Filho3
National Institute for Space Research (INPE), São José dos Campos, Brazil - (mpf; yosio)@dsr.inpe.br
Irstea, Land, environment, remote sensing and spatial information (TETIS), Montpellier, France, - (jean-baptiste.feret;
eloi.grau)@teledetection.fr
3
Institute of Geosciences, University of Campinas, Campinas, Brazil - beto@ige.unicamp.br
1
2
Abstract
Spatial and temporal information on biochemical and biophysical attributes of forest areas are key
for global and regional studies of ecosystem processes. In this purpose, hyperspectral remote sensing
combined with physically-based radiative transfer models (RTMs) has been widely used. However,
such approach requires an accurate simulation of the remote sensing signal, which is challenging
due to the complex structure and spectral diversity of forest canopies, particularly in tropical
environments. This issue can be addressed using three-dimensional (3D) RTMs such as the Discrete
Anisotropic Radiative Transfer (DART) model. In this study, the DART model was used to simulate the
spectral response of tropical tree species at the canopy level. We tested the influence of a simplified
crown structure and optical properties leading to realistic simulations. A look-up table of individual
tree crown (ITC) reflectance was built using combinations of DART parameters derived from in situ
measurements. Then, measured spectral characteristics of individual trees were compared to their
simulated counterparts in terms of Root Mean Square Deviation (RMSD). Finally, a sensitivity analysis
was performed to understand how the DART parameters affect the simulated canopy reflectance.
Differences between experimental and simulated data reached RMSD values from 0.2 to 0.9% in the
visible (VIS, 450-650 nm), 0.4 to 1.2% in the red edge (RE, 700-750 nm) and 0.2 to 1.8% in the nearinfrared (NIR, 750-920 nm). The sensitivity analysis showed that structural parameters are dominantly
responsible for the variation observed in the NIR range, where the highest disagreements between
simulated and experimental data were observed. Our study shows that the simulation of the spectral
response of tropical tree species is feasible with 3D RTMs, which is the first step to inversely retrieve
biophysical and biochemical attributes of tropical forest canopies using physically-based methods.
1. Introduction
Retrieval of biophysical and biochemical vegetation
properties using remotely sensed data has long been
performed with empirical models. These models
relate a measure of reflectance or a vegetation index
(VI) to the variable of interest measured in the field.
Although these methods are simple and fast, they
have some unresolved drawbacks such as: (i) the
validity of the model is limited to the environmental
conditions where it has been developed, (ii) variables
influencing the radiation regime such as the canopy
structure are neglected and (iii) the relationship
between VIs and the variable of interest is established
using costly fieldwork data. 3D Radiative Transfer
Models (RTMs) offer the possibility to encompass
72
these limitations by integrating the canopy
architecture and simulating remote sensing data in a
variety of acquisition and environmental conditions.
The retrieval of canopy variables from measured
signals is performed by inversion of the model. The
inversion basically consists of finding the best set of
model parameters that generate the most similar
spectrum to a given experimental spectrum (e.g. the
spectral response of a pixel). This procedure mainly
depends on a good agreement between simulated
and measured data. However, little is known about
the accuracy of simulations of hyperspectral images
as acquired over tropical forests.
Universidad Mayor
In this study we designed a methodology aiming
at comparing experimental imaging spectroscopy
acquired over a complex tropical forest with
DART simulations obtained after application of
a certain number of simplifying hypotheses. This
methodology is based on the generation of a look-up
table (LUT) of individual tree crown (ITC) reflectance
using DART simulations obtained by realistic
combinations of biophysical and chemical vegetation
properties derived from field observations, and a
simplified geometric architecture of the trees. We
then compared simulations with experimental data
in terms of spectral similarity. Finally, a sensitivity
analysis was performed to understand how each
DART parameter affects the simulated canopy
spectral response.
2. Material and Methods
2.1 Study area
The study area is a well-preserved tropical seasonal
semi-deciduous forest located in the municipality
of Campinas, São Paulo State, southeastern Brazil
(22°49’13.4”S 47°06’43.6”W). It is an old-growth
forest area, about 630 m a.s.l, subjected to a 5-month
dry season (May to September) and characterized
by deciduous and evergreen tree species.
2.2 Hyperspectral data
Hyperspectral data was acquired under clear sky
conditions on June 7, 2010, using the AisaEAGLE
(Spectral Imaging, Inc., Oulu, Finland) sensor that
covers the visible/near-infrared (VNIR, 400-970 nm)
wavelength range. 122 VNIR spectral bands spaced
by 5 nm were acquired with a radiometric resolution
of 12 bits. Ten flight-lines in the North-South direction
were necessary to cover the whole study area (251.8
ha). The data collected were atmospherically and
geometrically corrected according to the procedures
described in Ferreira et al. (2016). The spectral bands
located below 450 and above 920 nm were discarded
due to their low signal-to-noise ratio (SNR). Then, 99
bands covering the 450-920 nm spectral range were
retained for further use.
2.3 Field data
2.3.1 Individual tree crown and leaf optical
properties data
In this work, we used the ITC dataset of Ferreira et al.
(2016), which was produced by visual interpretation
of the hyperspectral data and field work. A total of
ForestSAT 2016 Abstracts Summary
234 ITCs were identified, corresponding to seven
tree species. Additionally, leaves of these species
were collected at the beginning of the dry season
(end May 2015) for reflectance measurements. The
spectroradiometer ASD/FieldSpec®4 (Analytical
Spectral Devices (ASD), Inc., Boulder, Colorado) with
the Plant Probe accessory, combined with the Leaf
Clip assembly, was used to measure leaf reflectance
in the 350 to 250 nm range.
3. Methods
3.1 DART radiative transfer
DART is a 3D radiative transfer model that simulates
radiation propagation in urban and natural
landscapes (Gastellu-Etchegorry et al., 2015). In this
work, to simulate the spectral response of tropical
tree species at the ITC level, a single tree located at
the center of a 10x10 m scene was used. The crown
was of round shape with a diameter of 10 m, centered
at 20 m height. The canopy was represented by
0.5 m³ cells filled with foliage and woody branches
modeled with two separated turbid media, defined
by two densities, spectra and proportion of full
cells. Branch reflectance spectra were set to a
measured bark spectra. The forest understory was
represented by a turbid plot of 2.5 m height (±0.5 m)
covering a flat ground. Standard spectra from the
DART database were used to characterize optical
properties of the understory vegetation, braches
and litter. Leaf optical properties were simulated
with the PROSPECT-5 model (Féret et al., 2008).
3.2 LUT generation
We generated a LUT by varying 11 DART parameters
(Table 2), which can be divided into three main
categories: canopy structural properties, leaf
biophysical properties (corresponding to the input
parameters of PROSPECT-5) and scene optical
properties.
3.3 Sensitivity analysis
We are interested to identify the most influential
DART parameters and quantify how much they
contribute to the variability in the simulated canopy
reflectance. To do this, we performed a one at a
time sensitivity analysis (OAT-SA) and computed
a set of sensitivity indices. The OAT-SA aimed to
understand how the DART parameters individually
affect the spectral response of the simulated tree.
Thus, we varied the DART parameters one at a time
73
Development of Methods
Universidad Mayor
Fig. 1 Root Mean Square Deviation (%) between the simulated and the measured spectral response of seven
tree species from a tropical seasonal semi-deciduous forest formation. The mean spectral response of the
species are plotted for clarity. AP=Aspidosperma polyneuron; AG=Astronium graveolens; CL=Cariniana legalis;
CP=Croton piptocalyx; DS= Diatenopteryx sorbifolia; HC=Hymenaea courbaril; PL= Pachystroma longifolium.
74
Universidad Mayor
ForestSAT 2016 Abstracts Summary
by specific increments, keeping all other parameters
at their base-case values, which yielded a factorial
design of size 56 (Table 2).
3.4 Comparison between experimental and
simulated data
Pixels from the manually delineated ITCs were
extracted from the hyperspectral image and
averaged to compose an experimental dataset of 234
spectra. We selected only sunlit crown pixels within
each ITC to compute its mean spectral response.
Sunlit pixels were defined as those presenting high
NIR reflectance (>0.2). This choice was motivated
by the fact that sunlit pixels are less influenced by
shadows and thus present high signal-to-noise ratios.
We averaged the reflectance corresponding to sunlit
pixels for each ITC and obtained an experimental
dataset of 234 spectra. Similarly, the mean spectral
response of each simulated tree was computed,
resulting in a simulated dataset of 3518 spectra. We
then compared simulated with measured spectra,
using the Root Mean Square Deviation (RMSD) as a
criterion for spectral similarity.
4. Results
4.1 Spectral similarity
Spectral differences between experimental
spectra and their simulated counterpart (based on
minimized RMSD criterion) and averaged per species
varied from 0.2 to 1.8% over the VNIR domain. More
specifically, RMSD reached values from 0.2 to 1% in
the visible (VIS, 450-700 nm), 0.5 to 1.8% in the rededge (RE, 700-750 nm) and 0.4 to 1.4% in the nearinfrared (NIR, 750-920 nm) (Fig.1). Regardless of the
species, there was an increasing trend in RMSD from
VIS to NIR, which can be noted starting at 750 nm,
and due to the lower overall values of reflectance in
the VIS domain due to saturating absorption from
photosynthetic pigments. Overall, simulated and
experimental spectra showed a better agreement in
the VIS than in the NIR.
4.2 Sensitivity analysis
The sensitivity analysis provided the opportunity
to better understand how each DART parameter
affected the crown spectral response (Fig. 2). The
VIS range is more influenced by the leaf optical
properties than by the canopy structural parameters.
Most notably, Cab variations changed the spectra
from 500 to 750 nm, impacting more severely the
green peak and the red-edge (Fig. 2).
Fig. 3 (a) Extreme (dotdashed lines), inter-quartile
(grey) and median (bold line) of DART simulated
reflectance values, obtained after varying each
parameter in turn keeping all other parameters fixed
at their reference values (one at a time sensitivity
analysis (OAT-SA)) (see Table 2). (b) Dynamics of the
sensitivity indices computed for the OAT-SA over
the 450-920 nm wavelength range.
More subtly, Cxc and N also affected the VIS, while N
impact the entire region by increasing the amplitude
of the spectra, Cxc produced variations only in the
vicinity of 525 nm. Cm and Cw made a negligible
contribution to variations in the visible, with the
former affecting intensely the NIR and the latter
producing no variations at all.
5. Conclusions
Our study demonstrates the feasibility of using 3D
radiative transfer modeling to simulate airborne
hyperspectral data acquired over high diversity
tropical forest areas. We performed simulations
of the spectral response of tree species with small
75
Development of Methods
Universidad Mayor
RMSD ( 1.8%), using a simplified geometrical
crown filled with turbid media characterized by
leaves and branches optical properties. The highest
disagreements between experimental and simulated
data were found at the NIR, which was influenced
mainly by canopy structural DART parameters.
Ferreira, M. P., Zortea, M., Zanotta, D. C.,
Shimabukuro, Y. E., & de Souza Filho, C. R.
(2016). Mapping tree species in tropical seasonal
semi-deciduous forests with hyperspectral
and multispectral data. Remote Sensing of
Environment, 179, 66-78.
Acknowledgements
Gastellu-Etchegorry, J-P, et al., (2015). Discrete
Anisotropic
Radiative
Transfer
(DART
5) for modeling airborne and satellite
spectroradiometer and LIDAR acquisitions of
natural and urban landscapes. Remote Sensing,
7, 1667-1701.
This work was supported by the São Paulo Research
Foundation (FAPESP) grant no. 2013/11.589-5.
References
Féret, J.B., François, C., Asner, G. P., Gitelson, A.
A., Martin, R. E., Bidel, L. P., Ustin, S.L., Marie,
G., & Jacquemoud, S. (2008). PROSPECT-4 and
5: Advances in the leaf optical properties model
separating photosynthetic pigments. Remote
Sensing of Environment, 112, 3030-3043.
76
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Single tree detection with weak canopy shape constraints
Ruedi Boesch, Mauro Marty
Swiss Federal Institute for Forest, Snow and Landscape Research WSL, Birmensdorf, Switzerland(ruedi.boesch, mauro.marty)@wsl.ch
Keywords: tree, detection, shape, gradient
Abstract:
Deriving object features like single trees from point cloud data is crucial to understand the structure
and function of forest ecosystems. The recognition process of single trees is still a challenge due to
different forest types, varying acquisition conditions and often difficult terrain conditions.
Nowadays acquisition campaigns using airborne laser scanning or aerial images allow to derive
georeferenced point clouds with high spatial resolution and nearly uniform point density, even over
large areas.
Due to established uniform point density we propose a raster-based method to derive single trees
from point cloud data with minimized spatial constraints for tree crowns.
Initially a candidate of a single tree crown is found at a local height maximum. Starting from this
candidate location, 8 compass directions (separated by 45 degrees) are investigated, if the height
gradient values conform to a tree crown shape.
The vast number of tree species at different growing seasons requires to be very restrictive concerning
shape assumptions in point cloud data. Therefore the single gradient values can be varying, either
convex or concave, but must be positive. Gradients are positive if the height is constantly decreasing
in one of the 8 compass directions starting from the center location.
Due to filtering or matching errors during point cloud generation and varying acquisition problems,
digital surface models of point clouds are not as smooth as expected or simply contain data generation
errors. Therefore the strict gradient criteria for the 8 directions must be relaxed. Following one of the
8 compass directions, the gradient must be positive, but can be interrupted by small local maxima.
Small local maxima are represented by a short sequence of negative gradients. Search directions with
positive gradients can therefore be interrupted by negative gradients and still be classified as valid
tree crown direction. This relaxed shape constraint in one dimension must be fulfilled by all search
directions.
The proposed method will be evaluated with the available data from the NEWFOR-benchmark,
published by Eysn, Hollaus et al.,Forests 2015, 6, 1721-1747; doi:10.3390/f6051721.
77
Universidad Mayor
Development of Methods
Study on Removal of Atmospheric Effect on Normalized
Difference Vegetation Index
Authors: Haitao Lv1, and Yong Wang1,2,3*
School of Resources and Environment, University of Electronic and Science Technology of China (UESTC)
2006 Xiyuan Avenue, West Hi-tech Zone, Chengdu, Sichuan 611731, China
2
Department of Geography, Planning, and Environment, East Carolina University
Greenville, NC 27858, USA
3
Institute of Remote Sensing Big Data, Big Data Research Center of UESTC
2006 Xiyuan Avenue, West Hi-tech Zone, Chengdu, Sichuan 611731, China
*
Corresponding author, wangy@ecu.edu
1
Keywords: Empirical analysis; Radiative transfer model; atmosphere effect removal.
Abstract:
The normalized difference vegetation index (NDVI) is a popular and widely-used indicator in the land
application using visible and NIR data. Unfortunately, the presence of atmosphere (especially clouds)
affects red and NIR reflectance. With the increasing of atmospheric optical thickness, NDVI values
are obtained in lower panels. An empirical and radiative transfer algorithm is developed to remove
atmosphere effect on NDVI. We first assume that under the atmosphere-free condition, the topof-atmosphere reflectance values from ground targets between any two visible bands have a linear
relationship. The relationship is independent from different terrain types. We then assume that the
influences on any two visible bands caused by the atmosphere are linearly related. Finally, under the
cloud-cover water pixels, the reflectance values of a visible band caused by atmosphere are linear
related to the reflectance values of a near infrared band (NIR). After validating the assumptions and
solving six unknown parameters, the NDVI values after removal of atmospheric effect are obtained.
The effectiveness of the algorithm has been evaluated in forest areas using the Landsat 8 data. In
comparison with the data before and after the algorithm, the effect of atmosphere on NDVI disappears
visually. The values of mean increases. The increase of the mean values is further supported by the
rightward shift of the histogram curve for NDVI. Overall, the removal of atmosphere effect on NDVI
was quantitatively and quantitatively satisfactory.
78
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Synergism of SAR and optical data for land use mapping
in the Amazonia transition landscape
João Arthur Pompeu Pavanelli¹*, João Roberto dos Santos¹, Lênio Soares Galvão¹, Maristela Ramalho Xaud², Haron
Abrahim Magalhães Xaud²
¹National Institute for Space Research - INPE, Astronautas Av., 1758, São José dos Campos, São Paulo, Brazil
²Brazilian Agricultural Research Corporation - EMBRAPA, BR-174, Boa Vista, Roraima, Brazil
*Corresponding author: joao.pompeu@inpe.br
Keywords: Random Forest, Land use, SAR data, Savannah, Tropical forest
Abstract
The aim of this study was to analyse the integration of OLI/Landsat-8 optical and PALSAR/ALOS-2
images to characterize the landscape of ecological tension between forest and savannah in northern
Brazilian Amazonia using the Random Forest classifier. Landsat-8/OLI surface reflectance and ALOS2/PALSAR-2 HH and HV amplitude were tested for classification purposes. In addition, NDVI and EVI
were calculated from OLI data, and GLCM metrics were determined from PALSAR-2 images for both
HH and HV polarizations. Five polarization PALSAR-2 indices were also extracted. Random Forest
was calibrated for optimal “ntree” and “mtry” parameters. Results showed an overall classification
accuracy of 82.41% and a Kappa of 0.8. The synergistic use of OLI and PALSAR-2 data improved the
classification of a great number of classes (17) in the study area.
Introduction
Land use and land cover mapping in tropical
landscapes are key components for management,
conservation and better understanding on the
anthropogenic impacts over natural ecosystems.
However, the complexity of these fragmented
landscapes due to the transition between vegetation
physiognomies (savannahs and forests), forming
a mosaic of land use for agriculture and/or pasture
often, affects their characterization using optical
remote sensing data (LU et al., 2007). In addition,
persistent cloud cover creates difficulties for
obtaining cloud-free optical data over the region.
In this sense, the synergistic use of optical and SAR
satellite data is relevant because it generally improves
the quality of land use/land cover maps (LAURIN et
al., 2013). The objective of this study was to analyse
the integration of Operational Land Imager (OLI)/
Landsat-8 and Phased Array type L-band Synthetic
Aperture Radar (PALSAR)/ALOS-2 images to map
landscapes of ecological tension between forests
and savannah in northern Amazonia, in Brazil, using
the Random Forest (RF) classifier.
Methodology
The study area is located in the Roraima state, in
northern Brazilian Amazon (Figure 1). The region is
characterized by the contact of tropical forests and
savannah physiognomies, including also classes
of anthropogenic land uses such as smallholder’s
activities and large-scale commodity-driven
agriculture. These classes form a complex landscape
from a remote sensing perspective. The forest and
savannah domains are separated by the Mucajaí
River, composing an ecological tension zone.
Seventeen classes were defined during fieldwork
and tested for RF classification: (1) agriculture; (2)
waterbodies; (3) campinarana; (4) wooded savannah;
(5) savannah grassland; (6) shrub savannah; (7)
initial secondary succession (SS1); (8) intermediate
79
Development of Methods
Universidad Mayor
Figure 1. Location of study area highlighting the transition between tropical forests and savannah
physiognomies, separated by the Mucajaí River.
secondary succession (SS2); (9) burned savannah;
(10) woodland savannah; (11) mature forest; (12)
clean pasture; (13) overgrown (dirty) pasture; (14)
silviculture; (15) clear-cut silviculture; (16) bare soil;
and (17) palm swamps.
Landsat-8/OLI CDR surface reflectance data (30
metres of spatial resolution) and PALSAR-2/ALOS
dual (HH+HV), level 1.5, CEOS format (10 metres
of spatial resolution) were used for classification.
From the OLI dataset, the Normalized Difference
Vegetation Index (NDVI) and the Enhanced
Vegetation Index (EVI) were generated. The
PALSAR-2 image was filtered for speckle using a 3
x 3 Lee filter. After that, Grey Level Co-occurrence
Matrix (GLCM) textures metrics were extracted
from each HH and HV polarization image. We also
calculated five polarizations indices: HH+HV, HHHV, HH/HV, HV/HH and a SAR Index = (HH*HV)/
(HH+HV) (Lu et al., 2013). In total, we tested 35
attributes for RF classification: 8 optical reflectancederived attributes and 27 SAR amplitude-derived
metrics. All the images were co-registered using an
80
ortho-rectified OLI image subsequently resampled
to 10 meters.
The thematic classification was performed using
the data mining RF algorithm. In order to improve
classification accuracy, a previous calibration step
was required to define the parameters “ntree”
(number of trees in the forest) and “mtry” (number
of variables at each split node) (MILLARD and
RICHARDSON 2015). By default, the “ntree”
parameter is 500 trees and “mtry” is the square root
of the number of input variables, in this case 6.
RF was processed by combining the “ntree” and
“mtry” in order to obtain the lowest classification
error, starting with “ntree” = 100 and “mtry” =1, and
changing “ntree” from 100, 200, 300… 1000 and mtry
from 1 to 35. The set of combined attributes with the
minimum error was used for land use mapping.
Validation of the classification results was assessed
using a separate set of samples supported by
ground-truth information.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Figure 2. Land use/land cover map derived from the Random Forest classification using the hybrid set of
PALSAR-2 and OLI attributes.
Results
References
The hybrid PALSAR-2 and OLI classification (Figure
2) produced an overall accuracy of 82.41% and a
Kappa of 0.8, which is a good result due to the large
number and complexity of the classes.
LAURIN, G. V.; LIESENBERG, V.; CHEN, Q.;
GUERRIERO, L.; FRATE, F. D.; BARTOLINI, A.;
COOMES, D.; WILEBORE, B.; LINDSELL, J.;
VALENTINI, R. Optical and sar sensor synergies
for forest and land cover mapping in a tropicals ite
in west africa. International Journal of Applied
Earth Observation and Geoinformation, v. 21,
p. 7–16, 2013.
LU, D.; BATISTELLA, M.; MORAN, E. Land-cover
classification in the Brazilian Amazon with the
integration of Landsat ETM+ and RADARSAT
data. International Journal of Remote Sensing,
v. 28, p. 5447–5459, 2007.
LU, D.; LI, G.; MORAN, E.; DUTRA, L.; BATISTELLA,
M. A comparison of multisensor integration
methods for land cover classification in the
Brazilian Amazon. GIScience & Remote Sensing,
Taylor & Francis, v. 48, n. 3, p. 345–370, 2013.
MILLARD, K.; RICHARDSON, M. On the importance
of training data sample selection in random forest
image classification: A case study in peatland
ecosystem mapping. Remote Sensing, v. 7, n. 7,
p. 8489–8515, 2015.
The parameters “ntree” and “mtry” that produced
the lowest error in the RF calibration were 900 trees
and 11 variables at each split node, respectively. This
result highlights the importance of the previous step
of RF calibration to find the optimal parameters
for land use classification. The most important
RF variables, according to the mean decrease in
accuracy of the variables, were the reflectance of
band 5, followed by the EVI. For PALSAR-2, the
most important attributes were the GLCM metrics
of contrast for the HV and HH polarizations,
respectively.
Conclusions
The synergistic use of OLI and PALSAR-2 attributes
improved the classification of a complex transition
landscape, formed by a large number of classes
(17). The most important optical attributes for RF
classification were the reflectance of band 5 and the
EVI, whereas the most important SAR metric was
the GLCM contrast for the HV and HH polarizations.
81
Universidad Mayor
Development of Methods
Tandem-L Global Observation of forests with Two L-Band SAR
Satellites - Tandem-L
Andreas Huth1, Rico Fischer1, Friedrich Bohn1, Nikolai Knapp1, Sebastian Lehmann1, Sebastian Paulick1, Edna Roedig1,
Peter Biber2, Kostas Papathanassiou3
2
1
Helmholtz Centre of Environmental Research – UFZ, Leipzig, Germany
Technical University of Munich, Chair of Forest Growth and Yield Science, Germany
3
German Aerosopace Center (DLR), Microwaves and Radar Institute, Germany
Tandem-L is a proposal for a highly innovative
L-band SAR satellite mission for the global observation of dynamic processes on the Earth’s surface
with hitherto unparalleled quality and resolution (10
m). Main mission goals are the global measurement
of 3-D forest structure and biomass for a better understanding of ecosystem dynamics and the carbon
cycle, and high-resolution measurement of variations in soil moisture close to the surface. In addition the satellite mission can be used for systematic
recording of deformations of the Earth’s surface for
earthquake research (with millimeter accuracy) and
quantification of glacier movements and melting
processes. For this mission different bio/geo-physical information products have been developed and
evaluated based on a larger number of field campaigns in the HGF Alliance “Remote Sensing and
Earth System Dynamics” (EDA) .
82
The presentation will give an overview on the Tandem-L project and main results of the field campaigns concerning forest structure and biomass.
Based on the German forest inventory (including
9000 forest plots, 4 km grid) and remote sensing
data (L-Band Radar, Lidar) we calculated two indices describing the vertical and horizontal structure
of forests (16 classes). Forests in Germany feature a
low heterogeneity in vertical structure and a high diversity concerning horizontal structures. Classification by remote sensing can be compared to groundbased classifications. For 89 % (57 %) of the forest
plots the correct vertical (horizontal) structure type
could be predicted by remote sensing. In addition,
we show that height-biomass relationships for forests can be improved by including forest structure
indices.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Temporal Albedo Dynamics in Boreal Forest Fire
Scars Using Higher Resolution Albedo Products from
Landsat and Sentinel 2A
Angela M. Erb 1, Zhan Li1, Yan Liu1, Yanmin Shuai1, Qingsong Sun1, Crystal B. Schaaf 1, Zhuosen Wang 2, Ian Paynter1
2
1
School for the Environment, University of Massachusetts Boston, Boston, MA, USA
NASA Postdoctoral Program Fellow, Goddard Space Flight Center, Greenbelt, MD, USA
Keywords: Albedo, Landsat, Sentinel-2, Boreal
The Landsat Albedo Product provides higher
resolution albedo values by coupling 30m Landsat
surface reflectances with concurrent 500m
Bidirectional Reflectance Distribution Functions
(BRDF) Products from the Moderate Resolution
Imaging Spectroradiometer (MODIS) (Shuai et al.,
2011, 2014, Wang et al 2015). This approach has
been extended to generate a 20-m resolution albedo
product from Sentinel-2A MultiSpectral Instrument
(MSI) reflectance bands with configurations
similar to Landsat sensors. In addition, the 12bit radiometric fidelity of both the Sentinel-2A
and Landsat-8 satellites allows for the inclusion of
high-quality, unsaturated surface reflectance and
albedo calculations over snow covered surfaces.
The validation of these products over spatially
representative tower sites has shown the albedo
values from both Landsat-8 OLI and Sentinel-2A
MSI agree well with the ground measurements,
with snow and snow-free RMSEs for OLI of 0.0426
(n=28) and 0.0191 (n=89) respectively (Wang et
al., 2015) and a snow-free RMSE of 0.004 (n=4)
for MSI (Schaaf et al., 2016). MSI snow albedo
is under assessment and will be presented at
ForestSAT 2016. The increased temporal resolution
provided by multiple instruments allows a better
understanding of albedo dynamics in historically
difficult imaging locations.
Figure 1: Validation of the Landsat-8 Blue Sky Albedo Product at six tower sites in North America. The
overall RMSE for the combined product is 0.0267
83
Development of Methods
North American boreal forest fires have increased
in both frequency and severity over the past four
decades. This trend is expected to continue as
surface temperatures continue to rise and fire
season lengthens. As such, understanding both
the immediate and long term effects of these fires
on radiative forcing and the carbon budget will
become increasingly more important. We focus
this investigation on the seasonal variation within
several recent boreal forest fires scars in Alaska
and central Canada. We examine the impact of
landscape heterogeneity and fire severity on the
albedo in the growing and dormant seasons and
quantify the effects of over story canopy loss and
changes in snow exposure on radiative forcing.
Recent work on early spring albedo of fire scars
has illustrated that both burn severity and early
spring albedo show significant post-fire spatial
heterogeneity at the landscape scale and highlights
the need for a continued high temporal and spatial
resolution data record. This synergistic approach,
combining multiple data sources to create a more
comprehensive data record of land surface albedo,
will enhance our understanding and projection of
the carbon cycling and radiative forcing of these
spatially heterogeneous and temporally dynamic
ecosystems under a changing climate.
References
Schaaf, C.B., Li, Z., Liu, Y., Wang, Z., Erb, A.M.,
Sun, Q., Shuai, Y., Paynter, I.L. (2016). The
Use of Higher Resolution Albedo Product from
Landsat and Sentinel 2A to assess Landscape
Heterogeneity and Temporal Albedo Dynamics.
International Living Planet Symposium.
Shuai, Y., Masek, J. G., Gao, F., & Schaaf, C. B. (2011).
An algorithm for the retrieval of 30-m snow-free
albedo from Landsat surface reflectance and
MODIS BRDF. Remote Sensing of Environment,
115(9), 2204–2216. http://doi.org/10.1016/j.
rse.2011.04.019
Shuai, Y., Masek, J. G., Gao, F., Schaaf, C. B., & He, T.
(2014). An approach for the long-term 30-m land
surface snow-free albedo retrieval from historic
Landsat surface reflectance and MODIS-based
a priori anisotropy knowledge. Remote Sensing of Environment, 152, 467–479. http://doi.
org/10.1016/j.rse.2014.07.009
84
Universidad Mayor
Román, M. O., Schaaf, C. B., Lewis, P., Gao, F., Anderson, G. P., Privette, J. L., … Barnsley, M. (2010).
Assessing the coupling between surface albedo
derived from MODIS and the fraction of diffuse
skylight over spatially-characterized landscapes.
Remote Sensing of Environment, 114(4), 738–760.
Wang, Z., Erb, A.M., Schaaf C.B., Sun, Q., Liu, Y.,
Yang, Y., Roman, M.O., Shuai, Y., Casey
K. (2015). Early spring post-fire snow albedo dynamics
in high latitude boreal forests using Landsat-8 OLI
data. Remote Sensing on Environment. In press.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Terrestrial LiDAR and 3D Reconstruction Models for
Estimation of Large Individual Tree Biomass in Tropics
Alvaro Lau Sarmiento1,3*, Jose Gonzalez de Tanago1,3, Harm Bartholomeus1, Martin Herold1, Pasi Raumonen2, Valerio
Avitabile1, Christopher Martius3, Rosa Goodman4 and Solichin Manuri5
Laboratory of Geo-Information Science and Remote Sensing, Wageningen University, Wageningen, the Netherlands
2
Tampere University of Technology, Tampere, Finland
3
Center for International Forestry Research – CIFOR, Bogor, Indonesia
4
Yale School of Forestry and Environment Studies, New Haven, United States of America
5
Fenner School of Environment and Society, The Australian National University, Canberra, Australia
*
Email correspondent author: alvaro.lausarmiento@wur.nl
1
Key words: terrestrial laser scanning, 3D Reconstruction Models, quantitative structure models,
aboveground biomass, large trees, tropical forest
Abstract:
Tropical forest biomass is a crucial component on the estimation of carbon emissions in context of
REDD+. However, calibration and validation of such estimates requires accurate and effective methods
for in-situ estimation of above ground biomass – AGB. The most common method uses allometric
equation to indirectly estimate AGB from tree parameters as diameter, species and height. This
approach has been reported as major source of uncertainty in tropical large trees. On the other hand,
terrestrial LiDAR (Light Detection and Ranging) has demonstrated to be more accurate to infer forest
AGB in a non-destructive and more direct approach. This method has been applied and validated in an
open Eucalyptus in Australia. Nevertheless, application on tropical forest trees still has its challenges,
mostly due to occlusion, tree structural complexity and large scale application. We propose a method
to estimate AGB from individual tropical trees by estimating tree volume from terrestrial LiDAR point
clouds.
Nine plots of 30 x 40 m were scanned with a Riegl VZ-400 terrestrial laser scanner – TLS following a
spatial grid covering three study sites (Peru, Indonesia and Guyana). We identified the largest tree
per plot, extracted its point cloud and calculated tree wood volume by modelling 3D tree architecture
using quantitative structure models (TLS-QSM method). To validate our method, we harvested the
scanned trees, took detailed measurements of stems and all branches up to 10cm and reconstructed
tree volume as well. Then, tree wood volume was converted to AGB using species-specific wood
density values. To compare TLS estimates with present methods, we estimated AGB using
85
Development of Methods
Universidad Mayor
The Global Ecosystems Dynamics Investigation:
Current Status
Spaceborne lidar has been identified as a key
technology by the international ecosystem science
community because it enables accurate estimates of
canopy structure and biomass and forms the basis
for fusion approaches with existing and planned
missions, such as the NASA's ICESat2, ECOSTRESS
and OCO3 missions, and extends the capabilities
of radar missions such as the NASA-ISRO SAR,
Tandem-X and the ESA BIOMASS missions. The
Global Ecosystems Dynamics Investigation (GEDI)
is a space-based lidar instrument scheduled for
launch in late 2018. From its vantage point on the
International Space Station, GEDI will provide highresolution observations of forest vertical. These data
will be used to address three core science questions:
What is the aboveground carbon balance of the
land surface? What role will the land surface play in
mitigating atmospheric CO2 in the coming decades?
86
How does ecosystem structure affect habitat quality
and biodiversity? GEDI informs these science
questions by making billions of lidar waveform
observations per year. These canopy measurements
are then used to estimate biomass and in fusion with
radar and other remote sensing data to quantify
changes in biomass resulting from disturbance and
recovery. GEDI further marries ecosystem structure
from lidar with ecosystem and habitat modeling
to evaluate the impact of changes in land use and
climate on carbon sequestration and biodiversity. In
this talk we present an overview of the GEDI mission
and its current implementation status. We first
review its major science objectives and planned data
sets. We then summarize GEDI algorithms and our
approach to calibration and validation. Lastly, we
provide the status of the instrument hardware build,
as well as expected technical performance details.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
The Potential of Multitemporal and Polarimetric Features
Derived From Sentinel-1 Data for Forest Parameters Retrieval
Ziolkowski Dariusz1, Hoscilo Agata1, Lewandowska Aneta1, Sterenczak Krzysztof2
Institute of Geodesy and Cartography, 2Forest Research Institute
1
Keywords: Sentinel-1, Above_ground biomass, multitemporal analysis, polarimetry
The aim of the study is to show potential of
multitemporal and polarimetric analysis of Sentinel-1
data for forest parameters retrieval within various
forests of temperate zone. The C-band SAR data are
not commonly used for forest parameters retrieval,
especially for biomass studies, due to the saturation
effect of backscattering coefficient even for low
biomass values. The launch of Sentinel-1 data with
12-day’s temporal resolution gives the possibility for
new methodological approaches to overcome this
problem, especially using multitemporal analysis of
the data.
The research was carried out on two study areas in
Poland: Supraśl and Pieńsk. There are various types
of forests on both test sites: deciduous, mixed and
coniferous, each of them located on fresh and wet
habitats. Two sets of ground truth data were used:
data from Forest Digital Maps which offers many
forest parameters determined for forest plots and
field measurements for 500 sample plots of 25m
radius for each study area. For each of them many
various forest parameters were available i.e. tree
height, DBH (diameter at breast height), stock
volume, AGB (Above-ground biomass), density and
species. Meteorological data were also collected for
the whole period.
The set of 30 dual-polarization (VV and VH) successive
Sentinel-1 images acquired in 2015 were used.
The data were co-registered, calibrated and then
divided into two sets of images (trees with leaves
and without leaves). Then each of sets of images
was filtered separately using multitemporal filters.
Then various temporal statistics were generated for
both sets of the data. Independently polarimetric
processing of the data was also performed.
The change of backscattering coefficient during
the whole season and its correlation with forest
parameters (AGB, tree height and DBH) were
analyzed for the all sample plots together and
also separately for the particular forest types. The
analysis was performed in conjunction with the
variability of meteorological conditions. Next spatial
variability of temporal statistics for individual pixels
and its relation with forest parameters were studied.
The results of temporal analysis were compared
with spatial and temporal behavior of polarimetric
signatures.
First results show that the correlation of
backscattering coefficient with forest parameters is
changing a lot during the year. They are in accordance
with variation of hydrothermal index. The best
correlation of VH polarization with stem volume was
observed at the beginning of October 2015 during
the very dry conditions. The character and strength
of the correlation is different for particular forest
types, what can be seen in the best way in the case of
coniferous forests located on fresh and wet habitats.
Both of them are characterized by the strongest
correlations with backscattering coefficient but the
second one is negative. Great spatial variation of
multi-temporal features even within the same forest
types suggests big heterogeneity of forests which
is confirmed by spatial and temporal variability of
scattering mechanisms derived from H/A/Alpha
decomposition. Both methods have great potential
for improvements of the results.
87
Universidad Mayor
Development of Methods
Towards an “all-in-one sensor” for forestry
applications – estimating forest density, species composition
and biomass from stereo WorldView-2 data
Fabian Ewald Fassnacht1, Daniel Mangold1, Jannika Schäfer1, Hooman Latifi2
Institute of Geography and Geoecology, Karlsruhe Institute of Technology (KIT), Kaiserstraße 12, 76131 Karlsruhe,
Germany
2
University of Wuerzburg, Department of Remote Sensing in Cooperation with German Aerospace Center, Oswald-KuelpeWeg 86, 97074 Wuerzburg, Germany.
1
Keywords: WorldView-2, photogrammetric height information, very high resolution data, biomass, forest
density, tree species.
Over the last years, efforts were made to estimate
forest inventory attributes from airborne LiDAR,
multispectral and hyperspectral data. Although,
airborne data were found to be suitable for many
applications they are still commonly regarded as
being expensive compared to the cost of satellite
products. Since recently, the application of stereo
pairs of very high-resolution multispectral sensors
(VHRMS) such as WorldView-2 has been discussed as
an alternative to spatially estimate forest inventory
parameters. The VHRMS enable the simultaneous
collection of spectral and structural information via
stereo photogrammetric approaches.
kappa > 0.95 when considering only four classes
(broadleaved, coniferous, shadow and soil) and
kappa = 0.77 when considering 7 tree species. The
subsequent estimation of aboveground forest
biomass using forest type, density and additional
height information derived from a photogrammetric
point cloud returned reasonable results (r² = 0.58
obtained with Random Forest). The wall-to-wall
predictions (Figs. 1 + 2) agreed well with the expected
In this study, we used VHRMS data to spatially
estimate forest biomass with an approach that
imitates the field measured biomass estimations. In
a first step, the three key-parameters of forest type
(tree species), forest density (stem count per ha) and
height (as a proxy for DBH) were estimated from a
pair of WorldView-2 images acquired in 2013 over a
temperate central European forest in Germany. In
the second step, the obtained parameters served as
an input to a regression model of aboveground forest
biomass. Corresponding field data were collected in
three surveys during the period between 2013 and
2015.
Preliminary results of Random Forest models and
few predictor variables representing spectral,
textural and structural information resulted in
moderate to good estimates of forest density (r2 =
0.68) . Classification of forest types using Support
Vector Machines led to high accuracies, including
88
Fig 1: Forest density map.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
spatial distributions of stand densities and biomass
within the study area.
Further refinements of the described methodology
by integrating pre-knowledge on the relationships
amongst forest density, forest type and height in the
biomass models are foreseen. In addition, variables
describing the average crown diameters of the
reference areas will be integrated to further improve
the biomass models.
Our preliminary results suggest VHRMS data
as viable alternatives to airborne data to obtain
useful estimations of forest inventory parameters.
Further investigations, including more diverse forest
scenarios and varying sensor settings are required to
draw more consistent conclusions on the suitability
of the presented approach.
Fig 2: Biomass estimated for the study area.
89
Universidad Mayor
Development of Methods
Tropical forest degradation monitoring: Radiometric issues
of using both Landsat 8 and Sentinel 2 in one time series
Manuela Hirschmugl, Mathias Schardt, Roland Perko
Joanneum Research Forschungsgesellschaft mbH
8010 Graz, Austria
manuela.hirschmugl@joanneum.at
Viktoria Schell, Carina Sobe
Graz University of Technology
Graz, Austria
Keywords: forest monitoring, remote sensing, geometric and radiometric adjustment
Abstract
For continuous monitoring of forest disturbance and re-growth dynamics with high temporal and spatial resolution, a dense time series of data is needed. Sentinel-2 and Landsat8 are the sensors, which
are currently delivering optical data globally and free of charge. This study investigates, how data from
both sensors can be integrated in a time series analysis for forest degradation monitoring in tropical
areas in Malawi and Peru. Geometric inconsistencies exist and need to be eliminated before considering a time series classification approach. The differences could be removed by an automated image
matching procedure. Radiometric differences in the raw data are reduced by using surface reflectance
products (Landsat8) and Sen2Cor results (for Sentinel-2). The still remaining differences, mainly in the
NIR and SWIR band, could be corrected by relative radiometric adjustment. First analysis of time series
data show, that the magnitude of these remaining differences is much smaller than the magnitude of
forest changes. Therefore, data from both sensors can be jointly used in a time series classification.
Introduction
The aim of the EOMonDis project is to develop a
system which allows spatial-explicit, wall-to-wall,
continuous monitoring of forest disturbance and
re-growth dynamics with high temporal and spatial
resolution. In order to do so, a dense time series is
needed, which can only be achieved by employing
multiple sensors in tropical conditions with frequent
cloud cover. Although Sentinel 2 will – once the 2B
satellite is in orbit too – deliver images every 5 days,
there are still some limitations. First, in order to monitor changes, the natural variations need to be understood and these variations can only be observed
in the past, where Sentinel data had not yet been
available. Further, even with a high revisit rate such
as the one from Sentinel 2, there is still the matter
of cloud cover, which may be reduced by more frequent images taken from different sensors. Therefo-
90
re, the objective of this study is to integrate Landsat
8 and Sentinel 2 data in one time series data stack
for follow-up joint processing. The issues are investigated in two testsites: Malawi and Peru.
Methods
Geometric adjustment
While in some areas and test sites, the geometric
accuracy between Landsat 8 and Sentinel 2
orthorectified images is already sufficiently high,
others still show significant geometric shifts,
which have to be corrected before moving on
to investigating radiometric issues. In order to
correct geometric differences, a multi-modal
matching technique was used. The developed fully
automatic image matching procedure 1 using the
mutual information method (based on [2]) leads to
Universidad Mayor
ForestSAT 2016 Abstracts Summary
improved geometric congruence between the two
data sets. The specific matching strategy proposed
for this project starts with a pre-processing step that
upscales the Landsat data to the reference Sentinel
– 2 data. Next, the histograms of the filtered images
are compressed using a quantile processing. A lower
and upper threshold representing a given quantile of
the data distribution (e.g. 1% and 99%) are calculated
and used to compress the histogram to a given
number of values. This technique helps the following
up matching procedure as the joint histograms are
then better populated and thus a larger number of
correct matches are found. A compression to 64 bins
yields best results. The areal matching paradigm
is based on mutual-information maximization.
The main idea is that the joint entropy of two
image patches is minimized, when the patches are
correctly aligned. Therefore, a maximization of
the MI measure corresponds to a maximization of
clusters in the joint entropy and to a minimization
of the joint entropy’s dispersion. To get normalized
similarity measurements the entropy correlation
coefficient is used, which maps the normalized
mutual-information to the domain [0,1],
ECC(X, Y) = 2 −
2H(X, Y)
H(X) + H(Y)
(1)
with H(X) and H(Y) being the individual entropies and
H(X,Y) the joint entropy of image patches X and Y (cf.
[2]). In contradiction to e.g. the normalized crosscorrelation (NCC) method, which is just invariant to
linear mappings between the values of the images
to be matched, the MI method can also handle
non-linear dependencies. In our tests best results
are achieved with 151x151 pixel windows for MI
calculation. A sub-pixel measurement is achieved by
fitting a polynomial using the 3x3 neighborhood of the
entropy’s correlation peak and analytical calculation
of its maxima. Since, a rigid transformation between
the two datasets is assumed, it is sufficient to match
points on a regular grid (dense image matching
would be an overkill for this application). Using
these vectors, a rigid transformation is estimated
by solving an over- determined equation system
in the least-squares sense. The parameters of the
estimated transformation are then used to register
the Landsat data to the reference Sentinel - 2 data.
In the last step of the adjustment, a resampling from
30 to 10m resolution is performed.
Radiometric evaluation
Radiometric differences need to be investigated, as
the gray level differences between intact and degraded forests can be very small. For Landsat 8, the
surface reflectance product [3] has been used, processed before July, 1st, 2016, i.e. there are still some
artifacts in the resulting images, which are mainly
problematic in the blue band, which has not been
used in the current analysis. For Sentinel 2, both the
raw data as well as the result of the Sen2Cor atmospheric correction [4] are analyzed. For the radiometric analysis, reference areas, which have not been
visibly changed between the image acquisition dates, were selected from VHR data and Landsat / Sentinel images for both test sites. For the three classes: water, bare soil and pine plantations (Malawi)
and natural forest (Peru), areas were selected and
mean values obtained from two Sentinel images and
one Landsat image. Different statistical values were
calculates such as: Coefficient of correlation, mean
absolute difference and difference in percent of
the variation of each land cover class. These values
were calculated both for the comparison between
the two Sentinel images as well as between Sentinel
and Landsat. Relative radiometric adjustment was
applied in order to see, if the remaining differences
can be reduced. The regression coefficients for the
relative radiometric adjustment are calculated within areas which define so-called pseudo-invariant
features (PIFs). The program uses an automated process where PIFs are derived by utilizing a correlation
coefficient within a certain window size. This allows
to use images with remaining cloud areas without an
effect on the radiometric adjustment.
Finally, for the Peru test site, a time series involving
both Sentinel and Landsat images was build. Several intact forest areas are compared to degradation
areas during the same time period in order to analyze
the magnitude of difference between the sensors in
comparison to the magnitude of the change to be
detected.
Results
Geometric adjustment results
The automatic geometric adjustment worked very
well in both test sites. Fig. 1 shows the Sentinal-2
image on the left and the Landsat image on the right side for the Peru test site. In the top comparison,
a clear north-south shift is visible, which could successfully be corrected by our automatic adjustment
procedure (bottom).
91
Universidad Mayor
Development of Methods
Fig. 1. top: Geometric shift between Sentinel (left)
and Landsat 8, bottom: Landsat image adjusted to
Sentinel 2 geometry
Radiometric evaluation results - Malawi
For the three classes: water, bare soil and pine
plantations, areas were selected and mean values
obtained from two Sentinel images (26.12.2015 and
05.01.2016) and one Landsat 8 image (12.11.2015).
The results show that the raw Sentinel 2 data
sets are already radiometrically stable. For the
shorter wavelengths, the difference is higher due
to atmospheric influences than for the longer
wavelengths (R² = 0.83 in blue band vs. 0.93 in the NIR
band). After atmospheric correction with Sen2Cor,
the R² value changed only marginally. Comparing
the two Sentinel images, the mean of absolute
differences for all selected areas is higher in the raw
data (blue band: 1152; NIR band: 1871) compared to
the Sen2Cor result (blue band: 477; NIR band: 285).
Despite the differences in band width, the
compliance between Sentinel 2 Sen2Cor result of
2015 to the Landsat8 data showed an R² of 0.97 for
the NIR band and 0.73 for the blue band. The mean
of absolute difference for all selected areas is 344
(blue) and 423 (NIR).
In order to get a better evaluation of the magnitude
of these values, we calculated the variance for each
class (water, bare soil and pine plantations) solely in
the Sentinel reference scene from 05.01.2016 (Table
I – second column). This variance is defined as the
difference between the maximum and the minimum
mean value per polygon for each land cover class in
the reference data set. It shows that bare soil has the
highest variance in all bands, followed by water areas
and pine plantation. This is logical, as plantations
tend to be very homogeneous. In the next step, the
92
difference values per polygon between the Sentinel
(Ref) and the Landsat8 respectively Sentinel (In) were
calculated. Next, each of these difference values
is expressed in percent of its land cover’s variance.
Some polygons show a high difference, some a low
one. This led to the values shown in Table I - columns
three and four. The table reads: the difference
between the same water areas in Landsat8 image
and Sentinel reference image in the red band shows
between 5 and 88% of the difference within all water
areas in the Sentinel reference scene alone. As long
as the values are below 100% it means that involving
an additional image would not add to the variance
of the land cover class. Considering this, it can be
seen that for the two Sentinel images, the values
remain below 100, while for Landsat, the differences
are higher. In case of the pine plantations, the
differences are about 20 % higher, for bare soil, it is
similar in red and SWIR, but higher in the NIR band.
The difference in water areas in the NIR and SWIR
bands are extremely high. One reason for these
larger differences may be found in the different time
of data acquisition, as the Landsat image and the
Sentinel reference image are almost two months
apart.
TABLE I.
DIFFERENCES PER LAND COVER BETWEEN SENTINEL &
SENTINEL AS WELL AS LANDSAT & SENTINEL FOR MALAWI
"
RED
NIR
SWIR
Variance per LC
(Sentinel
reference image
only)
Water: 877
Bare soil: 1929
Pines: 179
Water: 225
Bare soil: 1621
Pines: 2105
Water: 260
Bare soil: 3399
Pines: 774
Difference as % of Variance
"'" (In)
"&'
(26.12.2015)–
(12.11.2015) –
"'" (Ref)
"'" (Ref)
(05.01.2016)
(05.01.2016)
2 – 64 %
5 – 88 %
2 – 62 %
10 – 51 %
29 – 87 %
19 – 113 %
5 - 80 %
246 – 390 %
10 – 63 %
16 - 116 %
1 – 84 %
2 – 102 %
6 – 33 %
188 – 282 %
15 – 39 %
9 - 44 %
13 – 80 %
31 – 107 %
Radiometric evaluation results - Peru
In the Peru testsite, also two Sentinel 2 scenes and
one Landsat 8 scene were used for the radiometric
comparison. Sentinel data were acquired on
11.12.2015 and on 10.01.2016. The Landsat 8
scene was acquired on 06.12.2015. Thus the time
difference between Landsat and Sentinel is only six
days. The differences in percentage of the variance
per land cover are shown in Table II (same calculation
method as for Table I). The values indicate that the
time difference between the image acquisitions is as
Universidad Mayor
ForestSAT 2016 Abstracts Summary
important as the sensor. Also there is a difference in
the different bands: while the red band of Landsat
and Sentinel corresponds very well, the NIR band
is worse. The SWIR bands show similar difference
values for Landsat-Sentinel as for Sentinel-Sentinel.
After application of relative radiometric calibration,
the values change in different band differently.
While the red band becomes worse, especially in
forest areas, the NIR values are improved. Based on
this statistics, no definite recommendation on the
improvement by relative radiometric adjustment
can be given.
TABLE II.
DIFFERENCES PER LAND COVER BETWEEN SENTINEL &
SENTINEL; LANDSAT & SENTINEL AND LANDSAT CALIBRATED & SENTINEL
FOR PERU
"
RED
NIR
SWIR
Variance
per
LC
(Sentinel
reference
image
only)
Water: 86
Bare soil:
735
Forest: 21
Water: 39
Bare soil:
1060
Forest: 544
Water: 159
Bare soil:
1329
Forest: 222
Difference as % of Variance
"&'
"'" (In)
(06.12.2015) –
(10.01.2016)–
"'" (Ref)
"'" (Ref)
(11.12.2015)
(11.12.2015)
forest areas, because the spectral difference in a
change area is of a much larger magnitude than the
difference by the sensors. It also shows that relative
radiometric adjustment improves the continuity of
the data in the time series. This will be even more
important, if the degradation features are not as
clear as in Peru, where degradation is basically small
patches of deforestation (smaller than the minimum
mapping unit for deforestation and therefore
considered as degradation). In other tropical areas,
other degradation patterns such as selective logging
occur. They are more difficult to detect. In such cases,
a consistent time series is a prerequisite. In general,
integration of more scenes and testing in other areas
has yet to be applied to draw a final conclusion.
"&'
(06.12.2015) –
"'" (Ref)
(11.12.2015)
184– 265%
15 – 82 %
9 – 120 %
5 – 58 %
9 – 117 %
93 – 147 %
31 – 197 %
382 - 489 %
36 – 90 %
5 – 57 %
370 – 581 %
68 - 121 %
303 – 363 %
431 – 730 %
42 - 93 %
3 – 34 %
73 – 145 %
46 – 96 %
29 – 115 %
19 – 140 %
35 - 92 %
3 – 56 %
7 – 44 %
52 -115 %
5 – 37 %
5 – 66 %
2 – 46 %
Time series results
In order to apply the values in a time series, a first
attempt was performed in the Peru test site for
intact forests and changed areas. Normalized
difference infrared index 7 (NDII7) was calculated
for all images and printed in a timeline. The time
series consists of five Sentinel 2 images (22.10.2015,
11.12.2015; 10.01.2016; 10.03.2016; 18.06.2016)
and three Landsat scenes (06.12.2015; 15.06.2016;
01.07.2016). The timeline without additional relative
radiometric adjustment for the Landsat scenes is
plotted in Fig.2. It can be seen, that the values of
the Landsat image show clear differences from
the Sentinel time series, but the magnitude of the
difference is smaller than the magnitude of the
change. Fig. 3 shows the same timeline with the
calibrated Landsat images. The timeline is much
smother, the sharp turns mostly gone or at least
smaller in magnitude. From this first analysis it
appears that classification from a combined time
series is possible to detect degradation features in
Fig. 2. Forest and changed areas in a combined
Landsat – Sentinel timeseries
Fig. 3. Forest and changed areas in a combined
Landsat – Sentinel timeseries with additional
radiometric calibration of Landsat images
93
Universidad Mayor
Development of Methods
Conclusion and Outlook
References
In a preliminary conclusion, the combination of
Landsat8 surface reflectance product and the
Sentinel-2 Sen2Cor result in one time series is
possible. Remaining differences can be reduced
by application of relative radiometric adjustment.
Further work to be done consists of improved
cloud elimination, as the existing cloud masks from
Sentinel-2 and also Landsat 8 still do not cover all
cloud areas. These remaining cloud areas need to be
classified through the time series classification, as
manual delineation is too tedious and too expensive
for the large amounts of images to be processed.
Further, a temporal filtering approach to fill the
cloud gaps is already available and needs to be
applied before finally moving on to classification of
the change areas. This classification will be done in a
trajectory fitting approach [5].
[1] R. Perko, H. Raggam, K. Gutjahr, and M. Schardt.
Using worldwide available TerraSAR-X data to
calibrate the geo-location accuracy of optical
sensors. In IEEE International Geoscience and
Remote Sensing Symposium Proceedings, pages
2551–2554, Vancouver, Canada, 2011.
[2] J.P.W. Pluim, J.B.A. Maintz, and M.A. Viergever.
Mutual information based registration of medical
images: A survey. IEEE Transactions on Medical
Imaging, pages 986–1004, 2003.
[3] Department of the Interior U.S. Geological
Survey. PRODUCT GUIDE: PROVISIONAL
LANDSAT 8 SURFACE REFLECTANCE CODE
(LASRC) PRODUCT, version 3.2 edition, 2016.
[4] Uwe Mueller-Wilm. Sentinel-2 MSI - Level-2A
Prototype Processor Installation and User Manual,
2016.
[5] Manuela Hirschmugl, Martin Steinegger, Heinz
Gallaun, and Mathias Schardt. Mapping Forest
Degradation due to Selective Logging by Means
of Time Series Analysis: Case Studies in Central
Africa. Remote, 6 (1)(ISSN 2072-4292):756–775,
2014.
Acknowledgment
This project has received funding from the European
Union’s Horizon 2020 research and innovation
programme under grant agreement No 685761
(Project EOMonDis).
94
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Tropical forest height and structure estimation from
AfriSAR campaign PolInSAR data in Gabon
Marco Lavalle1, Ralph Dubayah2 and Lola Fatoymbo3
1
Jet Propulsion Laboratory, California Institute of Technology
2
University of Maryland
3
Goddard Space Flight Center
Keywords: Tropical forests, vertical structure, PolInSAR algorithms, radar tomography
Aim
The goal of this contribution is to present our first
results of canopy height and vertical structure
estimation
from
polarimetric-interferometric
radar (PolInSAR) data collected in Gabon during
the February 2016 AfriSAR campaign. The AfriSAR
campaign was a joint effort between NASA and the
European Space Agency (ESA) to acquire airborne
and field data over African tropical forests in support
to forthcoming spaceborne missions, including
NISAR, GEDI and BIOMASS. NASA acquired 39.6
flight hours of data with the JPL’s L-band UAVSAR
radar instrument and 32.4 flight hours of data with
the LVIS lidar instrument over 5 different sites in
Gabon for calibration, validation, and new algorithm
demonstration of various ecosystem science
products.
Materials
Here we focus on the UAVSAR dataset of 9 PolInSAR
images (~20 km x 40 km, 0.6m x 1.8 m level-1
resolution) acquired over the northern part of the
Lope National Park in Central Gabon to demonstrate
extraction of canopy height and vertical structure.The
gradient of forest biomass from the forest-savanna
boundary (< 100 Mg/ha) to dense undisturbed
humid tropical forests (> 400 Mg/ha) makes Lope an
ideal site for PolInSAR algorithm assessment. The
UAVSAR dataset is fully polarimetric and has been
acquired by incrementing the aircraft altitude by 20
m at each flight track in order to acquire data from
different look angles and resolve vertical structure.
Several LVIS lidar flights have been also conducted
contemporary to UAVSAR to retrieve canopy height
and structure. Field data are available from previous
field campaigns as well as from the current the 2016
campaign.
Methods
We apply model-based PolInSAR technique (Cloude
and Papathanassiou, 1998) and polarization
coherence tomography (PCT) technique (Cloude,
2006) to extract canopy and vertical structure of
tropical forests from UAVSAR data. In PolInSAR,
wave polarization (e.g, HH, HV, VV or arbitrary
combinations of these channels) is used to identify
distinct structural components of the forest
(understory, stem, foliage, etc.) and interferometry
is used to locate the vegetation component
along the vertical direction – thus providing a
measurement of 3D vegetation structure. PolInSAR
models such as the RVoG or RMoG models (Lavalle
and Khun, 2014; Lavalle and Hensley, 2015) allow for
direct estimation of canopy height. PCT technique,
instead, assumes arbitrary vertical structure with
enhanced flexibility in parameter retrieval at the
expenses of data processing.
Results
We will show maps of canopy height and vertical
structure estimated via PolInSAR and PCT
techniques from multi-baseline UAVSAR data. Error
on the radar-derived canopy height is expected to
be ~15-25% and will be accurately assessed against
LVIS lidar-derived canopy height. Radar-derived
vertical structure and lidar-derived vertical structure
will be compared to assess the relative differences
as well as explore complementarity that may lead
to new radar-lidar fusion algorithms. Field data will
be also considered for assessing the quality of the
95
Development of Methods
radar-derived structural parameters, as well as to
convert maps of radar-derived structural parameters
to above-ground biomass.
Cloude, S.R., “Polarization Coherence Tomography”
(2006), Radio Science, 41, RS4017.
Cloude, S. R., and Papathanassiou, K. P. (1998).
“Polarimetric SAR Interferometry”. Geoscience
and Remote Sensing, IEEE Transactions on, 36(5),
1551-1565.
Lavalle, M., and Hensley, S. (2015), “Extraction of
structural and dynamic properties of forests from
polarimetric-interferometric SAR data affected
by temporal decorrelation,” IEEE Transactions
on Geoscience and Remote Sensing, vol.53, no.9,
pp.4752-4767.
Lavalle, M., and Khun, K. (2014), “Three-baseline
InSAR Estimation of Forest Height,” Geoscience
and Remote Sensing Letters, IEEE, vol. 11, no. 10,
pp. 1737–1741.
96
Universidad Mayor
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Two Phase Assessment System for the Effective Monitoring of
Tropical Forests
Mathias Schardt, Forschungsgesellschaft Joanneum Research, Austria
Based on increasing demands for information on
forests new methods for effective forest inventories
in tropical countries are to be developed in order to
save costs and / or to increase accuracies. Research
in this field has been very dynamic in recent years
producing a wealth of new ideas and approaches
including state-of-the-art technologies such as
remote sensing, mapping as well as inventory
designs. The challenge, particularly in the context of
tropical and sub-tropical countries is how to make
such comprehensive assessments most cost-effective
while achieving desired international standards.
Behind this background a pilot project was initiated
by the Austrian Natural Resources Management and
International Cooperation Agency (ANRICA) and the
Surinamese Foundation for Forest Management and
Production Control (SBB) aiming at the development
of an effective forest inventory system. The proposed
approach is based on two main concepts: a)
comprehensive statistical sampling methodologies
(to dramatically reduce costs by meeting high
international accuracy standards at the same time)
and a two phase approach by combining aerial and
terrestrial data. The cost reduction is achieved by
substantially reducing field work due to a reduced
number of field plots to be measured.
The remote sensing phase was composed by CIR
aerial photos and 3D canopy models derived from
photogrammetry. Additionally a LiDAR campaign
will be carried out in autumn this year in order
to derive tree heights and digital terrain models.
Aerial based acquisition of remote sensing data
in countries without any local facilities for data
acquisition is very cost intensive since a special
airplane needs to be transferred to the country
which causes high additional costs. To overcome
this problem Joanneum Research has developed the
mapping system ADAM-C (airborne data acquisition
and mapping for Cessna) that can be mounted
on the wing strut of any Cessna 182 and 206 and,
therefore can be operated worldwide. The system
consists of a Phase One Camera (80MP), Riegl
Laserscanner 580, IMU (inertial measuring unit) and
a GPS system receiver. The terrestrial phase of the
inventory system comprises of terrestrial sampling,
whereas the sampling intensity was set up based
upon the national accuracy requirements for the
data collected on timber stocks and Above Ground
Biomass (generally set at no less than 5% accuracy).
That’s why for the pilot project, primarily data on
these two parameters was collected. The terrestrial
phase was carried out by measuring field plots
according to good forest inventory standards for
terrestrial field inventory defined by the FAO.
The proposed two phase sampling approach uses
parameters extracted from photogrammetric
models to be correlated to with AGB and timber
volume, thus helping to reduce the number of plots
which need to be measured in the field.The key
challenge for this investigation was to find the most
suitable parameters that can be extracted from the
aerial imagery. As a source for this investigation the
following basic data sets were used:
● CIR ortho image mosaic (20 cm resolution)
● DSM from multiple stereo images (20 cm
resolution)
● SRTM terrain estimation
● In future also LiDAR data
Overall the tested parameters can be grouped in the
following way:
● Vegetation
indices, Texture
parameters,
Forest canopy density, Volume of DSM – DTM
difference, Number of trees, Size of the crowns by
segmentation, Statistical features based on nDSM
Mathematical models combining airborne and field
data on above ground biomass were established and
statistical data analysis has shown that the number
of highly expensive ground plots can be substantially
reduced and, thus, costs of forest inventories
decreased.
97
Development of Methods
The presentation will concentrate on the methods for
the remote sensing based assessment of the forest
parameters described above, the correlation of
these parameters with terrestrial AGB assessments
(from terrestrial plot sampling) and on the selection
of the best performing parameters.
98
Universidad Mayor
Universidad Mayor
ForestSAT 2016 Abstracts Summary
UAV-Borne and Airborne Remote Sensing for Tree Disease
Symptom Detection
Magdalena Smigaj1, Rachel Gaulton1, Stuart L. Barr1, Juan C. Suarez2
1
School of Civil Engineering and Geosciences, Newcastle University, Newcastle upon Tyne, NE1 7RU, UK
Contact: m.smigaj@ncl.ac.uk
2
Forest Research, Northern Research Station, Roslin, Midlothian, EH25 9SY, UK
Keywords: stress detection, unmanned aerial vehicle, UAV, thermal, hyperspectral
Abstract
Climate change has a major influence on forest health by indirectly affecting the distribution and
abundance of pathogens, as well as the severity of tree diseases. Changing weather conditions may
also result in the introduction of non-native invasive pest species. The detection and robust monitoring
of affected forest stands is therefore crucial for allowing management interventions to reduce the
spread of infections.
Stress induced by an invasion of insects or onset of disease manifests itself in tree foliage, and
may result in a variety of changes to plant’s physiological processes. When a plant is under stress,
stomatal closure occurs to help reduce water losses and prevent the entry of microbes and host tissue
colonisation. This mechanism can cause an increase in leaf and canopy temperature.
Nevertheless, there has been limited research into the use of thermal remote sensing for tree health
monitoring as required high spatial resolution data is usually obtained with low temporal frequency.
Newly emerging technologies, such as unmanned aerial vehicles (UAVs), could supplement aerial data
acquisition, but sensor development is still in the early stages. This project investigates the use of
airborne and UAV-borne sensors for detection of disease symptoms, in particular low-cost UAV-borne
microbolometer thermal system for monitoring disease-induced canopy temperature rise.
The research is based in Queen Elizabeth Forest Park, Scotland, where research plots were established
in pine stands, exhibiting various stages of stress. Extensive structural measurements of sample trees
were collected, including visual estimation of red band needle blight infection level. Infection level
was expressed as percentage of total unsuppressed crown volume in 10% steps, where a score of 20
indicates that 20% of crown was diseased. To compliment this, a number of needle samples were
collected for leaf spectral reflectance measurements. These measurements accompany airborne
hyperspectral, thermal and LiDAR data, as well as a thermal UAV-borne imagery collected in 2014.
Initially, calibration of the microbolometer camera was performed in laboratory, revealing nonuniformity across the imagery, which was minimised using a two-point calibration technique. Further
laboratory trials involved altering the camera’s temperature throughout imaging. These indicated that
the imager maintains stable radiometric calibration across different temperatures, unless exposed to
very rapid changes (> 0.2 K/min).
The derived calibration parameters were applied to test datasets of UAV-borne imagery. These were
georeferenced by registration to a LiDAR-derived canopy height model. At canopy level, the comparison
of tree crown temperature recorded by the thermal camera suggests a small temperature increase
related to disease progression; indicating that UAV-borne cameras might be able to detect sub-degree
temperature differences induced by disease onset. The influence of acquisition timing on the signal
was tested with repeated flights over the survey plot during summer 2015 at different times of the day.
This paper will present results of UAV-borne thermal imaging for detection of disease-induced canopy
temperature increase, as well as preliminary analysis of the acquired airborne data.
99
Universidad Mayor
Development of Methods
Updating of lidar based forest attribute maps using digital
photogrammetry combined with the lidar data
H. Olsson1, N. Lindgren1, J. Bohlin1, J. Jonzen1, I. Bohlin1, E., Willén2, M. Nilsson1
Department of Forest Resource Management, Swedish University of Agricultural Sciences, Umeå, Sweden, firstname.
lastname@slu.se
2
Skogforsk, Uppsala, Sweden, firstname.lastname@skogforsk.se
1
Keywords: lidar, laser scanning, digital photogrammetry, forest attribute maps, updating
Abstract
The Swedish national land survey (Lantmäteriet) has scanned all of Sweden with airborne lidar. Based
on these data, the Swedish University of Agricultural Sciences (SLU) and the Swedish Forest Agency,
has made nationwide forest attribute maps, which are freely available over internet. These prediction
maps were made using the area based method and regression functions with national forest inventory
plot data as response variables (Nilsson et al., 2016). In addition, most large forest companies in Sweden
have made similar lidar based prediction maps from the same national lidar data, but using their own
reference plots and ancillary information from their existing forest maps. The existence of the national
lidar data has changed the practice in the forest sector, since essential variables such as stem volume,
basal area, mean stem diameter and Lorey’s mean height, can be automatically predicted with at least
as good accuracy as traditional methods based on field work and air photo interpretation. The national
lidar scanning did however start year 2009 and most forest areas were scanned during the years 2010 –
2012. There is no decision yet to repeat the national lidar scanning, but Lantmäteriet has an ambitious
air photo program. Point clouds from digital photogrammetry provides good height estimates, but
compared to lidar, less good estimates of the canopy density (Bohlin et al. 2016). Hence, for example
volume and basal area are estimated less accurately.
In this presentation, we will therefore present the results from a study were lidar data from 2010 and
2011 is combined with point clouds from digital air photos acquired year 2015. Among the evaluated
methods are the use the old lidar metrics in combination with the height difference obtained when
new photogrammetry data is added. The test area is 1 million ha area 100 km north Stockholm. The
estimations will be trained with national forest inventory plot data from 2015 and the validation will be
done for 100 stands which are field surveyed with 8 – 10 plots per stand. The products will be produced
in two versions: using only commonly available data which corresponds to a possible future nationwide
product; and using also company specific information, which corresponds to possible future products
to be used internally in the forest companies.
References
Bohling, J., Bohlin, I., Jonzén, J., Nilsson, M.
2016. National forest attribute map using
stereophotogrammetry of aerial images, the
national terrain model and data from the national
forest inventory. Submitted.
100
Nilsson, M., Nordkvist, K., Jonzen, J., Lindgren, N.,
Axensten, P., Wallerman, J., Egberth, M., Larsson,
S., Nilsson, L., Eriksson, J., Olsson, H. 2016. A
nationwide forest attribute map of Sweden using
airborne laser scanning data and field data from
the national forest inventory. Submitted.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Use of Hybrid Model-Based Inference with a
Sample of Lidar Measurements to Produce Gridded
Biomass Estimates
Paul L. Patterson*, Sean P. Healey, Göran Ståhl, Sören Holm, Steen Magnussen, Ralph O. Dubayah,
Steven Hancock, Hans-Erik Andersen
*Speaker: United States Forest Service, Statistician
Keywords: Space-based Lidar, Hybrid, Model-based Inference, Biomass
The forthcoming NASA GEDI (Global Ecosystem
Dynamics Investigation) mission will install a fullwaveform lidar instrument on the International Space
Station for the purpose of measuring global forest
structure. The resulting waveform data is expected
to be strongly correlated with aboveground forest
biomass, and one of the mission’s primary science
products will be a 1-km gridded biomass product.
Grid cell-level estimates must be accompanied by
formally estimated precision. Waveforms will be
collected in spatially discontinuous “footprints”
that will sample, instead of census, each 1-km cell.
Biomass will be modeled at each footprint using
relationships derived from sets of co-located field and
lidar measurements. GEDI’s spatially discontinuous
measurements, combined with the fact that biomass
will be modeled instead of measured at each
footprint, argues for methods based upon a hybrid of
design- and model-based inference.
their performance at the scale of 1-km grid cells has
not been thoroughly demonstrated. Two activities
are under way to assess such estimators for use with
GEDI waveforms. First, a simulation-based study
is investigating the general relationship between
estimator performance and variables such as the
size of an estimation unit and spatial autocorrelation
of model residual error. Second, an empirical
study is assessing proposed estimators using GEDI
waveforms simulated from small-footprint airborne
lidar data collected in six diverse sites in the United
States. This latter study addresses GEDI-specific
concerns such as density of instrument overpasses
and strength of the footprint-level biomass
relationship. Relevance of these studies extends
to estimation of biomass across irregularly shaped
areas (e.g. watersheds or countries), as well as to
other sensors that collect high-quality but spatially
discontinuous forest structure information.
Hybrid estimators (sensu Ståhl et al., 2016) have been
employed in large-area estimation problems, but
101
Development of Methods
Universidad Mayor
Use of partial-coverage UAV data as a sampling tool for large
scale forest inventories
Stefano Puliti*, Terje Gobakken, Liviu Theodor Ene, and Erik Næsset
Department of Ecology and Natural Resource Management, Norwegian University of Life Sciences,
P. O. Box 5003, NO-1432 Ås, Norway,
Stefano.puliti@nmbu.no, terje.gobakken@nmbu.no, liviu.ene@nmbu.no, erik.naesset@nmbu.no.
Keywords: Unmanned aerial vehicle; Structure-from-Motion; partial-coverage; hybrid;
large-scale forest inventory.
Abstract
Use of image based three dimensional data from unmanned aerial vehicles (UAV) has proven
effective for forest inventories; however limitations in the range of operations of UAV hinder their
use in large scale applications. Use of partial-coverage UAV data may increase precision of purely field
based estimates of forest resource parameters and offer a cost-effective alternative to wall-to-wall
acquisition. In this study, data from UAV collected in systematically distributed blocks and combined
with ground observations in a two-phase design with hybrid inference (UAVHYB) were used to estimate
mean volume and its standard error (precision) for a 7330 ha forest area of interest (AOI). Hybrid
inference (HYB) rests on probabilistic sampling of auxiliary data (the UAV data) and a prediction model
calibrated with field data that predicts volume across the areas covered with auxiliary data. In the
uncertainty analysis, the sampling variability of the UAV data collection is added to the uncertainty
induced by the prediction model. Because the field data need not come from a probability sample,
the method offers great flexibility in field data collection while it can benefit from the great value of
strongly correlated remotely sensed data to improve precision of estimates. The estimate of precision
of UAVHYB was compared with precision of estimates with from alternative inventory, in terms of design
(one-phase) and estimators (design-based and model-based). Relative efficiency (RE), calculated
as the ratio between the estimated variances of two alternative methods, was used as measure for
improvement in precision between two alternative methods. Additionally the study addressed the
flexibility of HYB inference by including external ground reference observations. The comparison
against a method where a one-phase, design-based estimate based purely on field data revealed that
the use of UAV increased the precision of the estimates (RE = 1.2). The alternative sampling design was
one-phase sampling and the alternative estimators were design-based (DB) and model-based (MB),
respectively. For DB only field-data (FIELDDB) was available while for MB also wall-to-wall airborne
laser scanning (ALS) data (ALSMB) were acquired. Relative efficiency (RE), calculated as the ratio
between the estimated variances of two different methods, was used as measure for improvement
in efficiency for one method over the other. Comparison of UAVHYB against FIELDDB revealed that
the use of the former four times more efficient than the latter (RE=4.4) when external observations
were used. This translates to a need for 4.4 times as many field plots under simple random sampling
for a FIELDDB estimate to be equally precise as the UAVHYB estimate. However, when including only
observations from the AOI the RE decreased to 1.2, indicating only a slight increase in precision. Also,
the comparison against the ALSMB revealed that for the latter the increase in efficiency compared to
UAVHYB was limited (RE = 1.6). The study also demonstrated that the precision under UAVHYB can be
improved when including additional field data from other inventories, highlighting the flexibility of
HYB. Cost estimates for each inventory approach were also compared and discussed, revealing that
UAV may be a cost-effective tool for landscape level forest resource assessment.
102
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Using LiDAR remote sensing for identifying suitable habitat.
Case: Darwin´s Fox
G. Carrasco1*, R. Briones2
Landscape ecology program, Ecosystemic management division, Bioforest S.A.
Fauna conservation program, Ecosystemic management division, Bioforest S.A.
1
2
Keywords: LIDAR, Camera-trap, Lycalopex fulvipes, Caramavida.
Abstract
The range of Nahuelbuta has an outstanding biogeographic importance and high endemism of flora
and fauna. In this range, the Valdivian temperate forest extends north along the maritime watershed,
while the sclerophyllous forest extends south along the continental slope and northern coastal plain.
The L. fulvipes that inhabits this habitat is considered at high extinction risk due to both demographic
and ecological factors, such as disease, predation by cougars or other fox species. Our study was
conducted in Caramávida using camera traps (N = 84) during October 2011-March 2012 to observe L.
fulvipes, this information was correlated with LIDAR data to generate: elevation model, forest height;
and the raw data were used to estimate the vertical vegetation coverage for seven different layers of
height, cover, leaf area index and vertical complexity index. Our results indicate the presence of 17
positive results with L. fulvipes watch. The presence data and vegetation analyzes indicate correlations
with tree cover larger than 20 meters high and with a high diversity of vegetation in the vertical profile
, which may propose the presence of sites with potential presence and corridors for this species. We
conclude that this methodology can generate highly accurate and relevant vegetation variables that
may provide some guidance regarding which are the areas where a species is potentially distributed
and the design of corridors that may enrich their habitat.
Introduction
The biogeographic situation of Nahuelbuta Range
and its geomorphological and climatic characteristics
have given rise to environmental heterogeneity
establishing a wide variety of habitats (Mardones
2005, Luebert & Pliscoff 2005). Its peaks higher than
1,200 masl are covered by Araucaria araucana and
Nothofagus pumilio forests, as well as wetlands. Both
of the areas protected by the State in Nahuelbuta
Range, the Nahuelbuta National Park (PNN) and
the Contulmo Natural Monument (MNC), besides
the protected area Piedra del Águila, are clearly
insufficient to preserve this biodiversity, due to their
small surface and location in high regions, above 600
masl (Contulmo with 82 ha) and above 1,000 masl
(Nahuelbuta with 6,800 ha) (Ibarra-Vidal et al. 2005,
Ortiz & Ibarra-Vidal 2005).
Lycalopex fulvipes (Martin, 1837) is an endemic canid
of Chile at high extinction risk (Macdonald et al. 2004,
Cofré & Marquet 1998). This species was originally
considered Vulnerable (Glade 1993) and then as
Endangered (MINSEGPRES 2007) and Critically
Endangered at global level by UICN (Jiménez et al.
2008). It is included in Appendix II of CITES, being
considered as one of the canids with most serious
conservation problems (Macdonald & Sillero-Zubiri
2004).
It was considered that its distribution was restricted
to the west part of the Great Island of Chiloé and
Nahuelbuta National Park located at approximately
600 km from Chiloé Island. This disjoint pattern is
now being questioned because of the recent finding
of individuals using camera traps in the Alerce
Costero National Park, Valdivian Coastal Reserve,
103
Universidad Mayor
Development of Methods
Oncol Park and Chanchan near Valdivia, as well as
the presence of a dead specimen in Lastarria, near
Gorbea (Farias et al. 2014, D´elía et al. 2013, Vilà et al.
2004, Medel et al. 1990, Jaksic et al. 1990, Jiménez et
al. 1990). This evidence leads to the hypothesis that
the Darwin fox distribution may be more continuous,
associated with the remaining coastal forests.
The biological information about this species is
scarce, it is found mainly in studies conducted in
Chiloé Island (Jiménez et al. 2008, Jiménez 2007,
Jiménez & McMahon 2004, McMahon 2002). On this
regard,Yahnke et al. (1996) describe home ranges and
they inform a population of 500 individuals for Chiloé
(Yahnke et al. 1996). Jiménez & McMahon (2004),
based on intensive capture inside the Nahualbuta
National Park (PNN), made a population estimate of
78 individuals by extrapolating from a density of 1.14
ind/km2. This estimate is based on captures in the
southeast section of the PNN (sectors: Pehuenco,
Piedra del Águila and Coimallín). This scenario has
become more complex because in general the areas
adjacent to the PNN have high degradation and
human impact levels (Armesto et al. 2010) turning
them unsuitable for the L. fulvipes to have a viable
population (Mella 1994, Shaffer 1981).
The species distribution is mathematically or
statistically associated with different independent
variables that describe the environmental conditions.
If it is so, this relationship is extrapolated to the rest
of the study area and then a value is derived for each
place which is usually construed as the presence
probability of the species in that spot. The “presence
probability” is, therefore, an abusive interpretation
of the environmental similarity measure which
should be construed, at the most, as a suitability
value for the species to develop. These models use
variables and among them the forest variables are
widely used; however, they are often estimated
categorically.
Using the LIDAR technology (Light Detection and
Ranging) and some process algorithms vegetation
variables can be generated (e.g. coverage, leaf area
index, vertical profile and height) with high precision.
Aerial LIDAR is a sensor installed in an airplane that
emits pulses while flying and these pulses hit an
object (e.g. bare ground, building, stone, vegetation,
water). Part of this energy is reflected by the ground
or the objects on the surface and this energy is
detected by the sensor.
104
The sensor calculates the distance to the ground or
object; and each one of these pulses is stored together
with coordinates by means of the differential GPS
system and the inertial navigation system installed
in the airplane, so that the position of a spot can be
determined in three dimensions with high precision
(Dubayah & Drake 2000). When these points hit the
vegetation, they can be intercepted at different
heights and if the pulse density is high (around 4
pulse/m2) the vertical structure of vegetation can
be determined very precisely by means of a set of
algorithms (Ko 2012, McGaughey 2007, McGaughey
2003).
The purpose of this study is to use a methodology
associated with the LIDAR technology in order
to generate variables intended to be used in the
biodiversity area and applied to endangered species
such as L. fulvipes.
Methodology
Study area and data
The study area is located in the Nahuelbuta Range
in the central zone of Chile, in the sector called
Caramavida, characterized by temperate forest
vegetation and the presence of a wide animal
diversity. The area is topographically steep and
scarped with 900 m mean height above sea level,
and the range going from 500 to 1200 m.
In 2010, flights were performed over the area among
the activities of a cartographic improvement project
of Forestal Arauco and the study area surface is
500 km2. The data were collected using the LIDAR
Optech sensor, the scanning angle was ± 15º and
the footprint was around 0.5 m. The final pulse
density was 3.5 pls/m2. The data were processed by
the owner company of the flight producing a highresolution digital elevation model (1x1m resolution),
surface model and orthorectified images (0.5x0.5
m resolution). The raw data were gathered in LAS
format including X, Y, Z coordinates and intensity.
Determination of distribution range
To estimate the quantity of stations, the monitoring
information of Quebrada Caramávida performed
by Forestal Arauco was used (Briones et al. 2011)
(Zuñiga 2012), based on which a minimum number
Universidad Mayor
of spots was determined in order to have significant
estimates of the parameters being studied (standard
error lower than 0.04; see Mackenzie et al. 2005),
using the native forest registry (CONAF-CONAMABIRF, 1999). Within each coverage, the stations
will be randomly distributed at a distance of 0.3
kilometres to promote spatial independence (based
on the home range radius by Jiménez 2007). The
proportion of sampling units in each coverage was
related with the size of each coverage, and the
design was balanced in order to minimize variance
in the results, which might occur in a coverage with
small surface.
A total of 84 sampling stations were installed being
made up by a camera trap. They were deployed in
micro-habitat conditions suitable for L. fulvipes,
as close as possible to the selected spot, but
considering the following restrictions:
300m
from houses, and
50m from trails and roads.
The cameras were installed at 0.5 m high and they
remained active during 15 consecutive days in each
station. Considering that detection rates could be
low for this species, baits will be used to attract them
(synthetic urine and jack mackerel). The camera
trap study was carried out in spring and summer
(from October 2011 to March 2012). Both seasons
are critical periods after winter for recovering the
energetic demands and reproduction, increasing
the carnivore activity (Muñoz-Pedreros et al. 1995,
Jaksic et al. 1990, Jiménez et al. 1990). On the other
hand, the previous study with cameras conducted
between 2009 and 2011 in the study area confirms
a higher activity of L. fulvipes in spring-summer,
compared to autumn-winter (Zuñiga 2012)(Fig. 1b).
Determination of core patches
With the elevation model (bare land) and the surface
model, the crown height model was calculated
using the difference between both surfaces (Fisk et
al. 2009). After that, the surface with a vegetation
height above 4 m was identified, in order to get the
core patches in the area; subsequently, the patches
with a surface larger than 100 ha. were isolated since
these are the areas that may potentially support a
sustainable habitat (Santos & Tellería 1998).
Additionally, a model of morphology patterns (Vogt
et al. 2007) was used to classify the patches in 7
shape classes in order to determine which patches
ForestSAT 2016 Abstracts Summary
are actual patches or which participate in other
connectivity functions.
Processing raw LIDAR data
The data were processes with the software program
FUSION (McGaughey 2007) and the Gridmetrics
algorithm. The resolution or cell size for calculation
was 20x20 m (container) since at lower resolution the
process tends to identify trees and generate gaps in
the vegetation coverage and therefore the coverage
estimate of the area cannot be determined. The
points intercepted at seven height ranges were
obtained (0 – 2, 2 – 4, 4 – 8, 8 – 12, 12 – 20, 20 – 32
and >32 m.), which are those used by the national
registry of native forest (CONAF-CONAMA-BIRF
1999).
The coverage (Cob) was estimated for each height
range as follows:
∑ xint veget stands for the pulses intercepted
by vegetation and ∑ xtotales is the total pulses
intercepted in that height range. Based on this,
the vegetation coverage for each layer could be
estimated. Besides, the total coverage for the entire
vertical profile was estimated.
The effective leaf area index (IAF) is calculated with
the Beer law equation:
The LIDAR system is classified as an active remote
sensing system, that is, it emits its own light source.
This characteristic means that vegetation can be
illuminated by means of pulses or infrared light
beams and comparing the light that is intercepted
with the light that reaches the forest ground it is
possible to apply the Beer law in order to estimate
the area where light was intercepted; and that area
is then the effective leaf area.
The vertical complexity index (ICV) is based on
diversity measurement indexes that measure the
heterogeneity within a specific system (Van Ewijk
et al. 2011) in this case, the vertical structure of
vegetation.
105
Development of Methods
Where HB is the total number of pulses in the
container and is the pulse ratio in the container at
height i.
Statistical analysis
Universidad Mayor
were found during the sampling period: L. fulvipes,
Lycalopex culpeus (Molina, 1782), Lycalopex griseus
(Gray, 1837), Puma concolor (Linnaeus, 1771),
Conepatus chinga (Molina, 1782), Galictis cuja
(Molina, 1782). The occupancy values, occupancy
corrected for detectability and detectability of L.
fulvipes show that the proportion of stations where
the species was detected was 11%, but the actual
proportion would reach 14% after correcting for
detectability.
There is a variety of models in order to predict the
potential habitat of species; however, a logistic
regression model was selected due to the nature
of the dependent variable and also because these
models can provide a probabilistic prediction of L.
fulvipes presence and in this way our analysis with
LIDAR can be prioritized. Additionally, unlike most
of the multivariate procedures, it does not require
variables to be normally distributed. The logistic
function is expressed as:
Where P is the estimated probability of occurrence
of an event, e is the inverse of the natural logarithm
and u is the linear model:
Where is the regression coefficients and is the
independent variables.
Nine variables based on surfaces were used:
understory vegetation coverage, mean, maximum
and modal height of vegetation, total coverage, leaf
area index, elevation and vertical complexity index.
Since there are different scales in the variables, some
of them were converted in order to generate a better
adjustment (elevt =elev/1000). A stepwise process
was performed to select the variables that generate
the best estimate of the L. fulvipes distribution.
The statistics used for the selection was the Chisquare test and the correct percentage of model
classification.
Results
From the 84 sampling stations consisting of camera
traps, 17 had positive results for the L. fulvipes
presence (Fig. 1b). A total of seven carnivore species
106
Figure 1: Study area in Caramávida zone located
at 30 km from Cañete. The grey zone is the area
that has raw LIDAR data and the hatched zone is
Caramávida, the area with high conservation value
belonging to Forestal Arauco. The points indicate
the 84 cameras installed in the study area and those
in red are the ones that detected the presence of L.
fulvipes.
The stepwise process carried out by means of
forward selection, backward elimination and
subset selection, gave as a result that the best
prediction model includes the following as main
variables: modal height, vertical complexity index
and elevation (Fig. 2). The chi-square value for the
model is 49.60 (Table 2).
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Figure 2: a) Modal height (m) which characterizes the
most frequent height, indicating the homogeneous
forest height. b) Vertical complexity index, derived
from diversity indexes that allow characterizing the
vertical structure of the forest. Closer to zero means
that only a few height layers have vegetation and
closer to 1 means that all the height layers have the
same quantity of vegetation or coverage, showing
that there is a high diversity in the vertical profile of
vegetation. c) Elevation above sea level.
These three variables were selected as the best
indicators or independent variables explaining the
presence of L. fulvipes.
The probability of occurrence increases as long
as there are core patches present and the most
important predictor variables are the modal height,
representing the most frequent height, that is the
most homogeneous height of the core patch, and
then the ICV variable that reflects the distribution
of vegetation at different heights; closer to zero
indicates areas where vegetation is concentrated at
certain heights, such as the case of forest plantation,
but when the value is near 1, vegetation is distributed
homogeneously in the entire vertical profile,
indicating the presence of forests with different
composition or age, which show a high diversity
in the zone. Subsequently, elevation indicates an
altitudinal gradient for the presence of L. fulvipes.
Table 2: Parameters derived from the best predictor
variables of the regression model.
The classification matrix showed that the model has
a high prediction rate (79.07%), which means that
from every 5 observation 4 are virtually successful,
also, the odds ratio was 16, indicating that there is a
high probability of occurrence or presence (16:1) of
L. fulvipes in the zones shown by the model (Fig. 3).
107
Development of Methods
Conclusions
The Caramávida cleft, Trongol and their surroundings
are the most relevant areas in Nahuelbuta, since
they still have primary and secondary forest
fractions. These native forests, preserved in different
degrees, constitute the laurifolia ecosystem in the
Nahuelbuta Range (Pauchard 2011). In its low areas
Gomortega keule (Molina, 1782) Baill., 1869 and
Berberidopsis corallina Hook, 1862, can be found,
and in its high areas Araucaria araucana (Molina) K.
Koch dominating the landscape. The Caramávida
ecosystems are considered as priority sites for
regional preservation by the Chilean governmental
bodies.
Generating reliable information about the
geographic distribution of species is one of the main
requirements to establish effective preservation
policies. However, after decades of taxonomic
and faunistic work, we only have approximate
data about the total of species that inhabit Chilean
lands and we do not have convincing information
that enables us to know the current distribution
of most species (Briones et al. 2012). These
insufficiencies become evident when we include
vegetable coverage in our study. For our study area
Universidad Mayor
we have generated information layers regarding the
understory vegetation coverage, mean, maximum
and modal height of vegetation, total coverage, leaf
area index, elevation and vertical complexity index.
This information gives us more variables making our
analysis statistically more robust.
The species distribution models are in full
development and expansion with new methods and
strategies for their treatment and interpretation.
Consequently, there is a large number of articles
building up with significant methodological and
theoretic contributions for the modelling of species
distribution (see Mateo et al. 2011). There are several
information restrictions for these models, such as
the lack of presence/absence data, cartography,
environmental variables. With this the predicting
capacity of the models is affected.
Acknowledgements
The authors thank Forestal Arauco centre zone for
their funding and support in the field. The authors
are also grateful to the researchers Patricio Viluñir,
Rodolfo Figueroa, Felipe Hernandez, Alfonso Jara,
Alfredo Zuñiga, Dario Moreira and Javier Cabello for
their support in the field and methodological inputs.
Figure 3: Map showing the
probability of L. fulvipes presence,
derived from the output of the
logistic model. The variables used
in the model were modal height,
ICV and elevation; the correct
classification
percentage
was
79.07%.
108
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Bibliography
GLADE, A. 1993. Libro rojo de los vertebrados
terrestres de Chile. Corporación Nacional
Forestal, Santiago, Chile. 68 pp.
ARMESTO, J.J., D. MANUSCHEVICHA, A. MORA, C.
SMITH-RAMIREZ, R. ROZZI, A.M. ABARZÚA &
MARQUET, P.A. 2010. From the Holocene to the
Anthropocene: A historical framework for land
cover change in southwestern South America in
the past 15,000 years. Land Use Policy 27 148–
160.
BRIONES, R., F. GARATE-FLORES & JEREZ, V. 2012.
Insectos de Chile nativos, introducidos y con
problemas de conservación. Guía de Campo. Ed.
Corporación Chilena de la Madera, Concepción,
Chile. 256 pp.
BRIONES, R. 2011. Diversidad de Carnívoros de
transectos con trampas cámaras en AAVC
Caramavida. Documento interno Forestal
Arauco. 30 pp.
IUCN 2012. IUCN red list categories and criteria,
version 3.1, 2nd edn. IUCN, Gland.
COFRÉ, H. & MARQUET, P.A. 1999. Conservation
status, rarity, and geographic priorities
for conservation of Chilean mammals: an
assessment. Biological Conservation 88: 53-68.
CONAF-CONAMA. 1999. Catastro y Evaluación
de Recursos Vegetacionales Nativos de Chile.
Informe Regional Décima Región. Proyecto
CONAF-CONAMA-BIRF. Santiago. Chile.
D’ELÍA, G., A. ORTLOFF, P. SÁNCHEZ, B. GUIÑEZ
& VARAS, V. 2013. A new geographic record of
the endangered Darwin’s fox Lycalopex fulvipes
(Carnivora: Canidae): filling the distributional gap.
Revista Chilena de Historia Natural 86:485–488.
FARIAS, A.A., M.A. SEPÚLVEDA, E.A. SILVARODRÍGUEZ, A. EGUREN, D. GONZÁLEZ,
N.I. JORDÁN, E. OVANDO, P. STOWHAS &
SVENSSON, G.L. 2014. A new population of
Darwin’s fox (Lycalopex fulvipes) in the Valdivian
Coastal Range. Revista Chilena de Historia
Natural 87: 1:3.
FISK, H., P. MAUS & DALE, S. 2009. Derived Lidar
products and forestry applications: Crawford
study area—White Mountain National Forest.
RSAC-0999-BRIEF3. Salt Lake City, UT: U.S.
Department of Agriculture Forest Service,
Remote Sensing Applications Center. 6 p.
IBARRA-VIDAL, H., C. SEPÚLVEDA, D. SAAVEDRA
& MALDONADO, E. 2005. Propuestas de
conservación de los bosques nativos en la
cordillera del Maule y Biobío (VII y VIII Regiones).
In Smith-Ramírez C., J. Armesto & Valdovinos,
C. eds. Historia, biodiversidad y ecología de
los bosques costeros de Chile. Santiago, Chile.
Editorial Universitaria. p. 617-631.
JAKSIĆ, F.M., J.E. JIMÉNEZ, R.G. MEDEL &
MARQUET, P. A. 1990. Habitat and diet of
Darwin´s fox (Pseudalopexfulvipes) on the
Chilean mainland. Journal of Mammalogy, 71(2),
246–248.
JIMÉNEZ, J.E. 2000. Viability of the endangered
Darwin´s fox (Pseudalopex fulvipes): assessing
ecological factors in the last mainland
population. Progress report for the Lincoln Park
Zoo Neotropic Fundation. Chicago, Illinois, USA.
68 pp.
JIMÉNEZ, J. E. & MCMAHON, E. 2004. Pseudalopex
fulvipes. In: Sillero-Zubiri, C., Hoffmann. M., &
Macdonald, D.W. (eds.). Canids: Foxes, wolves,
jackals and dogs. Status survey and conservation
action plan. IUCN/SSC Canid Specialist Group.
Gland, Switzerland and Cambridge, UK. Pg. 5055.
JIMÉNEZ, J. E. 2007. Ecology of a coastal population
of the critically endangered Darwin´s fox
(Pseudalopex fulvipes) on Chiloe Island, southern
Chile. Journal of Zoology, 271(1), 63-77.
JIMENEZ, J. E., YÁÑEZ, J.L., TABILO, E. & JAKSIC,
F. M. 1996. Niche-complementarity of South
American foxes: reanalysis and test of a
hypothesis. Revista Chilena de Historia Natural,
69: 113–123.
JIMÉNEZ, J.E. 1993. Comparative ecology of
Dusicyon foxes at the Chinchilla National Reserve
in northcentral Chile. Master of Science thesis.
University of Florida, Gainesville, FL.
JIMÉNEZ, J.E., LUCHERINI, M., & NOVARO, A.J.
2008. Pseudalopex fulvipes. IUCN Red List of
Threatened Species.Version 2010.4.<www.
iucnredlist.org>. Downloaded on 09 May 2011.
109
Development of Methods
JIMÉNEZ, J.E., MARQUET, P.A., MEDEL, R.G.,
& JAKSIC, F.M. 1990.Comparative ecology of
Darwin’s fox (Pseudalopex fulvipes) in mainland
and island settings of southern Chile. Revista
Chilena de Historia Natural, 63, 177-186.
JIMENEZ J.E., C. BRICEÑO, H. ALCAÍNO, P.
VÁSQUEZ, S. FUNK & GONZÁLEZ-ACUÑA D.
2012. Coprologic survey of endoparasites from
Darwin’s fox (Pseudalopex fulvipes) in Chiloé,
Chile. Archivos de Medicina Veterinaria 44: 93-97
KO, C., T.K. REMMEL, & SOHN, G. 2012. Mapping
tree genera using discrete LiDAR and geometric
tree metrics. BOSQUE 33(3):313-319.
LUEBERT, F. & PLISCOFF, P. 2005. Bioclimas de la
Cordillera de la Costa del centro-sur de Chile.
In: Historia, biodiversidad y ecología de los
bosques costeros de Chile (Smith, C., Armesto,
J. y Valdovinos, C. eds.), pp. 60-73. Editorial
Universitaria, Santiago.
SILLERO-ZUBIRI, C.,
M.
HOFFMANN
&
MACDONALD, D.W. (eds). 2004. Canids: Foxes,
Wolves, Jackals and Dogs. Status Survey and
Conservation Action Plan. IUCN/SSC Canid
Specialist Group. Gland, Switzerland and
Cambridge, UK
MCMAHON, E. 2002. Status and conservation of the
zorro chilote on mainland Chile. Progress report
for SAG and CONAF, Temuco, Chile.
MACKENZIE, D. I., & ROYLE, J. A. 2005. Designing
occupancy studies: general advice and allocating
survey effort. Journal of Applied Ecology, 42 (6),
1105-1114.
MARDONES, M. 2005. La Cordillera de la Costa:
caracterización físico-ambiental y regiones
morfoestructurales. En: Historia, biodiversidad y
ecología de los bosques costeros de Chile (Eds. C.
Smith-Ramírez, J.J. Armesto& C. Valdovinos), pp.
39-59. Editorial Universitaria, Santiago de Chile.
MCGAUGHEY, R. J. & CARSON, W.W. 2003. Fusing
LIDAR data, photographs, and other data using 2D
and 3D visualization techniques, In: Proceedings
of Terrain Data: Applications and Visualization
– Making the Connection, October 28-30, 2003;
Charleston, South Carolina: Bethesda, MD:
American Society for Photogrammetry and
Remote Sensing. pp. 16-24.
110
Universidad Mayor
MCGAUGHEY, R.J. 2007. USDA Forest Service,
FUSION manual, version 2.60.
MEDEL, R.G., J.E. JIMÉNEZ & JAKSIC, F. 1990.
Discovery of a continental population of the rare
Darwin´s fox, Dusicyonfulvipes (Martin, 1837) in
Chile. Biological Conservation, 51(1), 71–77.
MINSEGPRES (Ministerio Secretaría General de la
Presidencia). 2007. Decreto supremo 151/2007.
Oficializa primera clasificación de especies
silvestres según estado de conservación.
MUÑOZ-PEDREROS, A., J. RAU, M. VALDEBENITO,
V. QUINTANA & MARTINEZ, D. 1995. Densidad
relativa de pumas (Felis concolor) en un
ecosistema forestal del sur de Chile. Revista
Chilena de Historia Natural, 68: 501–507.
ORTIZ JC & IBARRA-VIDAL H. 2005. Anfibios y
reptiles de la cordillera de Nahuelbuta. En:
Smith-Ramirez C, JJ Armesto, C Valdovinos
(eds.). Historia, Biodiversidad y Ecología de los
Bosques Costeros de Chile. Editorial Universitaria,
Santiago, Chile. Pp 427-440.
SANTOS, T. & TELLERÍA, J.L. 1998. Efectos de
la fragmentación de los bosques sobre los
vertebrados de las mesetas ibéricas. MIMA,
Madrid.
SHAFFER, M.L. 1981. “Minimum population sizes
for species conservation”. BioScience (American
Institute of Biological Sciences) 31 (2): 131–134.
STOWHAS, P. 2012. Conflicto Entre Carnívoros
Silvestres y Campesinos en el sur de Chile, DVM
thesis. Universidad Mayor, Santiago, Chile.
MELLA, J. 1994. Conservación de mamíferos
en las áreas silvestres protegidas de Chile:
Representación y poblaciones viables. Ambiente
y Desarrollo, Vol X (3): 72-78.
VAN EWIJK, K. Y., P. M. TREITZ & SCOTT, N.A.
2011. Characterizing Forest Succession in
Central Ontario using Lidar-derived Indices.
Photogrammetric Engineering and Remote
Sensing 77:261-269.
VILÀ, C., J.A. LEONARD, A. IRIARTE, S.J. O´BRIEN,
W.E. JOHNSON & WAYNE, R. K. 2004. Detecting
the vanishing populations of the highly
endangered Darwin´s fox, Pseudalopex fulvipes.
Animal Conservation, 7(2), 147-153.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
VOGT, P., RIITTERS, K., IWANOWSKI, M.,
ESTREGUIL, C., KOZAK, J. & SOILLE, P. 2007.
Mapping Landscape Corridors Ecological
Indicators 7, issue 2, pp. 481-488.
YAHNKE, C. J., W.E. JOHNSON, E. GEFFEN, D.
SMITH, F. HERTEL, M.S. ROY, C.F. BONACIC,
T.K. FULLER, B. VAN VALKENBURGH & WAYNE
RK. 1996. Darwin´s Fox: A Distinct Endangered
Species in a Vanishing Habitat. Conservation
Biology, 10(2), 366-375.
ZUÑIGA, A. 2012. Ensamble de carnívoros de la
Cordillera de Nahuelbuta: ¿Condicionado por el
tipo de hábitat o las relaciones interespecificas?.
Tesis para optar al grado de Magíster en Ciencias,
mención Manejo, Producción y Conservación de
Recursos Naturales. Universidad de Los Lagos,
Chile. pp 106.
111
Universidad Mayor
Development of Methods
Why Your Next Mapping Project Will Probably Use Stacking
Sean P. Healey1*, Warren B. Cohen1, Zhiqiang Yang2, Todd Schroeder1, Noel Gorelick3, Gretchen Moisen1
Affiliations:
1-
*Speaker
United States Forest Service 2- Oregon State University
Ensemble prediction methods have become popular
for many classification tasks, including those related
to remote sensing of forest resources. Foremost
among these methods is a “bagging” technique
called Random Forests, which trains an ensemble of
decision trees with random subsets of the training
data, and uses a voting procedure to integrate
across output classifications. This approach is
thought to improve classification accuracy through
reduction of overfitting. An alternative approach
called “stacking” (or “stacked generalization,”) starts
with an ensemble of classifiers that use completely
independent methods instead of replicating the
same model with random subsets of training data.
Stacking is accomplished by comparing alternative
classification results with the training data to
explicitly weight ensemble members in producing a
meta-classification. Typically, parametric methods
such as logistic regression are used to identify the
classifier biases used in this process.
We suggest that alternative forest maps can be
effectively integrated by a stacking process governed
by a non-parametric model such as Random Forests
instead of a parametric approach. A Random
Forests integration rule eliminates the variableselection processes required with parametric
approaches, and theoretically would also improve
performance and robustness by reducing overfitting
112
3-
Google, Inc.
of ensemble member weights. This approach was
recently tested with several leading forest change
detection algorithms and a broad reference dataset
that included even very subtle and/or slow changes.
Judged against this definition of disturbance, which
is admittedly broader than the definition built into
some of the original classifiers, ensemble integration
with a Random Forests stacking rule outperformed
both the original maps and stacking with logistic
regression. Performance (rates of disturbance
omission and commission) also improved when
unrelated spatial products were added to the
ensemble, including: forest cover classifications,
topography, and pre- and post-disturbance Landsat
reflectance values.
Stacking provides a means to integrate the signal
carried by alternative maps in addressing complex
classification tasks, but as implemented here, it
also requires tremendous computing capacity and
the ability to make maps using several different
classification approaches. Increasingly popular
cloud-based remote sensing platforms are at once 1)
improving our ability to share mapping algorithms,
and 2) raising our capacity to run several mapping
processes in parallel. When implemented with a
Random Forests integration rule on a cloud-based
system, stacking can offer significant performance
gains in a variety of mapping tasks.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
3-D Model of a Mediterranean Tree-Grass
Ecosystem for Remote Sensing Applications
Pacheco-Labrador, Javier1, Gajardo, John2, Riaño, David11,3, Martín, M. Pilar1
Environmental Remote Sensing and Spectroscopy Laboratory (SpecLab), Institute of Economic, Geography and
Demography (IEGD-CCHS), Spanish National Research Council (CSIC), C/Albasanz 26-28, 28037 Madrid, Spain.
2
Facultad de Ciencias Forestales Universidad de Talca, Avenida Lircay S/N Talca, Chile.
3
Center for Spatial Technologies and Remote Sensing (CSTARS), University of California, Davis, One Shields Avenue, 139
Veihmeyer Hall, Davis, CA 95616, USA.
1
Keywords: Tree-grass, savanna, LiDAR, Remote sensing, BRDF, shadow
Abstract
Tree-grass ecosystems are mixed woody-herbaceous systems characterized by a low density and
disperse distribution of woody vegetation. Such heterogeneity in physiology, phenology and structure
represent a challenge for modelers of different communities. Remote sensing applications usually
need to determine the fraction of trees and grass canopies observed; whereas Proximal Sensing
systems have the challenge of measure separately woody and grass elements. This becomes more
difficult in the case of multi-angular hyperspectral systems dedicated to the retrieval of BRDF, since
usual kernel-based functions are no longer valid.
In this work we present a geometrical 3-D model of a Tree-grass ecosystem located in Majadas del
Tiétar, Spain. The site is subject of research from remote sensors, Eddy Covariance systems and
features multi-angular hyperspectral systems (AMSPEC-MED). The model is developed to contribute
to the abovementioned applications. The 3-D model was built from Airborne laser scanning (ALS)
data, provided by Spanish Program of Aerial Orthophotography (PNOA). TerraScan (Terrasolid Ltd.,
Finland) classified ground and tree canopy points to obtain a 0.5 m Digital Ground (DGM) and a
Surface (DSM) Model. Circular Hough transform-based algorithm identified single crowns in the DSM;
then an ellipsoid was fit to ALS returns identified in each crown. Additionally, terrestrial laser scanning
(TLS) data of some crowns and a voxel model were used to predict crown transmittance at different
illumination angles.
The model covered an area of 2,328 Km2, where tree density was ~ 21.6 trees\ha. It provides detailed
structure of tree crowns and grass surface, and an additional probability of transmission as a function
of observation/zenith angle. Current applications are the unmix of trees and grass BRDF from the
AMSPEC-MED system, estimation of the tree/grass/shadow fractions as observed from mid and coarse
resolution remote sensors and ecosystem shadow fraction estimation for biogeochemical modeling
and energy balances.
113
Development of Methods
Universidad Mayor
3D measurement of tree health using multispectral intensity
data from terrestrial laser scanners
Samuli Junttila1,2, Mikko Vastaranta1,2, Riikka Linnakoski1, Junko Sugano3, Harri Kaartinen2,4, Antero Kukko2,4, Markus
Holopainen1,2, Hannu Hyyppä2,5, and Juha Hyyppäb,4
1
Department of Forest Sciences, University of Helsinki, 00014 Helsinki, Finland.
Centre of Excellence in Laser Scanning Research, Finnish Geospatial Research Institute FGI, 02431 Masala, Finland.
3
Department of Biology, University of Turku, 20014 Turku, Finland.
4
Department of Remote Sensing and Photogrammetry, Finnish Geospatial Research Institute FGI, 02431 Masala, Finland.
5
Department of Built Environment, Aalto University, P.O.Box 15800, 00076 Aalto, Finland.
2
Keywords: Forest health, terrestrial laser scanning, LiDAR, multispectral laser scanning, leaf water
content, monitoring.
Abstract
Equivalent water thickness (EWT), a measure of leaf water content, and leaf chlorophyll content
(Cab) are important early indicators of tree stress caused by a variety of factors including pest insects,
pathogens and drought. Early detection of pest insects and pathogens is vital in reducing the damages
and costs of decreased tree growth and tree mortality. Recent investigations have explored the use
of backscatter intensity information from terrestrial laser scanners (TLSs) in order to obtain EWT and
Cab, but few studies have investigated the use of multiple wavelengths or TLSs.
Here we tested the ability of three-wavelength terrestrial laser scanning in detecting leaf water content
from Norway spruce seedlings (n = 90). To simulate a drought event, the seedlings were subjected to
different levels of watering during 8 weeks resulting in variation in tree health. During this period a
sub-sample of seedlings were randomly selected for laser scanning and destructive measurements of
EWT at 10 time intervals. The relationship between the measured EWT from needle samples and laser
intensity features, using 690 nm, 905 nm and 1550 nm wavelengths, were determined. The results
showed a relationship of R2 = 0.72 (RMSE = 0.006 g/cm2) between the measured EWT and the ratio
of mean backscattered laser intensity from 905 nm and 1550 nm wavelengths. In addition to EWT
measurements, needle samples from 30 seedlings were collected for analysis of chlorophyll a and b
content. These analyzes are under process and the results will be presented in ForestSAT 2016.
114
Universidad Mayor
ForestSAT 2016 Abstracts Summary
3D modeling of below-canopy global irradiance using
terrestrial LiDAR data and ray tracing
Renato Cifuentes La Mura1, 3, Dimitry Van der Zande2, Christian Salas3, Laurent Tits1, 4,
Jamshid Farifteh1, Pol Coppin1
KU Leuven, Department of Biosystems, Division of Crop Biotechnics, Willem de Croylaan 34,
BE-3001 Leuven, Belgium
2
Directorate Natural Environment, Royal Belgian Institute of Natural Sciences,
Gulledelle 100, BE-1200 Brussels, Belgium
3
Laboratorio de Biometría, Departamento de Ciencias Forestales, Universidad de La Frontera,
PO Box 54-D, Temuco, Chile
4
Flemish Institute for Technological Research (VITO), Remote Sensing Unit, Boeretang 200, BE-2400 Mol, Belgium
1
Keywords: canopy structure, 3D modeling, terrestrial LiDAR, ray tracing
Phase-based terrestrial LiDAR (PTL) measurements
were collected on two broadleaved forest to i) generate the 3D models of their canopies and estimate
leaf area index (LAI), clumping index (CI) and canopy
openness (CO), and to ii) feed a ray-tracing system to
analyze distribution of global irradiance, specifically
photosynthetically active radiation (PAR), below the
canopy. Data were collected on a pure beech (Fagus
sylvatica L.) and on a mixed (beech, oak (Quercus
robur L.) and birch (Betula pendula Roth)) forest. A
total of nine scans were taken per forest on a 400 m2
area using a FARO® LS880 PTL scanner (λ=785 nm,
maximum range=76 m). Digital hemispherical photographs were collected as reference data. Leaves
and trunks spectra were also recorded on a FieldSpec® 3 spectroradiometer. Similarly, global irradiance was measured in the field using a remote cosinecorrected receptor. This light measurements were
taken on every cross point of a 1m grid within the
400 m2 area. For three-dimensional (3D) canopy modeling, the PTL data was filtered (i.e., noise correction), classified in green and non-green elements,
and voxelized using four different voxel side lengths
(VS) i.e., 10 mm, 20 mm, 26 mm and 30 mm. A series
of forest scenes, emulating hemispherical images,
were then built using the open source Persistence of
Vision Raytracer (POV-Ray) for LAI, CI and CO calculation. These estimates were later compared with
the reference data. Likewise, the Physically-Based
Ray Tracer (PBRT) was used for light distribution
simulations and PAR analysis. Results after simple
linear regression modeling (α=0.01) indicate that
predicted values of canopy structural variables are
consistent with the observed ones at particular zenith angle ranges (ZAR) and parametrization for 3D
modeling. In the mixed forest, LAI was better predicted using VS 10 mm at 30°-60° ZAR, CI was not significantly influenced by VS but showed differences
in slope and intercept at different ZAR, and CO was
accordingly predicted using VS 10 mm at all ZARs. In
the beech forest, LAI predictions were consistent at
20°-60° ZAR using VS 20 mm, CI presented similar
trends as in the mixed forest but VS had an impact at
narrower ZAR (e.g., VS 20 mm at 40°-60°), while CO
was better predicted with several VS (e.g., 20 mm,
26 mm, and 30 mm) at both 5°-25° and 5°-30° ZAR.
Consequently, the parameters required for the suitable specification of the 3D model for light simulation
and analysis of PAR in PBRT were adopted after this
structural analysis. PAR results from both mixed and
beech forest reflected that predictions were biased
in relation to observed values. In the mixed forest,
predictions underestimated PAR up to a certain level, where the trend was then shifted to overestimation (RMSD = 0,23). On the other hand, predicted
values on the beech forest, systematically underestimated PAR (RMSD = 0,19).
115
Development of Methods
116
Universidad Mayor
Forest Mapping & Inventory
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
A web service proposal for forest inventory of fast-growing
species in areas with small-size property using free software:
GRASS GIS,WPS and LiDAR
Miguel Cordero , Eduardo Corbelle David Miranda
Land Laboratory – Department of Agroforestry Engineering, University of Santiago de Compostela, Escuela Politécnica
Superior, R/Benigno Ledo, Campus Universitario, 27002 Lugo, Spain
Miguel Cordero (miguel.cordero@usc.es)
Keywords (5-8), forest inventory, small property, lidar, WPS, free software
Abstract,
Estimating timber volume of very small stands (e.g. less than 2000 m2) of fast-growing species
by means of automated inventory techniques can be really challenging. Lack of adequate means
to easily estimate volume in those circunstances result in less transparent timber markets and
generally hampers planning and sustainable forest management. This is tipically the case of areas in
Northwestern Spain, where most property lots are usually very small. While methods for LiDAR-based
volume estimation in this geographic area have already been published (mainly for Pinus radiata D.
Don, and Eucaliptus globlulus Labill stands), their use has been largely restricted to research purposes
and they have only been used operationally in forest stands of medium to large area. In this work,
we outline the characteristics of a public web service based on free software that builds on existing
methods and publicly available LiDAR data to automate estimation of timber volume in small plots
of fast-growing species. The service would be implemented in GRASS GIS, use WPS protocol, and
build on public low-density LiDAR data, available filtering algorithms, and already published volumeassessment equations for P. radiata and E. globulus. We tested the system with real data from small
woodlots harvested in the same year of LiDAR data acquisition. A basic description of the proposed
design is presented, along with an assessment of the achieved degree of functionality, main obstacles
found, and an assessment of final error in volume estimation.
Introduction
Estimating timber volume of very small stands
(e.g. less than 2000 m2) of fast-growing species by
means of automated inventory techniques can be
really challenging. Lack of adequate means to easily
estimate volume in those circunstances result in less
transparent timber markets and generally hampers
planning and sustainable forest management. This
is tipically the case of areas in Northwestern Spain,
where most property lots are usually very small.
While methods for LiDAR-based volume estimation
in this geographic area have already been published
(mainly for Pinus radiata D. Don, and Eucaliptus
globlulus Labill stands), their use has been largely
restricted to research purposes and they have
only been used operationally in forest stands of
medium to large area. In this work, we outline the
characteristics of a public web service based on free
software that builds on existing methods and publicly
available LiDAR data to automate estimation of
timber volume in small plots of fast-growing species.
The service would be implemented in GRASS GIS,
use WPS protocol, and build on public low-density
LiDAR data, available filtering algorithms, and
already published volume-assessment equations for
119
Universidad Mayor
Forest Mapping & Inventory
P. radiata and E. globulus. We tested the system with
real data from small woodlots harvested in the same
year of LiDAR data acquisition. A basic description
of the proposed design is presented, along with an
assessment of the achieved degree of functionality,
main obstacles found, and an assessment of final
error in volume estimation.
Methodology, The service would be implemented
in GRASS GIS, use WPS protocol, and build on
public low-density LiDAR data, available filtering
algorithms, and already published volumeassessment equations for P. radiata and E. globulus.
We used data from two public sources; cadastral
plot and details of the lidar from PNOA (a natioanl
program of remote data acquisition). Both data sets
are freely downloadable and vector data supplied
high quality in terms of spatial accuracy. In the case
of lidar with low density points (0.5 points / m2).
As the tools we chose to use software for easy access
to internal documents and the gratuitous licensing
to implement the service; OpenLayers v3.0 as map
viewer, PyWPS v5.0 as implementation of the
protocol and Web Processing Service and GRASS
GIS v7.0 as computing GIS platform.
For the validation of the system results were
compared with values of timber companies that buy
wood in the area conducting field inventory plots.
In order to maximize the transparency of the
process, the system is based on data and public
access software and hardware.
For the estimation of timber in each plot, a GIS
system was prepared that from the polygon of land
cut the plot data, filter the corresponding points to
the terrain, and apply the existing volume estimation
equations to lidar data. This process was automated
from scripts in the GIS.
Through PyWPS was connected to the web server
so that the previous GIS script became available as a
web service hosted on a public server.
120
Through Open Layers an interface is provided to
the user to select the parcel of interest and to be
able to request its timber estimation. This interface
transmits to PyWPS the parameters that the service
needs (limit of the plot or zone to be measured).
Results, We tested the system with real data from
small woodlots harvested in the same year of lidar
data acquisition. In plots with raw data (without
information of age, forestry, etc.) the system can
estimate between 40-60 percent of the total value.
Taking into account that they are raw data can
be an important margin of improvement without
deteriorate user experience.
We also tested the system with real users, owners
of small plots, detecting that there may be a
problem, not of handling the tool but of continuity
or familiarity of use, since the measurements are
sporadic uses, with years without intermediate
use. Therefore, it may be necessary to take actions
towards their generalization in the sector.
References.
[1] E. Gonz ́alez-Ferreiro, U. Di ́eguez-Aranda, and D.
Miranda.
Estimation of stand variables in pinus radiata
d. don plantations using different lidar pulse
densities.
Forestry, 85(2):281–292, 2012.
[2] L. Gon ̧calves-Seco, E. Gonz ́alez-Ferreiro, U. Di
́eguez-Aranda, B. Fraga-Bugallo, R. Crecente,
and D. Miranda.
Assessing the attributes of high-density
eucalyptus globulus stands using airborne laser
scanner data.
International Journal of Remote Sensing,
32(24):9821–9841, 2011.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Above ground biomass related to field measurement errors
Trucíos C., R.1 ; Paudel, P.2, Kleinn, C. 3
PhD student Faculty of Forest Sciences and Forest Ecology, Georg-August-Universität, Göttingen.
M. Sc. formed in Tropical and International Forestry, Faculty of Forest Sciences and Forest Ecology,
Georg-August-Universität, Göttingen.
3
Prof. Dr. and Head of Chair of Forest Inventory and Remote Sensing, Faculty of Forest Sciences and Forest Ecology,
Georg-August-Universität, Göttingen.
1
2
Key words: measurement error, discrepancy, training, biomass
Measurement errors are generally not reported
in the dasometric data, in this case specifically
measurements in national forest inventories.
Measuring the field variables and know the
measurement error makes known the accuracy of
the measured variable and those estimates derived
variables expressing forest productivity as basal
area, volume, forest biomass, stored carbon, among
others. This document presents a case study in a
forests owned by the federal state of Lower Saxony,
north of Göttingen, Germany. In this study, we
compared the accuracy of above ground biomass
estimates developing skills through training teams
in field measurement variables (DBH, total height,
azimuth and horizontal distance). Similarly, the
accuracy between measurements of two instruments
for diameter and height variables used in the
estimation models of forest biomass was compared.
To analyze the random measurement errors, clear
protocols for the use of devices were established,
in addition to calibrated instruments used to avoid
deviations due to bias measurements. Part of the
analysis was to compare field measurements of the
two work teams: experienced and no experience in
dasometric measurement variables. After a training
period, the teams re-measured the plots and the
effects of training were analyzed on the accuracy
of the measurements. Measurements of the two
teams were made with the same instruments and
compared to a control data consisting of the average
of 5 mensurations (using the same instrument). The
second part of the analysis corresponds to the remeasurement of trees (Control two) with different
precision equipment and its comparison with Control
data. After second training, mean of the difference
in DBH measurement decreased for both groups,
-0.08 cm to 0.0002 cm for inexperienced and
-0.107 cm to 0.02 cm for experienced in first and
second measurements, respectively. However,
mean of difference in height measurement for
experienced observers was very low compared to
inexperienced observers in both measurements.
Large errors (outliers) in measurement of DBH in
both measurements were randomly distributed
regardless of tree size. A higher deviation was
noticed in measurement of height for taller
and bigger trees measured by inexperienced
observer in both measurement, however such
deviation only found in first measurement in case
of experienced observer. One of the parameters
used to compare control data was the relative
standard deviation of the analyzed variables which
was less than 6%, indicating acceptable accuracy
measurements. Using the law of error propagation
will be estimated uncertainty of the measured
variables and their impact on the estimation of
above ground biomass through allometric models
used in this study.
121
Universidad Mayor
Forest Mapping & Inventory
Above-ground biomass estimation using calibrated
multispectral aerial images in grasslands. Do calibration
targets matter?
Miguel Marabel, Flor Alvarez-Taboada*
*corresponding author. flor.alvarez@unileon.es
Affiliation of all the authors: GEOINCA research group. Department of Mining Technology, Topography and Structures,
University of León, Ponferrada Campus, Avenida de Astorga s/n, 24400, Ponferrada, León, Spain.
Key words: grasslands; vicarious calibration; spectro-radiometer; Partial least squares regression;
Ultracam XP/WA
Abstract
The main objectives of this research work were: (i) to establish a procedure for the radiometric
calibration of images from the airborne photogrammetric camera UltraCam-Xp WA and (ii) to
estimate the above-ground biomass (dry weight of the green fraction) in grasslands using those
calibrated images. The empirical line calibration and three multispectral images (A, B, C) were used.
Three different types of calibration targets were used: (i) 1 m x 1 m portable rigid targets (gray
scale), (ii) 2 m x 2 m non rigid targets made of fabric (gray scale and colour) and (iii) target areas (not
portable). 27 different combinations of these targets were tested to calibrate the images. The most
suitable calibrations were used to obtain the at-surface reflectance images (Ar, Br, Cr), which were
used to estimate the biomass. The calibration targets consisting of portable rigid panels provided the
most accurate radiometric calibration (R2=0.99; RMSE 0.014 %; RMSE (%)=3.64). The most accurate
estimation of above-ground biomass was obtained by using Partial Least Squares Regression (PLSR)
and at-surface reflectance of the Cr image (R2=0.90; RMSE=4.096 g/m2; RMSE (%)=12.92), although
the differences between images were not significantly different. Using the vegetation indices NDVI
(R2=0.80; RMSE=5.593 g/m2; RMSE (%)=18.81), SR (R2= 0.78; RMSE=5.864g/m2; RMSE (%)=18.49), NLI
(R2=0.88; RMSE=4,267g/m2; RMSE (%)=13.46) or SAVI (R2=0.85; RMSE=4.867 g/m2; RMSE (%)=15.35)
the estimates were less accurate.
When PLRS was used for the biomass estimation, the differences between using different calibration
sets ranged from 0.1% to 2.5% for the RMSE (%). However, when vegetation indices were used (NDVI,
SR, NLI, SAVI), the differences derived from using different calibration sets ranged from 1.6% to 19.2%
for the RMSE. NDVI and SAVI were the least sensitive to the changes in the calibration (<5% difference),
while SR was the most affected (10%-19% difference). These results were obtained in the tree images.
The calibration equations adjusted for image A and for image C were applied to image B, and the
differences in the biomass estimation after using those calibrations or the calibration obtained
specifically for B were smaller than 1% (RMSE%). Thus, the calibration of one image can be successfully
used for adjacent images.
The images of the UltraCam-Xp WA are therefore suitable for estimating the above-ground biomass
once they have been calibrated to at-surface reflectance, using PLSR and the spectral information of
the 4 bands of the image.
122
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Introduction
radiometric calibration of images from the airborne
photogrammetric camera UltraCam-Xp WA and (ii)
to estimate the above-ground biomass (dry weight
of the green fraction) in grasslands using those
calibrated images.
Above-ground biomass refers to the part of
vegetation which grows above the ground, and some
of it corresponds to grasslands. The importance of
grasslands is remarkable, since they cover 37% of
the terrain and they are able to capture CO2 and
transform it into biomass through photosynthetic
processes. It is also noticeable the role of grassland
biomass in silvopastoralism, as well as its energetic
value, since they are biofuels. Moreover, these areas
are crucial to protect soils from erosion. On the
other hand, digital aerial cameras are suitable to
record multispectral images, with high radiometric
and geometric accuracy, which need however to be
calibrated to extract qualitative and quantitative
thematic information (Honkavaara et al., 2004).
Along those lines, the estimation of biophysical
variables like biomass requires an absolute
radiometric calibration of those images, so that atsurface reflectance values are obtained (Jensen,
2005) and they can be used an input data in the
estimation models.
Therefore, the main objectives of this research
work were: (i) to establish a procedure for the
Methodology
The general workflow is showed in Figure 1. It starts
with the radiometric calibration of the multispectral
images using an empirical line calibration and
its validation, and it is followed by the biomass
estimation using the calibrated image and field
measurements, by means of reflectance values or
vegetation indices.
Radiometric calibration
The empirical line calibration and three multispectral images (A, B, C) were used. The images were
gathered with the metric aerial camera UltraCam-Xp
WA, with a spectral resolution covering the Red (R),
Green (G), Blue (B) and Near Infrared (NIR) regions
of the electromagnetic spectrum, a GSD of 18 cm,
and a radiometric resolution of 16 bits. The flight
was carried out in 23/07/2012, and the three adjacent
photograms were taken following a N-S flight-line.
Figure 1. General workflow.
123
Universidad Mayor
Forest Mapping & Inventory
To avoid possible bilinear reflectance effects, the
geometric center of the images was close to the calibration targets. The images were processed from level 1 to level 2 and orthorectified using ground control points measured with a GNSS receiver.
Three different types of calibration targets were
used: (i) 1 m x 1 m portable rigid targets (gray scale),
(ii) 2 m x 2 m non rigid targets made of fabric (gray
scale and colour) and (iii) target areas (not portable).
27 different combinations of these targets were tested to calibrate the images (Table 1). Each target was
radiometrically characterized by 30 measurements
done by using the ASD Fieldspec3 spectro-radiometer. The same procedure was followed to characterize thirty 1 m x 1 m plots, which consisted mainly
of Lolium perenne, Poa pratensis and Trifolium repens. The aboveground biomass (defined as the dry
weight of the green fraction in g/m2) was obtained as described
in Marabel &
Álvarez-Taboada (2013, 2014). The biomass
average value was 31.71 g/m2 ±12.63.
The digital numbers obtained for each target in the
multispectral image were compared with the at-surface reflectance values gathered with the spectro-radiometer. The difference in spectral resolution
of the data (4 wide multispectral bands for the la UltraCam-Xp WA versus 1nm spectral bands for the radiometer) was taken into account by calculating the
weighed reflectance corresponding to 20 nm simulated bands for the UltraCam-Xp WA, considering the
sensibility curve for each multispectral band. Therefore the weighed reflectance for the R-G-B-NIR
bands was obtained for each target.
A linear regression was carried out between the
at-surface reflectance values (dependent variable)
and the digital numbers (independent variable)
for each target and for each multispectral band, as
showed in Del Pozo et al. (2014). A full cross-validation was carried out (Geladi & Kowalski, 1986). The
criteria to choose the most suitable calibration set
were to minimize the RMSE and to maximize the R2
in the validation.
Biomass estimation using radiometrically
corrected images
The most suitable calibrations were used to obtain
the at-surface reflectance images (Ar, Br, Cr), which
were used to estimate the biomass. Two approaches
were followed, (i) using the reflectances for bands R,
G, B, NIR or (ii) using vegetation indices as independent variables. The vegetation indices considered in
this study were: NDVI (Rouse et al., 1974), Simple
Ratio (Birth & McVey, 1968), Soil Adjusted Vegetation Index (Huete, 1988), and Non-Lineal Vegetation Index (Goel & Qin, 1994). The prediction models
based on reflectance bands were adjusted by using
Partial Least Squares Regression (PLSR) (Geladi &
Kowalski, 1986), due to its suitability for collinear
data (Marabel & Álvarez-Taboada, 2013). Later on
a Leave-One-Out Cross validation was carried out.
The criteria to choose the most suitable model were:
minimize the number of factors, maximize R2, minimize RMSE and %RMSE of the validation. The biomass estimation using vegetation indices was done
by using Ordinary Least Squares Regression (OLSR),
using as criteria to maximize R2, and minimize RMSE
and %RMSE of the validation. All statistical analyses were performed by Unscrambler® X10.2 (CAMO
Software Inc., Woodbridge).
Results
Radiometric calibration
The calibration targets consisting of portable rigid
panels provided the most accurate radiometric
calibration for the three images (R2=0.99;
RMSE 0.014 %; %RMSE (%)<3,80%), therefore that
set can be used for calibration (Table 1). This set
consists of portable rigid targets with 0%, 25%, 50%,
75%, 100% nominal reflectance. These results are
similar to the ones obtained by Álvarez et al. (2010)
and Honkavaara & Markelin, (2007) for the ADS40,
DMC, UltraCamD and DSS multispectral sensors.
Tabla 1. Empirical line calibration accuracy for each band for the suggested set. RMSE values in % (reflectance)
124
.
0.996
0.999
0.997
0.998
0.997
0.0178
5.31
0.0092
2.69
0.0146
4.33
0.0107
2.34
0.0131
3.67
.
0.996
0.999
0.995
0.998
0.997
0.0169
5.03
0.0086
2.52
0.0174
5.13
0.0108
2.36
0.0134
3.76
.
0.996
0.999
0.996
0.997
0.997
0.0169
5.03
0.0084
2.48
0.0150
4.42
0.0121
2.64
0.0131
3.64
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Table 2. Results for the cross-validation of the models for biomass estimation in the three images, using
reflectances, vegetation indices and the recommended calibration set. P=Image; F: latent factors; E=RMSE:
g/m2; %E=%RMSE: %.
) 0,873
) 0,878
) 0,898
4,5741 2
4,4842 2
4,0958 2
14,42 0,793 5,6500 17,82 0,725 6,5181 20,56
14,14 0,767 5,9964 18,91 0,710 6,6939 21,11
12,92 0,797 5,9635 18,81 0,701 6,7929 21,42
0,851
0,830
0,839
4,7869 15,10
5,1158 16,13
4,9887 15,73
0,843
0,830
0,841
4,9163 15,50
5,1163 16,13
4,9541 15,62
Tabla 3. Validation of the biomass estimation models obtained when the radiometric calibration equations
adjusted for image B were applied to image A and C, using reflectances, vegetation indices and the recommended calibration set. P=Image; F: latent factors; E=RMSE: g/m2; %E=%RMSE: %.
0.878
0.884
0.884
4.4842 2
4.4938 2
4.4938 2
14.14 0.767 5.9964 18.91 0.710 6.6939 21.11 0.830 5.1158 16.13 0.830 5.1163 16.13
14.17 0.775 5.8883 18.47 0.729 6.4613 20.38 0.841 4.9582 15.64 0.835 5.0497 15.92
14.17 0.778 5.8575 18.47 0.747 6.2529 19.72 0.844 4.9044 15.47 0.835 5.0441 15.91
Biomass estimation
The most accurate estimation of above-ground
biomass was obtained by using PLSR and at-surface
reflectance of the Cr image (R2=0.90; RMSE=4.096
g/m2; %RMSE (%)=12.92), although the differences
between images were not significantly different.
Using the vegetation indices the estimates were less
accurate (Table 2).
These results were obtained in the tree images (Ar,
Br, Cr). However the vegetation indices NLI or SAVI
can be used if a simplified prediction model is aimed,
since the similar accuracy in the results, with differences in %RMSE smaller than 1.6% (regardless the
considered image) showed the suitability of the calibration method and the model to estimate biomass.
When PLRS was used for the biomass estimation,
the differences between using different calibration
sets ranged from 0.1% to 2.5% for the RMSE (%)
(results not showed). However, when vegetation
indices were used (NDVI, SR, NLI, SAVI), the differences derived from using different calibration sets
ranged from 1.6% to 19.2% for the RMSE. NDVI and
SAVI were the least sensitive to the changes in the
calibration (<5% difference), while SR was the most
affected (10%-19% difference). These results were
obtained in the tree images.
The calibration equations adjusted for image B were
applied to image A and C, and the differences in the
biomass estimation after using those calibrations
or the calibration obtained specifically for B were
smaller than 1% (RMSE%) (Table 3).
Conclusions
The images of the UltraCam-Xp WA are therefore
suitable for estimating the above-ground biomass
once they have been calibrated to at-surface
reflectance, using PLSR and the spectral information
of the four bands of the image. Moreover, the
calibration of one image can be successfully used for
adjacent images.
References
Álvarez, F., Catanzarite, T., Rodríguez-Pérez, J.R.,
Nafría, D., 2010a. Radiometric Calibration and
Evaluation of the Ultracam Xp using Portable
Reflectance Targets and Spectroradiometer
Data. Application: to Extract Thematic Data from
the Imagery Gathered by the National Plan of
Arial Orthophotography (PNOA). En: Colomina, I.
Birth, G.S., and McVey, C. 1968. Measuring the color of
growing turf with a reflectance spectroradiometer,
Agronomy Journal, 60: 640-643.
CAMO, Technologies Inc., 2013. The Unscrambler
appendices: method references. PDFdocument.
Available at: http://www.camo.com/ (09/11/2016).
Del Pozo, S., Rodríguez-Gonzálvez, P., HernándezLópez, D., Felipe-García, B. 2014. Vicarious
Radiometric Calibration of a Multispectral
Camera on Board an Unmanned Aerial System.
Remote Sensing, 6: 1918-1937.
Geladi, P., Kowalski, B.R. 1986. Partial leastsquares
regression: a tutorial. Analytica Chimica Acta,
185: 1-17.
125
Forest Mapping & Inventory
Goel, N.S., Qi, W., 1994. Influences of canopy
architecture on relation- ships between various
vegetation indices and LAI and FPAR: A computer
simulation. Remote Sens. Rev., 10: 309-347.
Honkavaara, J., Siitari, H., Viitala, J. 2004. Fruit
color preferences of redwings (Turdus iliacus):
experiments with hang-raised juveniles and wildcaught adults. Ethology, 110: 445-457.
Honkavaara, E., Markelin, L., 2007. Radiometric
Performance of Digital Image Data Collection.
A Comparison of ADS40/DMC/UltraCam and
EmergeDSS, Photogrammetric Week. (Fritsch,
D., Ed.), Wichmann Verlag, Heidelberg, Germany,
pp. 117-129.
Huete, A.R., Jackson, R. D., Post, D.F., 1988. Spectral
response of a plant canopy with different soil
backgrounds. Remote Sensing of Environment,
17: 37-53.
Jensen, J. R., 2005. Introductory to Digital Image
Processing: A remote sensing perspective (3rd
ed.). Upper Saddle River, NJ, USA, Prentice Hall,
526 p. ISBN: 0-13-145361-0.
Marabel, M. y Álvarez-Taboada F. 2013. Spectroscopic
Determination of Aboveground Biomass in
Grasslands Using Spectral Transformations,
Support Vector Machine and Partial Least
Squares Regression. Sensors, 13: 10027-10051.
126
Universidad Mayor
Marabel-García, M. y Álvarez-Taboada F. 2014.
Estimación de biomasa en herbáceas a partir
de datos hiperespectrales, regresión PLS y la
transformación continuum removal. Revista de
la Asociación Española de Teledeteción (AET)
42:49-60.
Rouse, J.W., Haas, R.H., Schell, J.A., y Deering,
D.W. 1974. Monitoring vegetation systems in
the Great Plains with ERTS, In: S.C. Freden,
E.P. Mercanti, and M. Becker (eds) Third Earth
Resources Technology Satellite–1 Syposium.
Technical Presentations, NASA SP-351, NASA,
Washington, D.C., 1: 309-317.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
ALS-based forest inventory data and stem data from harvester
to predict timber quality of Norway spruce structural timber
Carolin Fischer1, Marius Hauglin1, Terje Gobakken1, Olav Høibø1, Geir I. Vestøl1
1
Norwegian University of Life Sciences, Department of Ecology and Natural Resource
Management, P.O. Box 5003, NO-1432 As, Norway.
Keywords: Strength grading, timber properties, forest inventory
The properties of structural timber vary significantly
at different levels, making strength grading of each
single board crucial. Norway spruce (Picea abies)
is the main species for the production of structural
timber in Norway, and most of the strength grading
machines used account for only parts of the great
variation. In order to meet the requirements of
timber properties from all subsamples, the grading
has to be conservative, resulting in a poor utilization
of the properties of the timber resource. The
yield of higher strength classes can potentially be
improved by combining machine strength grading
with knowledge about the timber resource, and
thereby improve the competitiveness for timber
as a structural material. Different methods to
predict timber properties prior to sawing have been
investigated, and one possibility is to use more
information about the origin of the timber. The aim
of the current study is to predict the distribution
of timber properties at the sawmill based on ALS
data, forest inventory data and stem data from the
harvester.
We collected timber from four sites in southeastern
Norway. On each site, 20 trees were selected and the
timber was graded in a local sawmill. We developed
single-tree models predicting timber strength using
variables derived from ALS and harvester data. These
models provide valuable insight in the potential of
forest inventory and remote sensingdata to predict
important timber properties to improve strength
grading.
127
Forest Mapping & Inventory
Universidad Mayor
Characterization of forest structures through the synergetic
fusion of airborne LiDAR and multispectral sensor-derived
difference vegetation index.
Manzanera, JA., García-Abril, A., Pascual, C, Tejera, R, Martín-Fernández S, Martínez-Falero E, Valbuena, R2.
Affiliation(s) & Corresponding Author(s):
Research Group for Sustainable Management SILVANET FoReStLab, College of Forestry and Natural Environment,
Technical University of Madrid, Ciudad Universitaria, 28040 Madrid, Spain.
2
University of Eastern Finland. Faculty of Forest Sciences. Yliopistokatu 7. 80100 Joensuu, Finland; and: University of
Cambridge. Department of Plant Sciences. Downing St. CB2 3EA Cambridge, UK.
Corresponding author: J.A. Manzanera (joseantonio.manzanera@upm.es).
Keywords (5-8) Airborne laser scanning; Data fusion; Difference Vegetation Index; Forest Structural Types;
multispectral imagery; Stand structure.
Abstract
The aim of this contribution is the incorporation of complementary multispectral information from
optical sensors to Light Detection and Ranging (LiDAR), particularly through the incorporation
of complementary multispectral information from optical sensors by back-projecting the LIDAR
points onto the multispectral image. Canonical Correlation Analysis (CCA) was performed to relate
multivariate data sets of both LIDAR and multispectral metrics with structural variables measured
in a Scots pine stand. The correlation coefficient between pairs of canonical variables ranged from
0.997 to 0.949 for the four significant pairs. The amount of explained variance in the response
component ranged from 99.39% to 90.1%, and an 86.2% accumulated variance was explained. We
related indicators of stand development, i.e. height and volume, with LIDAR elevation metrics, tree
size and stand density, representing maturity of the stand, and LIDAR elevation metrics and three red
edge index-based multispectral metrics with Lorenz curve-derived attributes, which represent size
heterogeneity within the stand, in the first three pairs of canonical variables. The third pair of canonical
variables was interpreted as an indicator of size variability, discriminating even-sized from unevensized stands, which also benefited from the multispectral information. In the fourth pair of canonical
variables, four red edge index metrics significantly explained the heterogeneity of the stand. We may
conclude that metrics from the optical sensor complemented information of the LIDAR sensor.
Introduction
Initial studies of forest structure inferred from
LIDAR advanced incorporating predictors derived
from satellite multispectral datasets (Pascual et
al., 2010). Data fusion of both LIDAR and optical
passive sensors, such as multispectral photography,
have synergic capabilities for providing reliable
inventories for operational forestry (Packalén et al.,
2009; Valbuena et al., 2013). Estimation of forest
parameters from ALS can also be assisted by image
analysis, increasing the potential of multispectral
128
imagery for thematic classification and index
calculation (St-Onge and Achaichia, 2001).
Appropriate automated geometric correspondence
of simultaneously acquired LIDAR and aerial
multispectral imagery, plus the positioning of LIDAR
points on the conical projection of uncorrected aerial
pictures is a method known as back-projecting,
which successfully permitted allocating the original
radiometric information in the LIDAR point cloud
(Valbuena et al 2011, 2013).
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Therefore, our purpose is to analyze the canonical
relationship between two coincident multivariate
datasets, one containing both LIDAR and passive
optical sensor measurements, and the other
containing field measurements of forest stands.
to find that metrics derived from the multispectral
sensor showed significant explanatory potential for
the prediction of these structural attributes.
Materials and Methodology
In light of the obtained results, we concluded
that metrics derived from the optical sensor have
potential for complementing the information from
the LiDAR sensor in describing structural properties
of forest stands. We therefore recommend the use of
back-projecting for jointly exploiting the synergies
of both sensors using similar types of metrics as they
are customary in forestry applications of LiDAR.
The study area is the Valsaín forest (Spain; latitude:
40°53’ – 41°15’ N; longitude: 3°59’ – 4°18’ W; altitude
1300 – 1500 m above sea level), which is attached to
the Sierra de Guadarrama National Park, with Scots
pine (Pinus sylvestris L.) as the main species. Field
survey consisted of 37 circular plots of 20 m-radius
sampled in a cluster design, where the standard
dasometric variables were inventoried.
Remote sensing data was simultaneously acquired
with a discrete-pulse multi-return airborne laser
scanner ALS50-II (Leica Geosystems, Switzerland),
and a digital mapping camera DMC (ZeissIntergraph, Germany) system consisting of four
charge-coupled device (CCD) frame array sensors.
With Red and NIR data we generated the Red edge
(redge) index (Gautam et al., 2010):
redge = NIR – Red
(1)
Canonical Correlation Analysis (CCA) was performed
to assess the interrelationships between two data
sets of variables: LIDAR and multispectral metrics,
on one side (the predictor dataset), and forest
structure indicators, on the other.
Results
A multivariate data set of both LiDAR and
multispectral metrics could be related with a
multivariate data set of stand structural variables
measured in a Scots pine forest through Canonical
Correlation Analysis. Four statistically significant
pairs of canonical variables were found, which
explained 84.2% accumulated variance. The first
pair of canonical variables related indicators of stand
development, i.e. height and volume, with LiDAR
height metrics. The second pair related stand density
to LiDAR variables determining canopy coverage.
The third and fourth pairs of canonical variables
pertained to Lorenz curve-derived attributes, which
are measures of within-stand tree size variability and
heterogeneity, able to discriminate even-sized from
uneven-sized stands. The most relevant result was
Conclusions
References
Gautam B.R., Tokola T., Hamalainen J., Gunia M.,
Peuhkurinen J., Parviainen H. (2010) - Integration
of airborne LiDAR, satellite imagery, and field
measurements using a two-phase sampling
method for forest biomass estimation in
tropical forests. In: International Symposium
on “Benefiting from Earth Observation”.
Kathmandu, Nepal.
Packalén P, Suvanto A, Maltamo M (2009) A two
stage method to estimate species-specific
growing stock. Photogramm Eng Remote Sens
75:1451–1460
Pascual C., A. Garcıa-Abril, L.G. Garcıa-Montero,
S. Martın-Fernandez, W.B. Cohen. Objectbased semi-automatic approach for forest
structure characterization using lidar data in
heterogeneous Pinus sylvestris stands. Forest
Ecology and Management 255 (2008) 3677–3685.
St-Onge, A., and Achaichia, N. (2001). - Measuring
forest canopy height using a combination of
LIDAR and aerial photography data. International
Archives of Photogrammetry and Remote
Sensing, 343 W4, pp.131-137
Valbuena R., Mauro F., Arjonilla F. and Manzanera J.
A., Comparing Airborne Laser Scanning-Imagery
Fusion Methods Based on Geometric Accuracy in
Forested Areas. Remote Sensing of Environment.
115 (2011) 1942–1954.
Valbuena, R., De Blas, A., Martín Fernández, S.,
Maltamo, M., Nabuurs, G.J., & Manzanera, J.A.
(2013) Within-species benefits of back-projecting
laser scanner and multispectral sensors in
monospecific Pinus sylvestris forests. European
Journal of Remote Sensing 46: 491 – 509.
129
Universidad Mayor
Forest Mapping & Inventory
POSTER
Characterization of forest structures through the synergic
fusion of airborne LiDAR and multispectral sensor-derived
difference vegetation index
Manzanera, JA., García-Abril, A., Pascual, C, Tejera, R, Martín-Fernández S, Martínez-Falero E, Valbuena, R.
Affiliation:
Research Group for Sustainable Management SILVANET FoReStLab, College of Forestry and Natural Environment,
Technical University of Madrid, Ciudad Universitaria, 28040 Madrid, Spain.
Key words Airborne laser scanning; Data fusion; Difference Vegetation Index; Forest Structural Types;
multispectral imagery; Stand structure.
Abstract
The aim of this contribution is the incorporation of complementary multispectral information from
optical sensors to Light Detection and Ranging (LiDAR), particularly through data fusion by backprojecting the LiDAR points onto the multispectral image.
Materials and methods used:
A multivariate data set of both LiDAR and multispectral metrics was related with a multivariate data set
of stand structural variables measured in a Scots pine forest through Canonical Correlation Analysis.
Main Results:
Four statistically significant pairs of canonical variables were found, which explained 84.2%
accumulated variance. The first pair of canonical variables related indicators of stand development,
i.e. height and volume, with LiDAR height metrics. The second pair related stand density to LiDAR
variables determining canopy coverage. The third and fourth pairs of canonical variables pertained
to Lorenz curve-derived attributes, which are measures of within-stand tree size variability and
heterogeneity, able to discriminate even-sized from uneven-sized stands. The most relevant result was
to find that metrics derived from the multispectral sensor showed significant explanatory potential for
the prediction of these structural attributes. Therefore, we concluded that metrics derived from the
optical sensor have potential for complementing the information from the LiDAR sensor in describing
structural properties of forest stands. We therefore recommend the use of back-projecting for jointly
exploiting the synergies of both sensors using similar types of metrics as they are customary in forestry
applications of LiDAR.
130
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Effect of flying altitude, scanning angle and scanning mode on
the accuracy of ALS based forest inventory
Juha Keränen1, Matti Maltamo1, Petteri Packalen1
Affiliation and address (including e-mail address) for each author
Faculty of Science and Forestry, School of Forest Sciences. University of Eastern Finland, P.O. Box 111,
FI-80101 Joensuu, Finland. FAX: +358 29 4457 316.
Email: juha.keranen@uef.fi; matti.maltamo@uef.fi; petteri.packalen@uef.fi
1
Keywords: ALS, flying altitude, scanning angle, scanning mode, bootstrap approach, forest inventory
Abstract
Many countries are nowadays collecting airborne laser scanning (ALS) data nation-wide to produce
high quality elevation data. While ALS data is used in the forest inventory, the ALS data acquisition
costs form a remarkable proportion of the total inventory costs. Also the accuracy of the ALS based
forest inventory is closely dependent on ALS data. That is the reason why ALS data collection
parameters must be carefully defined. These extrinsic parameters have been assessed in numerous
studies about a decade ago, but since then ALS devices have developed and it is possible that previous
findings do not hold true with newer technology.
In this study the effect of flying altitudes (2000, 2500 or 3000 m), scanning angles (±15° and ±20° off
nadir) and scanning modes (single- and multiple pulses in air) with the area-based approach using a
Leica ALS70HA-laser scanner was studied. The study was conducted in a managed pine-dominated
forest area in Finland, where eight separate discrete-return ALS data were acquired with the same
aerial coverage. With these ALS data sets and field data comprising 47 field sample plots the estimates
for plot-level volume and mean height were calculated. The comparison of results of the different data
sets was based on the bootstrap approach with 5-fold cross validation.
Results indicated that the narrower scanning angle (±15° i.e. 30°) led to slightly more accurate estimates
of plot volume (RMSE%: 21-24 vs. 22.5-25) and mean height (RMSE%: 8.5-11 vs. 9-12). In addition, the
results indicated that a moderate increase in flying altitude does not directly decrease the accuracy
of the prediction. This is the situation not only with the narrower scanning angle but also with the
wider scanning angle. This can be considered as a positive finding because it means that (almost) the
same RMSE% can be obtained at a lower cost. Our analysis also indicated that contemporary ALS
technology enables the use of multiple pulses in air mode.
131
Universidad Mayor
Forest Mapping & Inventory
Forest aboveground biomass mapping in Mexico using SAR,
optical and airborne LiDAR data
Mikhail Urbazaev1, Christiane Schmullius1, Christian Thiel1, Bruce Cook2, Ralph Dubayah3,
Mirco Migliavacca4, Markus Reichstein4
1
Friedrich-Schiller-University, Jena, Germany
NASA Goddard Space Flight Center, Greenbelt, MD, USA
3
University of Maryland, College Park, MD, USA
4
Max-Planck-Institute for Biogeochemistry, Jena, Germany
2
Keywords: aboveground biomass, Mexico, remote sensing, ALOS PALSAR, Landsat, airborne LiDAR
Abstract
Information on the spatial distribution of aboveground biomass (AGB) over large areas is needed for
understanding and managing the processes involved in the carbon cycle, and supporting international
policies for climate change mitigation and adaption. Furthermore, these products provide local
stakeholders with important baseline data for the development of sustainable management strategies.
Using remote sensing techniques it is possible to provide spatially explicit information of AGB from
local to global scales. In this work we present a twostage upscaling approach to estimate forest
aboveground biomass in Mexico at national scale based on multi-sensor remote sensing data. For this,
we estimate firstly AGB along the airborne LiDAR transects using Mexican National Forest Inventory
data collected by CONAFOR and very high resolution NASA G-LiHT LiDAR data. We calculated from
discrete-return LiDAR data more than 80 LiDAR metrics that are then related to field-estimated AGB.
In the next step, we calibrate active (ALOS PALSAR) and passive satellite imagery (Landsat) with
LiDAR-based AGB estimates in a non-parametric Random Forest model to create a national wall-towall AGB map. Finally, the generated AGB product is validated using independent Mexican National
Forest Inventory data that were not used for model training. Furthermore, we modeled AGB at national
scale using satellite imagery and National Forest Inventory data only and compared to the results from
the two-stage up-scaling approach in order to show a benefit of using airborne LiDAR data for largearea mapping. With airborne LiDAR data we increase a number of calibration data, which leads to a
robust estimation of AGB. Moreover, since field measurements are limited to point measurements,
they cannot adequately describe patterns at different spatial scales. In contrast, airborne LiDAR data
captures spatial variability of forest structure and improves estimation of AGB.
132
Universidad Mayor
ForestSAT 2016 Abstracts Summary
GEDI Biomass Model Development in Tropical Forests
Laura Duncanson1, Jim Kellner2, Steve Hancock3, John Armston3, Hao Tang3, Suzanne Marselis3, Ralph Dubayah3
Biospheric Sciences Laboratory, Goddard Space Flight Center (GSFC), NASA
2
Institute at Brown for Environment and Society, Brown University
3
Department of Geographical Sciences, University of Maryland, College Park
1
Key Words: GEDI, waveform LiDAR, biomass, tropics
The Global Ecosystem Dynamics Investigation
(GEDI) is a NASA funded, University of Maryland
(UMD) led mission to measure forest structure
from the International Space Station. GEDI will
be launched in early 2019 and provide waveform
LiDAR measurements of Earth’s forests between
~51 degrees North and South. One of the mission’s
primary goals is to produce an empirically derived
map of aboveground biomass (AGB). To estimate
biomass from GEDI data, models are being developed
from existing spatially and temporally coincident
airborne LiDAR and field measurements. Airborne
LiDAR are processed to simulate GEDI waveforms,
from which waveform metrics are extracted. These
metrics, (e.g. height percentiles), are empirically
related to field estimates of AGB. Here, we present
first results from tropical forests in Africa, Central
and South America. This presentation includes
results from Tanzania, Gabon, the Democratic
Republic of Congo, Costa Rica, Panama, Borneo,
Brazil, French Guyana, and Columbia. We develop
models both at the footprint (25 m diameter circle)
and multi-footprint level to assess the sensitivity
of modeling results to spatial scale. We test crossvalidated model accuracy using different regression
modeling frameworks including multiple linear,
stepwise linear, hierarchical Bayesian, and random
forests. Accuracies are assessed as a function of
topography, height, canopy cover, biomass, and
eco-region. Preliminary results show that modeling
accuracy decreases with increasing biomass, and
increases with increasing plot size.
133
Universidad Mayor
Forest Mapping & Inventory
Improved forest cover mapping based on MODIS time series
and landscape stratification
Jean-Philippe Denux1, Véronique Chéret1*, Emmanuelle Cano1, Mar Bisquert2 and Laurence Hubert-Moy3
1
Dynafor, Université de Toulouse, INRA, INPT, INP-EI Purpan, Toulouse, France
2
Universidad de Castilla-la-Mancha, Ciudad Real, Spain
3
Laboratoire LETG COSTEL, Université de Rennes 2, Rennes, France
*
Corresponding author, e-mail address: veronique.cheret@purpan.fr
Keywords: classification; stratification; MODIS; time series; object based image analysis;
forest cover mapping
Introduction
In Europe most of the forest inventories focus on
forest only without giving detailed information
other natural vegetation. Forest maps use mainly
a vector-based approach where polygons consider
forest stands as homogeneous. They updated
approximatively every 10 years. Nomenclatures and
classes definition change over different countries,
even over different provinces. We aim to use
remote sensing image as complementary source
of information to take into account more types of
forests and natural vegetation, to consider mixed
forest stands and gradients, to propose updates
regularly, and to homogenize nomenclatures. But
detailed forest cover mapping at a regional scale
by supervised classification is technically limited
by various factors. In a heterogeneous mountain
area, this study presents three points: the ability
of an OBIA landscape stratification method to
improve classification accuracy, the impact of time
compositing and finally the interest of studying
temporal consistency for accuracy assessment.
Materials
The study area is the Pyrenees mountain range
located at the border of France, Spain and Andorra,
covering a 75000 km² area. The elevation rises from
sea level to over 3000 m, with climatic influences
from Atlantic Ocean and Mediterranean Sea. We
processed MOD13Q1 products of Terra-MODIS,
using both NDVI (Rouse et al. 1973) and EVI (Huete
et al. 2002) in 16-day composite images, at 250 m
spatial resolution. The time series spans from 2000
134
to 2014, with 23 synthesis per year. We used as a
reference a synthesis of existing forest inventories,
namely the “Inventaire Forestier National” (IFN,
http://inventaire-forestier.ign.fr/spip/) for France,
the “Mapa Forestal de España” (MFE, www.
magrama.gob.es/es/biodiversidad/servicios/bancodatos-naturaleza/informacion-disponible/mfe50.
aspx) for Spain and the “Mapa Forestal del Principat
d’Andorra”
(MFPA,
http://www.iea.ad/mapaforestal-del-principat-dandorra) for Andorra. An
harmonized forest type nomenclature was created
from this database, composed of 22 forest and
natural vegetation classes.
Methodology
Stratification
The supervised classification was based on
Maximum Likelihood algorithm (ML). ML is
parametric approach, based on the hypothesis that
the values of each class are normally distributed in
a multidimensional features space (Richards 2013).
This may be a limitation to classify heterogeneous
natural vegetation over wide areas. To deal with this
problem an approach using spatial strafication to
divide each class into more homogeneous sub-classes
is frequently used (Richards and Kingsbury 2014).
The first step of the stratification was to delineate
radiometrically homogeneous regions considered
as landscape units using the object segmentation
method developed by (Bisquert et al. 2015). Principal
component analysis was apply to monthly synthesis
of EVI to select the 3 most representative dates.
Haralick textural indices (homogeneity, contrast,
dissimilarity, entropy and second moment) were
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Frequency
Table 1: Combination of variety and frequency
Maximum
(every year)
Intermediate
values
0
Always classified as
the reference
Never classified as the
reference
processed on a 5x5 pixels window. Then, the best
combination of EVI and textural index and optimal
scale parameters for segmentation, were chosen
with unsupervised evaluation methods (Johnson
and Xie 2011; Zhang et al. 2012). The second step
was to simplify the 56 homogeneous landscape units
created. Non-forested units were eliminated. Each
units with too small an area was merged with the
most similar nearby unit. Finally 25 homogeneous
landscape units were kept as strata. Classification
accuracies with and without stratification were
compared globally and per strata. Indices describing
the topographical and landscape of the strata
were used to identify characteristics related with
improvement of classification’s quality (Cano et al.
2017).
Input data and time compositing
Previously to the stratification, we analyzed the
effect of the organization of input data on the quality
of the classification, we compared:
● NDVI and EVI
● Annual temporal compositing: 3 dates, all the
dates, all dates except winter
● Inter-annual compositing: all the images for 1,
2 or 3 years, and calculating average images for
each date over 2 or 3 years.
Classifications with different combinations were
compared.
Temporal consistency
Kappa, producer’s and user’s accuracy and reject
fraction were used to assess the classification
quality (Congalton 1991). These indicators were
calculated using the synthesis of forest inventories
as a reference, selecting the pixels fully included in
forest polygons. Particular emphasis was made on
the temporal consistency of the results (Darren et
Variety
1
2 to maximum
Stable
Unstable
Stable and classified as
the reference
Errors and change
Stable affected to a
class different from the
reference
Errors and change
al. 2012). Frequency, variety and their combination
were caculated as indicator of temporal consistency
using all the classifications processed from 2000 to
2014. Frequency indicates the number of years a
pixel is classified as the reference. Variety shows the
number of different classes are assigned to a pixel.
Their combination allows to identify 3 categories of
results:
● Stable and classified as the reference
● Stable assigned to a class different from the
reference
● All others cases considered as error or change
Results
Input data and time compositing
Kappa values presented (figure 1) show clearly the
interest of using all the available dates, aside from
winter, for 3 years. Kappa value rises from 0.35 to
more than 0.50 between a data set composed of
3 dates of NDVI for 1 year and the one composed
of all available dates, aside from winter, for 3
years. Furthermore classifications based on NDVI
systematically show better results than the one
based on EVI.
Stratification
We compare classification for 2007-2009 and 20092011 three years period, for NDVI. Kappa index values
increased similarly for both periods, respectively,
from 0.53 to 0.65 and from 0.52 to 0.64. Meanwhile,
reject fraction (RF) value increased from 6 to 13%.
Results per strata (figure 2) show that higher Kappa
are usually accompanied by high reject fraction.
Most of the topographical and landscape indices
have no particular influence on the classifications
quality (table 2). The area and the forest cover rate for
each stratum are the only indices correlated with the
135
Universidad Mayor
Forest Mapping & Inventory
Figure 1: Effect of time
compositing on Kappa
Figure 2: Effect of stratification on Kappa and reject fraction
per strata
increase in classification quality. The interpretation
of these indices allows to identify thresholds which
leads to differentiate 3 groups:
● Strata with area >450 000 ha and forest area >120
000 ha: with no effect on classification quality
● Strata with area <100 000 ha and forest area <5
000 ha: very high reject fraction
● Intermediate strata presents an enhancement of
the classification quality.
Temporal consistency
Mapping the results of the combination of frequency
and variety displays a spatial organization of these
indicators (figure 3). Groups of pixels identified as
stable and misclassified seems to indicate errors in
the reference map based on the forest inventories.
136
Table 2: Coefficient of determination (R²)
between the stratum indicators and the
perstratum K and RF differences for the 20072009 period (bold: significant for α=0.01; (+):
positive correlation; (-): negative correlation).
Unstable pixels are often localized on the border
between polygons of two different classes. Our
hypothesis is that we may have some noisy data
with misregistration or acquisition angle problems.
On wider area these unstable pixels may indicate
a gradient and mixed pixels where the forest
inventories drew a linear limit.
Conclusion
Time series composing shows clearly the advantage
of using a three year series as input data for the
classification.
The object based stratification is easy to process
and useful to improve Maximum Likelihood results.
The landscape analysis stratification approach has
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Figure 3: Extract of the map showing the combination of frequency and variety
the same efficiency level as the usual approaches
encountered in literature. Only easily accessible
remote sensing data have been used and this
method does not have to rely on existent thematic
map or on expert knowledge. The analysis of the
strata landscape characteristics leads to strata
definition operational items that could be used
to adjust the parameters while creating HLU and
defining the strata. Strata area and the forest cover
rate are determining attributes to strata delineation.
Segmentation parameters can be adjusted to
produce an optimized stratification.
Accuracy assessment indicators processed from
confusion matrix are limited to the available
references. Reject fraction can be used for algorithm
quality assessment, furthermore it could be useful
for training evaluation and update. Temporal
consistency is an original point developed here. It
is complementary to existing accuracy assessment.
Variability can always be calculated. Frequency
is limited to reference availability but allows to
benefit from imperfect reference data. Thematic
interpretation of their combination needs to be
developed to distinguish errors from change.
References
Bisquert, M., Bégué, A., & Deshayes, M. (2015).
Object-based delineation of homogeneous
landscape units at regional scale based on MODIS
time series. International Journal of Applied Earth
Observation and Geoinformation, 37, 72-82
Cano, E., Denux, J.-P., Bisquert, M., Hubert-Moy,
L., & Chéret, V. (2017). Improved forest cover
mapping based on MODIS time series and
landscape stratification. International Journal
of Remote Sensing, accepted Congalton, R.G.
(1991). A review of assessing the accuracy of
classifications of remotely sensed data. Remote
Sensing of Environment, 37, 35-46
Darren, P., Rasim, L., Ian, O., & Robert, F. (2012).
Supervised Classification Approaches for the
Development of Land-Cover Time Series. Remote
Sensing of Land Use and Land Cover (pp. 177-190):
CRC Press
Huete, A., Didan, K., Miura, T., Rodriguez, E.P.,
Gao, X., & Ferreira, L.G. (2002). Overview of the
radiometric and biophysical performance of
the MODIS vegetation indices. Remote Sensing
of Environment, 83, 195-213 Johnson, B., & Xie,
Z. (2011). Unsupervised image segmentation
evaluation and refinement using a multi-scale
approach. ISPRS Journal of Photogrammetry and
Remote Sensing, 66, 473-483
Richards, J.A. (2013). Supervised Classification
Techniques. Remote Sensing Digital Image
Analysis (pp. 247- 318): Springer Berlin Heidelberg
Richards, J.A., & Kingsbury, N.G. (2014). Is There a
Preferred Classifier for Operational Thematic
Mapping? Ieee Transactions on Geoscience and
Remote Sensing, 52, 2715-2725 Rouse, J.W.,
Haas, R.H., Schell, J.A., & Deering, D.W. (1973).
Monitoring Vegetation Systems in the Great
Plains with ERTS. In NASA (Ed.), Third ERTS
Symposium. Washington, DC, USA
Zhang, X., Xiao, P., & Feng, X. (2012). An Unsupervised
Evaluation Method for Remotely Sensed Imagery
Segmentation. Geoscience and Remote Sensing
Letters, IEEE, 9, 156-160
137
Forest Mapping & Inventory
Universidad Mayor
Improving assessment of fire risk in Yunnan Province,
China using remote sensing
Jacqueline Rosette1, Iain Bye1, Bowei Chen2, Yongjie Xia2, Juan Suárez4, Shengxi Liao3, Huaiqing Zhang2,
Yong Pang2, Peter North1
1
Swansea University, Singleton Park, Swansea SA2 8PP, UK
Chinese Academy of Forestry, Institute for Forest Resources and Information Techniques (IFRIT), Beijing, China.
3
Chinese Academy of Forestry, Resource Institute for Resources and Insects (RIRI), Kunming, China.
4
Forest Research, Northern Research Station, Roslin EH25 9SY, UK
2
Key words: fuel model, fire risk, lidar, forest structure, fuelbed characterisation
The Newton Fund aims to strengthen research
and innovation partnerships between the UK and
emerging knowledge economies, and forms part
of the UK’s Official Development Assistance (ODA)
commitment. A research partnership was developed
between collaborators from Swansea University and
Forest Research in the UK, and the Chinese Academy
of Forestry (IFRIT and RIRI), and funded through a
Newton Agritech Project and the Royal Society
University Research Fellowship.
Lidar remote sensing was used to produce forest
structure, density and canopy profile-based inputs
for a fuel model. These enabled potential fuels to
be better-characterised using spatial variability
of vegetation and canopy structure, reducing
assumptions from field data sampling, and improving
the resolution of fire risk information.
The project included an intensive, collaborative
month-long field campaign in Yunnan Province,
China. Remote sensing analysis was used to segment
and stratify the landscape using representative
land cover classes and age distribution, enabling
the impartial selection of field plot locations. The
four vegetation classes comprised broadleaf forest,
conifer forest, bamboo stands, and economic shrubs
(tea and coffee). Broadleaves and conifers were
further subcategorised into young, pole and mature
classes.
138
The project adapts and applies the US Forest
Service Fuel Characteristic Classification System
(FCCS) to the Chinese context. Lidar permitted
direct observations and estimates of the spatial
variability of canopy structure, fractional cover, and
dynamic vegetation parameters which are key to
the successful estimation of fire risk. Understorey
shrubs, ladder fuels, herb layer and litter were less
easily estimated using remote sensing and were
therefore characterised using mean static values
from field investigation for each vegetation class.
SMAP data were additionally incorporated in order
to assess seasonal variability of susceptibility of a
fuelbed to surface and crown fires, and to consume
fuels.
The methodology reduces the assumptions based
purely on vegetation type that would otherwise
be typically formed from a coarse, ground-based
sampling scheme and conventional field procedures.
The results can inform management decisions
regarding risk potential and containment, allowing
specific at-risk stands to be managed or monitored
accordingly e.g. using fire breaks or managing
undergrowth. The proximity of shrub and forest
agriculture to rural dwellings, and the spatial
assessment of fire susceptibility and behaviour
potential, can have societal implications, both for
economic security (i.e. loss of potential revenue) and
for population safety.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Industrial forest mapping: a Landsat Spatial and
Temporal Approach
Luigi Boschetti*, Lian-Zhi Huo*, Nuria Sanchez Lopez*, Andrew Hudak**, Alistair Smith*, Robert Keefe*
* College of Natural Resources, University of Idaho, Moscow, ID 83844, USA
** Rocky Mountain Research Station, US Forest Service, Moscow, ID 83844, USA
Keywords: Forest Mapping and Inventory, Landsat, Automatic processing, Object Oriented,
Deforestation, Industrial Forestry
Abstract
Remote sensing has been widely used for mapping and characterizing changes in forest cover, but
the available remote sensing forest change products are not discriminating between deforestation
(permanent transition from forest to non-forest), industrial forest management (logging followed by
regrowth, with no land use change) and natural disturbances such as insect damage and fires (forest
cover loss followed by regrowth, also with no land use change) (Hansen et al, 2010). Current estimates
of carbon-equivalent emissions report the contribution of deforestation as 12% of total anthropogenic
carbon emissions (van der Werf et al., 2009), but accurate monitoring of forest carbon balance should
discriminate between land use change related to forest natural disturbances, and forest management.
The total change in forest cover (Gross Forest Cover Loss, GFLC) needs to be characterized based on
the cause (natural/human) and on the outcome of the change (regeneration to forest/transition to
non/forest)(Kurtz et al, 2010). Industrial forestry is nowadays highly optimized: economic profitability
forces the adoption of standard practices resulting in very clear spatial patterns evident to human
interpreter, but hardly detectable with traditional satellite mapping approaches. To overcome these
challenges, we propose a methodology for post-processing forest cover change maps (Hansen et
al., 2013) by classifying each forest cover loss detection as either (a) deforestation, (b) fire and insect
disturbances or (c) forest management practices. The classification methodology combines the use
of multitemporal Landsat data time series, and object-oriented analysis of shapes, textures, and
spatial relationships of the areas of forest cover loss. The methods are demonstrated by wall-to-wall
classification of the forest cover loss in the conterminous United States for the 2002-2011 period.
References
Hansen, M.C., Stehman, S.V., & Potapov, P.V. (2010).
Reply to Wernick et al.: Global scale quantification
of forest change. Proceedings of the National
Academy of Sciences, 107, E148-E148
Hansen, M. C., Potapov, P. V., Moore, R., Hancher,
M., Turubanova, S. A., Tyukavina, A., ... &
Kommareddy, A. (2013). High-resolution global
maps of 21st-century forest cover change.
science, 342(6160), 850-853.
Kurz, W.A. (2010). An ecosystem context for global
gross forest cover loss estimates. Proceedings
of the National Academy of Sciences, 107, 90259026
van der Werf, G.R., Morton, D.C., DeFries, R.S.,
Olivier, J.G., Kasibhatla, P.S., Jackson, R.B.,
Collatz, G.J., & Randerson, J. (2009). CO2
emissions from forest loss. Nature Geoscience, 2,
737-738
139
Universidad Mayor
Forest Mapping & Inventory
POSTER
Improving merchantable timber volume accuracy for
balsam fir plots by analyzing the spatial distribution of
airborne LiDAR returns
Sarah Yoga B., Jean Bégin, Benoît St-Onge, Martin Riopel
Sarah Yoga B. (sarah.yoga-bengbate.1@ulaval.ca), Department of Wood and Forest Sciences,
Laval University, Quebec City, QC, Canada.
Jean Bégin (jean.begin@sbf.ulaval.ca), Department of Wood and Forest Sciences,
Laval University, Quebec City, QC, Canada.
Benoît St-Onge (st-onge.benoit@uqam.ca), Department of Geography, Université du Québec
à Montréal, QC, Canada.
Martin Riopel (martin.riopel.1@ulaval.ca), Department of Wood and Forest Sciences,
Laval University, Quebec City, QC, Canada.
Keywords: Lidar, merchantable volume, balsam fir, spatial distribution, density, residual variance.
Abstract
Airborne Laser Scanning data (ALS) are now regularly used to characterize forest structure. In this
study, we determined the effects of the scanning angle settings of an ALS system and those of the
point spacing on the accuracy of merchantable timber volume estimates of balsam fir plots in eastern
Canada. We used the ALS point cloud to compute predictor variables of the merchantable volume in
a nonlinear model. Our best model included ALS mean height, the proportion of first returns below 2
m and the rumple index. Our analysis showed a high correlation (pseudo-R2 = 0.91, RSE = 27.5 m3.ha-1)
between ALS and field data of 119 plots. A more accurate merchantable volume estimate was obtained
from ALS data by focusing in the explanation of the residual variance of a standard prediction model.
We reduced the RSE (residual standard error) from 27.5 m3.ha-1 to 13.5 m3.ha-1 by including a variable
accounting for the spatial distribution of the ALS returns in our model. There was, however, no effect
of the ALS returns density on the merchantable volume models (p=0.74). This suggests that the
spatial distribution of ALS returns rather than the point density should be considered when deriving
merchantable volume estimates from ALS data.
140
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Industrial forest mapping:
a Landsat Spatial and Temporal Approach
Luigi Boschetti1, Lian-Zhi Huo1, Nuria Sanchez Lopez1, Andrew Hudak2, Alistair Smith1, Robert Keefe1
2
1
College of Natural Resources, University of Idaho, Moscow, ID 83844, USA
Rocky Mountain Research Station, US Forest Service, Moscow, ID 83844, USA
Keywords Forest Mapping and Inventory, Landsat, Automatic processing, Object Oriented, Deforestation,
Industrial Forestry
Abstract
Remote sensing has been widely used for mapping and characterizing changes in forest cover, but
the available remote sensing forest change products are not discriminating between deforestation
(permanent transition from forest to non-forest), industrial forest management (logging followed by
regrowth, with no land use change) and natural disturbances such as insect damage and fires (forest
cover loss followed by regrowth, also with no land use change) (Hansen et al, 2010). Current estimates
of carbon-equivalent emissions report the contribution of deforestation as 12% of total anthropogenic
carbon emissions (van der Werf et al., 2009), but accurate monitoring of forest carbon balance should
discriminate between land use change related to forest natural disturbances, and forest management.
The total change in forest cover (Gross Forest Cover Loss, GFLC) needs to be characterized based on
the cause (natural/human) and on the outcome of the change (regeneration to forest/transition to
non/forest)(Kurtz et al, 2010). Industrial forestry is nowadays highly optimized: economic profitability
forces the adoption of standard practices resulting in very clear spatial patterns evident to human
interpreter, but hardly detectable with traditional satellite mapping approaches. To overcome these
challenges, we propose a methodology for post-processing forest cover change maps (Hansen et
al., 2013) by classifying each forest cover loss detection as either (a) deforestation, (b) fire and insect
disturbances or (c) forest management practices. The classification methodology combines the use
of multitemporal Landsat data time series, and object-oriented analysis of shapes, textures, and
spatial relationships of the areas of forest cover loss. The methods are demonstrated by wall-to-wall
classification of the forest cover loss in the conterminous United States for the 2002-2011 period.
References
Hansen, M.C., Stehman, S.V., & Potapov, P.V. (2010).
Reply to Wernick et al.: Global scale quantification
of forest change. Proceedings of the National
Academy of Sciences, 107, E148-E148
Hansen, M. C., Potapov, P. V., Moore, R., Hancher,
M., Turubanova, S. A., Tyukavina, A., ... &
Kommareddy, A. (2013). High-resolution global
maps of 21st-century forest cover change.
science, 342(6160), 850-853.
Kurz, W.A. (2010). An ecosystem context for global
gross forest cover loss estimates. Proceedings
of the National Academy of Sciences, 107, 90259026
van der Werf, G.R., Morton, D.C., DeFries, R.S.,
Olivier, J.G., Kasibhatla, P.S., Jackson, R.B.,
Collatz, G.J., & Randerson, J. (2009). CO2
emissions from forest loss. Nature Geoscience, 2,
737-738
141
Universidad Mayor
Forest Mapping & Inventory
Inventory of Small Forest Areas Using an
Unmanned Aerial System
Stefano Puliti, Hans Ole Ørka, Terje Gobakken, and Erik Næsset
Department of Ecology and Natural Resource Management, Norwegian University of Life Sciences,
PO Box 5003, NO-1432 Ås, Norway
Keywords: UAS; forest inventory; area-based approach; structure from motion; photogrammetry
Abstract
Acquiring high spatial and temporal resolution imagery from small unmanned aerial systems (sUAS)
provides new opportunities for inventorying forests at small scales. Only a few studies have investigated
the use of UASs in forest inventories, and the results are inconsistent and incomplete. The present
study used three-dimensional (3D) variables derived from UAS imagery in combination with ground
reference data to fit linear models for Lorey’s mean height (hL), dominant height (hdom), stem number
(N), basal area (G), and stem volume (V). Plot-level cross validation revealed adjusted R2 values of 0.71,
0.97, 0.60, 0.60, and 0.85 for hL, hdom, N, G, and V, respectively, with corresponding RMSE values of 1.4
m, 0.7 m, 538.2 ha−1, 4.5 m2·ha−1, and 38.3 m3·ha−1. The respective relative RMSE values were 13.3%,
3.5%, 39.2%, 15.4%, and 14.5% of the mean ground reference values. The mean predicted values did
not differ significantly from the reference values. The results revealed that the use of UAS imagery
can provide relatively accurate and timely forest inventory information at a local scale. In addition,
the present study highlights the practical advantages of UAS-assisted forest inventories, including
adaptive planning, high project customization, and rapid implementation, even under challenging
weather conditions.
142
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Large scale timber volume prediction with digital aerial
photogrammetry and national forest inventory data
Johannes Rahlf¹, Johannes Breidenbach¹, Svein Solberg¹, Erik Næsset², Rasmus Astrup¹
¹ National Forest Inventory, Norwegian Institute of Bioeconomy Research, Postboks 115, NO-1431
Ås, Norway
² Norwegian University of Life Sciences, Postboks 5003, NO-1432 Ås, Norway
Digital aerial photogrammetry (DAP) is a relatively
new source of high resolution 3D information for
remote sensing assisted forest inventories. Forest
parameters have been predicted with similar
accuracy as with airborne laser scanning (ALS) on
small study areas with homogeneous conditions.
Accuracies for large scale applications, however,
have not yet been reported. In our study we used
DAP and national forest inventory (NFI) data to fit a
timber volume prediction model on a large area with
varying terrain and forest vegetation. Based on this
model, we tested the influence of various factors on
the accuracy.
Aerial images were acquired in 2010 and 2013 over
two adjacent areas in central Norway. The total
area was 24 473 km2. The two image blocks were
processed separately to point clouds using different
matching strategies. A digital terrain model based
on ALS data was used to extract vegetation heights.
Timber volume measurements of 513 National
Forest inventory plots which had been collected
during a period of five years were used as reference
data. We calculated numerical metrics describing
the height distribution of the DAP data at the
sample plots. A linear model was fitted to the timber
volume measurements using DAP height metrics as
explanatory variables. Cross validation was used to
calculate R² and RMSE (root mean squared error)
of the model. Additionally, we calculated a range
of variables describing the time difference between
inventory and image acquisition, the color of the
point cloud, and terrain and light conditions. The
effect of these additional variables on the model
accuracy was used to investigate the influence of
the heterogeneous conditions caused by large area
coverage.
The model based on the DAP height metrics
had a R² of 0.83 and a RMSE of 38 m3 ha-1, which
corresponded to 49% of the mean observed timber
volume. Including relative sun inclination during
the image acquisition at the inventory plot, as well
as the time difference between the inventory and
the aerial image acquisition showed the largest
influence on the model accuracy. The R² increased to
0.85 and the RMSE was reduced to 35 m3 ha-1 (45%).
The time difference of the two image acquisitions in
combination with different matching strategies had
only a marginal influence on the accuracy.
143
Universidad Mayor
Forest Mapping & Inventory
Large-Scale Prediction of Aboveground Biomass in Mountain
Forests Utilizing Airborne Laser Scanning
Maltamo, M.1, Bollandsås, O.M.2, Gobakken, T.2 and Næsset E.2
2.
1.
University of Eastern Finland, Finland. P.O.Box 111, 80101 Joensuu, Finland.
Norwegian University of Life Sciences, Department of Ecology and Natural Resource Management, Norway. P.O.Box
5003, N-1432 Ås, Norway.
Keywords: LiDAR, forest biomass, montane forests, allometry, climatic variation, elevation
Abstract
For most forest attributes (timber volume, biomass, mean height etc.) airborne laser scanning (ALS)
has been shown to be the remotely sensed source of auxiliary information that provides the most
accurate estimates (e.g. Zolkos et al. 2013). In the case of aboveground biomass (AGB), variation in
crown allometry is largely accounted for by the ALS data since the echo distribution resulting from
the collection of ALS data is directly dependent on the three-dimensional biomass distribution of the
different crown components (e.g., Magnussen & Boudewyn 1998). ALS based forest biomass studies
has reported considerably larger errors for the biomass estimates than those found in studies conducted
in productive low-land forests (e.g. Montagnoli et al. 2015). The reason may be that montane forests,
on small spatial scales, are more heterogeneous with respect to growing conditions and terrain. On a
larger scale, climate will also affect growing conditions and allometry.
This study consider ALS based AGB prediction in mountain forests. The objectives of this study were 1)
to test the accuracy of ALS-based AGB models calibrated on data comprising high levels of small-scale
variation in tree allometry as well as large-scale variation in the underlying growing conditions, 2) to
test the additional effect of proxies for growing conditions in models for AGB based on ALS variables,
and 3) to test the additional effect of species information in models for AGB based on ALS variables
and proxies for growing conditions.
The study area consisted of a long transect from southern Norway to northern parts of the country
with wide ranges of elevation and latitude. The transect was covered by ALS data and field data from
238 plots. AGB was modeled using different types of predictor variables, namely ALS metrics, variables
related to growing conditions and tree species information. To represent growing conditions, we
used climatic metrics, elevation and latitude as predictor variables. We applied both linear regression
modeling and k nearest neighbor (k-nn) imputation. The k-nn approach was applied to examine the
predicting power of different types of variables, whereas linear regression analysis was performed to
construct general and dominant species level ALS-based biomass prediction models.
Modelling of AGB in the long transect covering diverse mountainous forest conditions was challenging,
the RMSE values being rather large (37-70%). The effect of growing conditions was minor in predictions.
Furthermore, we did not find good proxy variables for different tree forms and climatic conditions that
potentially could have provided additional information beyond the tree allometry-related information
reflected by the ALS metrics. However, species information was essential to improve accuracy. The
analysis revealed that when doing inventories of spruce dominated areas, all plots should be pooled
together when the models are developed, whereas if pines or deciduous dominate the area in question,
separate dominant species-wise models should be constructed.
144
Universidad Mayor
ForestSAT 2016 Abstracts Summary
References
Magnussen, S., and Boudewyn, P. 1998. Derivations
of stand heights from airborne laser scanner data
with canopy-based quantile estimators. Can. J.
For. Res. 28: 1016-1031.
Montagnoli, A., Fusco, S. Terzaghi, M., Kirsrchbaum,
A., Pflugmacher, D, Cohen, W.B., Scippa, G.S.,
and Chiatanrte, D., 2015. Estimating forest
aboveground biomass by low density lidar data
in mixed broad-leaved forests in Italian Pre-Alps.
Forest Ecosystems 2. doi:10.1186/s40663-0150035-6.
Zolkos, S.G., Goetz, S.J., and Dubayah, R. 2013.
A meta-analysis of terrestrial aboveground
biomass estimation using lidar remote sensing.
Remote Sens. Environ. 128: 289–298.
145
Universidad Mayor
Forest Mapping & Inventory
Mapping Amazonian biodiversity and geology using
basin-wide fern species inventories and Landsat imagery
Jasper Van doninck, Gabriela Zuquim, Hanna Tuomisto
Amazon Research Team, Department of Biology, University of Turku, Finland
Keywords: Amazonia, Landsat TM/ETM+, species composition, soil, classification
Fern species composition has been found to be
a predictor of species composition of other plant
groups, as well as soil properties. The University of
Turku Amazon Research Team and the Brazilian
Program for Biodiversity Research have developed
rapid fern species inventories methods based on
long plots. Collecting 1000 plots over the past
30 years, we have accumulated an exceptionally
large and internally consistent data set of fern
species composition and soil properties. These
field inventories have allowed us to document high
local-scale plant heterogeneity, and to identify
the interface of geological formations. Evidently,
these field inventories are labour intensive and can
only cover certain selected regions of the Amazon
basin. Satellite remote sensing is an indispensable
tool in order to achieve a comprehensive mapping
of Amazonian forest biodiversity and geology. Due
to the high heterogeneity at the local scale, such
146
mapping requires high resolution images like those
acquired by the Landsat satellites. Previous research
has shown that spectral patterns in individual
Landsat images reflect patterns in fern species
composition and geology, but radiometric artefacts
in Landsat imagery –caused by, e.g., atmospheric
contamination or the bidirectional reflectance
distribution function– have for a long time hampered
the use of Landsat over larger areas. Recent advances
in image preprocessing (atmospheric correction,
directional normalization and image compositing)
and processing capabilities have now allowed us to
generate a radiometrically consistent, cloud-free
Landsat TM/ETM+ image composite covering the
entire Amazon basin. We here combine this Landsat
mosaic with our entire dataset of fern species
inventories to obtain a wall-to-wall mapping of plant
biodiversity and the underlying soil properties at an
unprecedented spatial resolution.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Mapping certified forests for sustainable
management - a tool for information
improvement through citizen science
Florian Kraxner (International Institute for Applied Systems Analysis - IIASA, Laxenburg, Austria), Dmitry Schepaschenko*
(IIASA, Laxenburg, Austria), Sabine Fuss (Mercator Research Institute on Global Commons and Climate Change, Berlin,
Germany), Anders Lunnan (Norwegian University of Life Sciences, Aas, Norway), Georg Kindermann (Austrian Research
and Training Centre for Forests, Natural Hazards and Landscape, Vienna, Austria), Kentaro Aoki (Shinshu University,
Nagano, Japan), Martina Dürauer (IIASA, Laxenburg, Austria), Anatoly Shvidenko (IIASA, Laxenburg, Austria), Linda See
(IIASA, Laxenburg, Austria)
*
Presenting author
Keywords: forest certification mapping, FSC, PEFC, citizen science, Geo-Wiki
About 10% of forest area has been certified globally,
however, the speed of certification has slowed down
and the vast majority of certified forests are located
in the northern hemisphere. There are currently no
spatially explicit, openly accessible data available
on forest certification below national level so
understanding the drivers of these developments,
examining the scope for further certification and
using this information for development of future
sustainable forest management strategies are
challenging. We present a methodology for the
development of a spatially explicit global map
of certified forest areas as well as an online tool
(http://forest.geo-wiki.org) for visualization and
interactive improvement of the map, which is aimed
at a range of stakeholders including certification
bodies, third-party certifiers, green NGOs, forestry
organizations, decision-makers, scientists and
local experts. A new methodology for downscaling
national forest certification statistics has resulted in
the first spatially explicit global forest certification
map at a 1km resolution. Regional validation
(Russia) suggests an overall accuracy of 89%. By
building such a community-based online tool, more
accurate information on forest certification will
become available, promoting the sharing of data
and encouraging more transparent and sustainable
forest management. Such an approach is intended
to encourage transparency in the forest certification
arena but will also provide benefits to multiple
users, e.g. in monitoring, marketing and in the
development of targeted policy strategies.
147
Universidad Mayor
Forest Mapping & Inventory
POSTER
Mapping forest degradation caused by fires in 2010 in
Mato Grosso State, Brazilian Amazon using Landsat
TM fraction images
Yosio E. Shimabukuro1,*, Egidio Arai1, Liana O. Anderson2, Luiz Eduardo Aragão1, Valdete Duarte1
Instituto Nacional de Pesquisas Espaciais – INPE
Caixa Postal 515 - 12245-970 - São José dos Campos - SP, Brasil
2
Centro Nacional de Monitoramento de Desastres Naturais – CEMADEN
Parque Tecnológico de São José dos Campos, Estrada Doutor Altino Bondensan, 500,
São José dos Campos - São Paulo, 12247-016
*
Corresponding author – yosio@dsr.inpe.br
1
Keywords: Remote Sensing, Image Processing, Forest Degradation, Burned Areas,
Forest Fires, Fraction Images.
Abstract
The objective of this work is to assess the extent of forest degraded areas due to fires in 2010 in Mato
Grosso State, southern Amazonia. Mato Grosso State has experienced high deforestation rates and
therefore forest degradation activities due to fires do occur. For this work, we selected 42 Landsat
TM images acquired during the dry season of year 2010. The proposed method is based on (i) linear
spectral mixing model applied to TM images to derive soil, vegetation and shade fraction images
and (ii) image segmentation and classification applied to the shade fraction images. In a first step,
a map of forest/non forest areas are derived from Hansen et al. (2013) dataset. Then burned areas
are identified and mapped from the shade fraction images. These mapped areas are then distributed
in the three land cover types, i.e., forest, non forest (cerrado and old deforestation), and deforested
areas from 2001 to 2010. Our results showed that 32% of forests in Mato Grosso were burned during
year 2010 (22,633 km2) likely degrading the forest ecosystem. In addition, 5,175 km2 (7%) of burning
occurred in the areas deforested from 2001 to 2010 and 42,510 km2 (61%) occurred in the Cerrado and
old deforestation areas. The proposed method is efficient for mapping degraded forest areas due to
fires. The information is important for the carbon emission estimation.
Introduction
A large part of the gross carbon emissions into
the atmosphere due to land cover changes is
attributable to deforestation in the tropics (Achard
et al., 2014). Forest degradation, defined as longterm disturbance in forested areas, is considered
to represent up to 40% of the gross emissions from
deforestation in the Brazilian Amazon (Aragão et
al., 2014). In this region deforestation is defined as
forest clear cut with conversion to other land uses
(INPE, 2008), while forest degradation is related to
a combination of selective logging and forest fires
(Asner et al., 2009, Berenguer et al., 2014).
148
Hansen et al. (2013) mapped global tree cover extent,
loss, and gain for the period from 2000 to 2012
at a spatial resolution of 30 m, with loss allocated
annually. Then from this dataset, the forest/non
forest map can be extracted for the year 2000 to
2012 by selecting some tree cover threshold value.
Fraction images derived from different remote
sensing sensors have been used for many tropical
forest applications, especially in the Brazilian
Amazon. Specifically the ones derived from the
Landsat Thematic Mapper were tested for this
region for mapping areas of degraded forests due to
the following characteristics: a) vegetation fraction
images highlight the forest cover conditions and allow
Universidad Mayor
ForestSAT 2016 Abstracts Summary
differentiating between forest and non-forest areas
similarly to vegetation indices such as the Normalized
Difference Vegetation Index (NDVI) and the Enhanced
Vegetation Index (EVI); b) shade fraction images
highlight areas with low reflectance values such as
water, shadow and burned areas; and c) soil fraction
images highlight areas with high reflectance values
such as bare soil and clear-cuts (Shimabukuro et al.,
2014). In this context the main purpose of this study
is to present an estimate of the extent of degraded
forest due to fires in Mato Grosso State, Brazilian
Amazon, using a semi-automated procedure based
on fraction images derived from TM sensor.
and then an unsupervised classification algorithm
was applied in order to derive individual and explicit
objects (polygons). Secondly, the resulting clusters
of objects were assigned as unburned or burned
areas from their spectral and textural properties
derived from shade fraction images.
Methodology
Results
The study area corresponds to Mato Grosso State
located in the Brazilian Amazon region. Due to
variable climate, terrain relief, precipitation patterns,
and length of the dry season, the State of Mato
Grosso comprises parts of three Brazilian biomes,
the Amazon, the Cerrado, and a portion of the
Pantanal, and has a naturally very high biodiversity
with vegetation types ranging from dense evergreen
forest to deciduous open forest, savannas, natural
grasslands and seasonal wetlands. This region has
taken place a high rate on conversion of vegetation
cover not only due to the recent use of mechanized
agriculture but also due to deforestation processes,
selective logging and burning (INPE, 2005, 2008).
The forest/non forest map derived from Hansen et al.
(2013) shows that forest cover (>50% of tree cover) is
458,677 km2 and 74,129 km2 of forest loss from 2000
to 2010 (5,403 km2 from 2010). Our results (Figure 1)
showed that 70,317 km2 was burned in the State of
Mato Grosso during the year 2010. From this amount
5,175 km2 was deforested and burned and 22,633
km2 was forest burned, i.e. forest degraded by fire.
The most burned areas (42,510 km2) occurred in the
non-forest areas (Cerrado and old deforested areas).
For this work, we selected 42 Landsat TM images
at 30 m spatial resolution acquired during the
dry season of year 2010 to adapt a methodology
developed by Shimabukuro et al. (2009) for burned
area estimation.
The proposed method for mapping the forest
degradation by fire consists in four steps. The first
step (i) of our approach is to obtain a forest mask. In
this case, we used the Hansen et al. (2013) dataset.
The second step (ii) consisted in the generation of
fraction images (Shimabukuro and Smith, 1991) for
the 42 Landsat TM images selected during the dry
season of year 2010. The burned areas are assessed
from the shade fraction images considering that
the shade fraction images highlight areas with low
reflectance corresponding to burned areas.
The third step (iii) of our approach was divided in two
parts. First, a shade fraction image was segmented
The fourth step consists in producing the forest
degradation areas due to forest fires (burned forest),
by combining the resulting maps of forest/nonforest areas (first step) with the burned area maps
(from the third step), i.e. the forested areas that
were burned without clear cut during the year 2010.
The depicted burned areas in Amazonia are related
to management activities using fire, to deforestation
process, or to a degradation process: in the case of
deforestation the forest cover is first clear cut and
then the remaining vegetation is burned to allow
using the land for agriculture (cropland or grassland).
In the case of forest degradation, the forest cover is
burned through an uncontrolled fire without removal
of wood nor conversion to another land use. This
makes the use of a forest/non-forest map essential
for differentiating between deforestation and
forest degradation processes. Deforested areas will
appear as non-forest areas (cropland or grassland)
in the successive months or years after the initial
deforestation event while burned forests (degraded
forest) will recover as forest regrowth (Shimabukuro
et al., 2014).
Conclusions
The proposed method is efficient for mapping
burned forest areas (degraded forest areas due to
fires). An initial forest/non forest map is essential for
developing a procedure for mapping degradation
areas due to forest fires.
149
Forest Mapping & Inventory
Universidad Mayor
Fig.1 - Burned areas depicted by our method: in red over the forest areas and in blue
over the non-forest areas.
The shade fraction image generated from the
application of spectral linear mixing model on
Landsat TM images allowed to identify and to map
burned areas semi automatically. Then combining
the burned areas result with the forest/non forest
map derived from Hansen et al. (2013) dataset
allowed to estimate the degradation forest areas due
to forest fire. These information are very important
for the carbon emission estimation.
References
Achard, F., Beuchle, R., Mayaux, P., Stibig, H.-J.,
Bodart, C., Brink, A., Carboni, S., Desclée, B.,
Donnay, F., Eva, H. D., Lupi, A., Raši, R., Seliger,
R., Simonetti, D. Determination of Tropical
Deforestation Rates and Related Carbon Losses
from 1990 to 2010.” Global Change Biology 20:
2540–2554, 2014. doi:10.1111/gcb.12605.
Aragão, L. E. O. C., Poulter, B., Barlow, J. B.,
Anderson, L. O., Malhi, Y., Saatchi, S., Phillips,
O. L., Gloor, E. Environmental change and the
carbon balance of Amazonian forests. Biological
Reviews (2014).
Asner, G. P., Knapp, D. E., Balaji, A., Páez-Acosta, G.
150
Automated mapping of tropical deforestation
and forest degradation: CLASlite. Journal of
Applied Remote Sensing, 3, 033543, 2009.
Berenguer, E., Ferreira, J., Gardner, T. A., Aragão,
L. E. O. C., Camargo, P. B., Cerri, C. E., Durigan,
M., Oliveira, R. C., Vieira, I. C. G., Barlow, J. A
large-scale field assessment of carbon stocks in
human-modified tropical forests. Global Change
Biology, doi: 10.1111/gcb.12627, 2014.
Hansen, M., Potapov, P. V., Moore ,R., Hancher,
M., Turubanova, S. A., Tyukavina, A., Thau, D.,
Stehman, S. V., Goetz, S. J., Loveland, T. R.,
Kommareddy, A., Egorov, A., Chini, L., Justice, C.
O., Townshend, J. R. G. High-Resolution Global
Maps of 21st-Century Forest Cover Change.
Science, 342, 850-853, 2013.
INPE (Instituto Nacional de Pesquisas Espaciais).
Monitoramento da floresta Amazônica
brasileira por satellite. Seminário de avaliação.
Disponivel on-line: <http://www.obt.inpe.br/
prodes/seminario2005/>(Acessado em 12 de
Agosto de 2005), 2005.
INPE(Instituto Nacional de Pesquisas Espaciais).
Monitoramento da cobertura florestal da
amazônia por satélites: sistemas PRODES,
Universidad Mayor
ForestSAT 2016 Abstracts Summary
DETER, DEGRAD E QUEIMADAS 2007-2008.
INPE, São José dos Campos. 47p., 2008.
Shimabukuro, Y. E., Smith, J. A. The least squares
mixing models to generate fraction images
derived from remote sensing multispectral data.
IEEE Transactions on Geoscience and Remote
Sensing, 29, 16-20, 1991.
Shimabukuro, Y. E., Duarte, V., Arai, E., Freitas,
R. M., Lima, A., Valeriano, D. M., Brown, I. F.,
Maldonado, M. L. R. Fraction images derived
from Terra MODIS data for mapping burnt areas
in Brazilian Amazonia. International Journal of
Remote Sensing, 30, 1537-1546, 2009.
Shimabukuro, Y.E., Beuchle, R., Grecchi, R., Achard,
F. Assessment of forest degradation in Brazilian
Amazon due to selective logging and fires using
time series of fraction images derived from
Landsat ETM+ images. Remote Sensing Letters,
Vol. 5, No. 9, 773–782, http://dx.doi.org/10.1080/2
150704X.2014.967880, 2014.
151
Forest Mapping & Inventory
Universidad Mayor
Mapping forest height and biomass of the Chocó region,
Colombia, combining stratified random sampling of Lidar data
and spaceborne remote sensing data
Victoria Meyer1,2, Sassan, Saatchi1, Antonio Ferraz1, Alan Xu3, Alvaro Duque4, Mailyn Gonzalez5, Jérôme Chave2
Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA. USA
Laboratoire Evolution et Diversité Biologique UMR 5174, CNRS Université Paul Sabatier, Toulouse, France
3
Institute of the Environment and Sustainability, University of California, Los Angeles, CA 90095-1496
4
Departamento de Ciencias Forestales, Universidad Nacional de Colombia, Calle 59A No. 63-20, Medellín, Colombia
5
Instituto de Investigación de Recursos Biológicos, Alexander von Humboldt, Bogotá, D.C., Colombia
1
2
Keywords: Lidar, Biomass, Forest Height, Tropical Forest, Remote Sensing, Random Forest
Abstract
Mapping aboveground biomass using remote sensing has been proven to be accurate at local scales,
especially when relying on small footprint lidar data and their derived height metrics. Extending these
biomass estimations to larger areas that are not covered by lidar is a challenge that has yet to be resolved.
In this study, we show that developing a forest height map based on multiple remote sensing sources
can be the first step to producing a biomass map at a regional scale. Focusing on mapping forest height
prior to estimate biomass circumvents the issues of having errors related to local forest structure and
differences in allometry.
We are testing a random forest algorithm based on spaceborne remote sensing layers over the Chocó
region, located along the Pacific coast of Colombia, combining information on topography (SRTM), forest
structure (ALOS) and spectral signature (Landsat). Forty-seven small footprint lidar scenes of 20km2
each, based on stratified random sampling over the whole Chocó region as part of a BioREDD project,
are used as the training data. The result is a forest height map covering the Chocó region, for which the
uncertainties are reported at the pixel level. The effect of spatial resolution is also being analyzed. The last
step of the analysis consists in determining allometric equations based on available ground data in the
region and local parameters such as wood density. We test the hypothesis that biomass can be estimated
at a regional scale if based on an accurate forest height map and wood density information.
152
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Mapping of forest attributes across Canada using Landsat
pixel composites and LiDAR plots
Giona Matasci1, Michael A. Wulder2, Joanne C. White2, Geordie W. Hobart2, Harold S. J. Zald3,
Txomin Hermosilla1, Nicholas C. Coops*
1
Integrated Remote Sensing Studio, Department of Forest Resources Management, University of British Columbia,
Vancouver, British Columbia V6T 1Z4, Canada
2
Pacific Forestry Centre, Canadian Forest Service, Natural Resources Canada, Victoria,
British Columbia V8Z 1M5, Canada
3
Department of Forest Engineering Resources and Management, College of Forestry,
Oregon State University, Corvallis OR 97331, USA
Keywords: LiDAR, Landsat, forest mapping, large-scale, random forest, monitoring
Large-area mapping of forest structural attributes
is a critical information need for national forest
reporting. Landsat time series data provide key
long-term information on land cover and land cover
change, including forest cover and disturbance
characteristics. Passive optical data alone however
are limited in their capacity to estimate forest
structural attributes such as height and biomass.
In this contribution, we present a large-scale mapping
approach to estimate key forest structural variables
across the forested area of Canada by leveraging the
Landsat archive (TM/ETM+ sensors) and LiDAR plots
(airborne laser-based observations of the forest
structure). A cross-country LiDAR transect acquired
in 2010 provided the response variables describing
the vertical structure of the forest. Spatially
comprehensive pixel-based image composites
(cloud-free, radiometrically and phenologically
consistent surface reflectance composites), derived
vegetation indices and forest disturbance history
combined with topographic variables provided the
predictors. The random forest framework, a widely
accepted, robust and accurate approach was used
to impute stand height over the entire forested
area. Such a statistical model was developed and
validated on a set of more than 80,000 LiDAR plots.
With the full availability of the Landsat archive,
these wall-to-wall predictions can inform models of
stand development over time, and provide critical
information to national monitoring programs and
forest management policies.
153
Universidad Mayor
Forest Mapping & Inventory
Mapping the efficacy of fuel reduction burns using
image-based point clouds
Luke Wallace1,2, Karin Reinke1,2, Christine Spits1, Bryan Hally1,2 and Simon Jones1,2
1.
School of Sciences, Mathematical and Geospatial Sciences, RMIT University, Melbourne, Australia,
Luke.Wallace2@rmit.edu.au
2.
Bushfire and Natural Hazards Cooperative Research Centre, Melbourne, Australia
Keywords: Fuel reduction burns, Image-based point clouds, Terrestrial Laser Scanners, multi-temporal
monitoring
Fuel reduction burns are commonly used in fireprone forests to reduce the risk of wildfire, or to
maintain ecosystem condition and diversity. As
such, producing quantified assessments of fireinduced change is important to understanding the
outcomes of the intervention. This study aims to
quantify the changes in fuel structure generated
by a fuel reduction burn using image-based point
clouds. Colocated image sets are collected pre and
post burn within a dry sclerophyll forest in South
East Australia. These image sets are used to produce
point clouds, from which metrics representing
fuel volume, horizontal connectivity, and vertical
stratification are derived. Fire-induced change in
these metrics is assessed and the efficacy of the burn
in relation to fuel hazard is described. The method
is assessed against two sets of data, (1) the results
of visual fuel hazard and fire severity assessments
following standard guidelines used by Australian
154
land managers, and (2) the point clouds generated
using Terrestrial Laser Scanning (TLS) data collected
from a similar fire-altered forest plot. It is shown
that a similar correlation between fuel load and
image-based point cloud estimated volume (r2 =
0.70 to 0.87) and TLS estimated volume (r2 = 0.69 to
0.81) exists. Furthermore, similarities between the
image-based and TLS point clouds in mapping fuel
depths (root mean square deviation (rmsd) less than
0.03 m) and horizontal coverage (rmsd less than
5%) indicate that the fire-induced change can be
mapped using image-based point clouds with similar
detail and accuracy to that achieved with TLS. The
method utilises a low-cost consumer grade camera
and has similar in field requirements to the visual
assessments. As such the method presented can be
easily adapted for use by land managers to provide
routine and quantified assessments of fuel hazard
and fire severity.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Prediction of height, basal-area and stem volume in boreal
forest using Pléiades or WorldView-2 acquisitions
Henrik J. Persson1*, Håkan Olsson1, Johan E.S. Fransson1
Department of Forest Resource Management, Swedish University of Agricultural Sciences, Umeå, Sweden
*Corresponding author: henrik.persson@slu.se
1
Keywords: stereo, satellite, forest, WorldView, Pléiades
Abstract
This paper evaluates predictions of Lorey’s mean height (HL), basal-area (BA) and stem volume (VOL)
of boreal forest at two Swedish test sites, Krycklan and Remningstorp. Forest heights derived from
very high resolution stereo matched satellite data from the Pléiades sensor or the WorldView-2
sensor were trained with 10 m field plots and then predictions and evaluations were performed on
independently inventoried 40 m field plots. The best prediction results were found in Krycklan with
WorldView-2 data, with HL = 3.4% RMSE, VOL = 9.8% RMSE and BA with 9.7% RMSE. In conclusion,
both sensors delivered robust and accurate imagery for both the evaluated test sites. Moreover, the
presented approach appears suitable for operational forestry planning, especially at remote locations
with limited or no other remote sensing data.
Introduction
There is a steady need of establishing and updating
information about the forest. Airborne laser
scanning (ALS) has for more than one decade been
considered the most accurate remote sensing
technique that in combination with field samples can
be used to create wall-to-wall estimations of typical
forest variables, like Lorey’s mean height (HL), stem
volume (VOL) or basal-area (BA). Many countries,
including Sweden, can now offer complete national
terrain models based on ALS, which has enabled
also other techniques for estimations of this type
of variables. One such technique uses images
acquired from at least two directions which enables
stereogrammetric image matching to derive heights.
For this purpose, aerial images or very high resolution
(VHR) satellite images are suitable, as they often
possess resolutions below one meter. The image
matched result is a digital surface model (DSM) from
which the terrain model can be subtracted to obtain
the forest canopy height, which can be correlated
with different forest variables. One main advantage
of using stereo matched VHR images over ALS data,
is the significant lower price ($40/km2 compared to
>$200/km2), and the high repetition frequency of
the satellites passing the region of interest (often
within days or a few weeks). The VHR imagery have
hence appeared as an attractive option to the more
expensive ALS data. The stereo matched heights
have been investigated in a number of papers, but
very few have compared the sensors Pléiades and
WorldView-2, both with image ground sampling
distances of about 0.5 m, in boreal forest (Persson &
Perko 2016; Persson 2016; Immitzer et al. 2016; Yu
et al. 2015; Shamsoddini et al. 2013).
This work evaluates and compares the estimation
of the forest variables Lorey’s mean height, stem
volume and basal-area, using stereogrammetrically
matched VHR imagery from the Pléiades or
WorldView-2 sensors.
Methods
VHR imagery have been stereogrammetrically
matched, using the software Remote Sensing
Package Graz, which utilizes the semi-global
matching algorithm, to derive height rasters. The
processing is further described in (Persson & Perko
2016; Persson 2016). The terrain height (obtained
from a national laser scanning for the duration of
155
Universidad Mayor
Forest Mapping & Inventory
2009 to 2016) was subtracted to obtain canopy
heights. From the derived height rasters, different
metrics, e.g., mean height, standard deviation,
and numerous height percentiles, were extracted
for field plots located on the two test sites. These
metrics were furthermore used as explanatory
variables in multiple linear regression models,
where the models were trained on plots with 10 m
radius, and hereafter the models were evaluated on
independently inventoried 40 m plots.
The performance of respective VHR sensor was
evaluated and compared at two Swedish test sites
illustrated in Figure 1 (Krycklan, Lat. 64°16’N, Long.
19°46’E, and Remningstorp, Lat. 58°30’N, Long.
13°40’E). Systematic 10 m field plots were distributed
at the respective test sites, from which the field
variables in question have been computed, using
established allometric equations. The estimated
variables were HL, BA and VOL. In Krycklan, the
five hundred 10 m plots were inventoried during
the fall 2015, while the thirty-two 40 m plots were
inventoried primarily before the 2016 vegetation
season. In Remningstorp, both the two hundred
sixty 10 m plots and the forty 40 m plots were
inventoried during the fall 2014. At all inventoried
plots, only trees with diameter breast height 0.04
m were calipered. Both the Pléiades and WV2 data
were acquired during 2015.
(a)
Results
From inspection of the model coefficients and
extracted heights, it appears that most WV2 height
percentiles are similar to each other, located close to
the top height, while the Pleiades height percentiles
differ more. That is, the Pléiades sensor appears to
catch a larger dynamic range compared to the WV2
sensor. Moreover, the detected top height appears
generally higher for the WV2 sensor compared to
the Pleiades. However, this might be due to possible
differences on how the acquired bands are used in
the matching. The WV2 images used in this study
were acquired as panchromatic images at 450-800
nm, while the Pléiades images were acquired in four
spectral bands, blue, green, red, and near-infrared,
with the possible (but not used) panchromatic range
of 480-830 nm, and possibly only a single band was
used in the image matching, which might cause the
larger height differences. This is to be clarified.
Estimation results from both sensors at the two
test sites indicate robust and similar results at both
test sites, despite the differences in the forests.
The heights were estimated with an RMSE below
one meter in Krycklan, corresponding to 3% to 6%
RMSE, while the accuracy was almost identical in
Remningstorp, with 5% RMSE (Table 1). As height
was the only variable directly derived from the VHR
(b)
Figure 1. Ortho-rectified Pléiades images of the two test sites superimposed in red. Krycklan (a) and
Remningstorp (b), located in northern (64°N) and southern (58°N) Sweden, respectively, projected in the
UTM 33N coordinate system on the WGS84 reference ellipsoid. ©CNES_2015, distribution Astrium Services
/ Spot Image S.A, France, all rights reserved
156
Universidad Mayor
ForestSAT 2016 Abstracts Summary
imagery, the same source of explanatory variables
(sometimes in transformed forms) were used to
predict HL, VOL and BA. Therefore, the similarity in the
height results is reflected also for the variables VOL
and BA, with almost identical results for respective
sensor, at the respective test site. However, the
across test site values differ slightly, which might be
due to several reasons, including sampling errors,
differences in time for the image acquisitions and the
field samples, and moreover due to the differences
of the forest types. One observation was that the
estimation accuracy of basal-area and stem volume
(a)
was highly correlated, despite that stem volume is
considered a threedimensional unity (including both
height and density of the forest) while basal-area
often functions as one measure of forest density.
The scatter plots were similar between the variables
and sensors, and hence an example on scatter plots
of the HL, VOL and BA estimated from Pléiades
are illustrated in Figure 2 a,c,d and in addition, the
height estimation from WV2 data is also included
(Figure 2 b) for comparison.
(b)
(c)
(d)
Figure 2. Scatter plots of the evaluated predictions from the Krycklan test site. a,c,d are estimations with the
Pléiades sensor and b from the WV2 sensor. a) HL, b) HL, c) VOL, d) BA
157
Universidad Mayor
Forest Mapping & Inventory
Table 1 –Results from evaluation of the independently inventoried plots with 40 m radius.
Test site
Krycklan
Krycklan
Krycklan
Krycklan
Krycklan
Krycklan
Remningstorp
Remningstorp
Remningstorp
Remningstorp
Remningstorp
Remningstorp
Sensor
Pléiades
Pléiades
Pléiades
WorldView-2
WorldView-2
WorldView-2
Pléiades
Pléiades
Pléiades
WorldView-2
WorldView-2
WorldView-2
Variable
HL
VOL
BA
HL
VOL
BA
HL
VOL
BA
HL
VOL
BA
Conclusions
The conclusion is that stereo matching of VHR
satellite images is a promising method for
estimating forest variables, when a high-resolution
terrain model is available. The results from this
study, indicated that the dynamic range of Pléiades
heights is larger than the heights obtained fromWV2
data. The BA and the VOL were highly correlated,
considering prediction accuracy, and both sensors
are giving robust and accurate acquisitions at both
the evaluated test sites.
Acknowledgments
The authors would like to thank the AdvancedSAR
FP 7 project under Grant 606971 for financial support
and to enable access to WorldView-2 imagery. The
Swedish National Space Board is acknowledged for
providing the Pléiades images. This work has also
received funding from the Hildur and Sven Wingquist
foundation.
References
Immitzer, M. et al., 2016. Forest Ecology and
Management Use of WorldView-2 stereo imagery
and National Forest Inventory data for wall-towall mapping of growing stock. Forest Ecology
and Management, 359, pp.232–246. Available at:
http://dx.doi.org/10.1016/j.foreco.2015.10.018.
Persson, H., 2016. Estimation of Boreal Forest
Attributes from Very High Resolution Pléiades
Data. Remote Sensing, 8(9), p.736. Available
at:
http://www.mdpi.com/2072-4292/8/9/736/
[Accessed September 6, 2016].
Persson, H.J. & Perko, R., 2016. Assessment of boreal
158
Unit
m
m3/ha
m2/m2
m
m3/ha
m2/m2
m
m3/ha
m2/m2
m
m3/ha
m2/m2
RMSE
0.876
29.1
1.69
0.502
27.9
1.51
1.09
36.1
3.06
1.10
38.4
3.41
RMSE%
5.88
10.2
7.61
3.38
9.80
9.67
4.84
10.9
10.1
4.88
11.6
11.2
forest height from WorldView-2 satellite stereo
images. Remote Sensing Letters, 7(12), pp.1150–
1159. Available at: https://www.tandfonline.com/
doi/full/10.1080/2150704X.2016.1219424.
Shamsoddini, A., Trinder, J.C. & Turner, R., 2013. Pine
plantation structure mapping using WorldView-2
multispectral image. International Journal of
Remote Sensing, 34(11), pp.3986–4007. Available
at: <Go to ISI>://WOS:000315384500016.
Yu, X. et al., 2015. Comparison of laser and stereo
optical, SAR and InSAR point clouds from airand space-borne sources in the retrieval of forest
inventory attributes. Remote Sensing, 7(12),
pp.15933–15954.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Regional predictive mapping of paludification black spruce
forests in the north eastern Canada using remote sensing and
statistical modeling
Osvaldo Valeria1*, Nicolas Mansuy1, Ahmed Laamrani2, Nicole Fenton1 and Yves Bergeron1
2
1
Forest research institute, Université du Québec en Abitibi Témiscamingue, Québec, Canada.
Agriculture and Agri-Food Canada: Guelph Research and Development Centre, Ontario, Canada.
*Osvaldo.valeria@uqat.ca
Abstract
The black spruce forests in north eastern Canada are considered as a potential source of wood fibre.
However, a considerable volume of timber in this region is located in areas that are prone to gradual
accumulation of surface organic matter deposits over time (paludification). Paludification is a natural
process where organic material accumulates on the ground surface over time, resulting in higher
soil moisture levels and elevated water tables (Crawford et al., 2003; Vygodskaya et al., 2007). These
conditions alter dynamic succession and favour the invasion of Sphagnum moss species (Fenton et al.,
2005), which can lead to the development of forested peatlands and substantial decreases in forest
productivity (Simard et al., 2007). Many parts of the world, including interior Alaska, the western Siberian
plain, and the Hudson Bay-James Bay Lowlands of Canada, are prone to paludification. In the black spruce
forests of the Abitibi and north of Quebec (A/NQ) regions, time-since-last fire, superficial deposit and flat
ground surface topography have been reported as the main factors that cause paludification (Laamrani
et al., 2014). For effective sustainable land resource management and policy making in northern black
spruce forests, an up-to-date and reliable predictive spatial map of paludification at the regional scale
is needed. The aim of this study was to predict paludified areas at a regional scale by parametrizing soil
organic layer-landscape models using a machine learning method (random Forest) and Maxent models.
Organic layer thickness (OLT) was predicted at a regional scale to map paludification. Most important
predictive variables were environmental data representing terrain attributes derived from digital
elevation models, landscape data derived from remotely sensed data, provincial Forest Inventory data
derived from Maps, and data on soil moisture parameters derived from Radarsat-2 data. Training sites
were constructed from OLT experimental sites; and the resulting regional map were validated with a set
of punctual OLT field-measured observations (~2500 points) spread over the study area. Knowledge of
the spatial extent of paludification at the regional scale is of great importance to forest industry and its
forest managers because it would allow them to make better cost-effective management decisions that
optimize forest productivity and ensure sustainability of the forest.
topography on the spatial distribution of organic
layer thickness in a paludified boreal landscape.
Crawford, R.M.M., Jeffree, C.E., and Rees, W.G. Geoderma, 221-222: 70–81.
(2003). Paludification and forest retreat inSimard, M., Lecomte, N., Bergeron, Y., Bernier, P.Y.,
northern oceanic environments. Annals of and Pare, D. (2007). Forest productivity decline
Botany, 91: 213–226.
caused by successional paludification of boreal
Fenton, N., Lecomte, N., Légaré, S., and Bergeron, soils. Ecological Applications, 17(6): 1619–1637.
Y. (2005). Paludification in black spruce (PiceaVygodskaya, N.N., Groisman, P.Y., Tchebakova,
mariana) forests of eastern Canada: Potential N.M., Kurbatova, J.A., Panfyorov, O., Parfenova,
factors and management implications. Forest E.I., and Sogachev, A.F. (2007). Ecosystems and
Ecology and Management, 213(1-3): 151–159.
climate interactions in the boreal zone of northern
Laamrani, A., Valeria, O., Fenton, N., Bergeron, Y., Eurasia. Environmental Research Letters, 2(4):
and Cheng, L.Z. (2014). The role of mineral soil 1–7.
References:
159
Universidad Mayor
Forest Mapping & Inventory
Scale-dependent mapping of stand structural
heterogeneity from airborne LiDAR data
Collins Byobona Kukunda1 and Philip Beckschäfer1
Chair of Forest Inventory and Remote Sensing, Georg August Universität Göttingen,
Germany Corresponding email: ckukund@gwdg.de
1
Key words: Scaling problems, Forest structure, Mixed effects models,
airborne LiDAR, Spatial cofounding
Abstract
Heterogeneity in forest structure, naturally occurring or induced by disturbance, is continuous in space
and time. In practice, heterogeneity in structure is quantified from ecological or forest inventory data;
often bound to observations made on sample plots. For any given location, the plot-based quantities
of structure are known to vary for different plot sizes due to differences in unobserved neighborhoods
at plot boundaries. In regression, this scale dependence may confound relationships between the
forest structural indices and their predictors affecting the resulting map accuracies. Our study first
investigated for plot size effects on variability in forest structure described by forest structural indices.
We thereafter explored the relationship between the indices of structure and heterogeneity and the
plot sizes in forest stands with varying degrees of structural complexity. Finally, the observed plot size
effects were modeled statistically and multiple regressions used to map structural complexity and
heterogeneity over unobserved parts of the study area from airborne LiDAR data. Three forest structural
indices were considered: the aggregation index of Clark and Evans, the Structural Complexity Index
and the Enhanced Structural Complexity Index. We used inventory data from one fully mapped 28.5
ha plot in a semi-natural mature deciduous forest stand and 23 fully mapped one-hectare inventory
plots spread in different temperate forest types in central Germany. To study the plot size effects, the
structural indices were quantified on the basis of 18 plot sizes from 0.1 to 9.8 ha simulated on the 28.5
ha plot and 10 plot sizes from 0.1 to 1 ha simulated on the 23 one-hectare inventory plots. A fixed effects
analysis was used to model plot size effects across levels of stand complexity. In addition, a structural
equation model was used to explain the effect of differences in the plot sizes on multiple regressions
between LiDAR derived canopy metrics and the forest structural indices. Resultant map accuracies
were assessed using the Root Mean Square Error (RMSE) and the True Skills Statistic (TSS) in a leaveone-out cross validation procedures for larger plot sizes and an independent set of data collected on
500 m2 inventory plots. Preliminary results show that all structural indices were influenced by the
plot size and LiDAR data is a good predictor of forest structural complexity and heterogeneity (RMSE
≈ 20%). The highest map accuracy so far (TSS ≈ 0.71) has been obtained at the scale shown by the
structural equation model to have minimal effects of spatial cofounding. These findings are relevant
to optimize plot sizes for efficient inventory of components of forest structure as well as for the design
of natural resource inventories. The structural complexity and heterogeneity map produced for the
study area will be relevant for guiding further ecological and forest management planning.
160
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Stereo matched very high-resolution satellite images for
predictions of forest variables
Persson, H.J., Olsson, H, Fransson, J.E.S.
Department of Forest Resource Management, Swedish University of Agricultural Sciences, Umeå, Sweden, firstname.
lastname@slu.se
Keywords: WorldView-2, Pléiades, stereo, forest, stem volume, tree height, basal area
There is a steady need of establishing and updating
information about the forest. Airborne laser
scanning (ALS) has for more than one decade been
considered the most accurate remote sensing
technique that in combination with field samples
can be used to create wall-to-wall estimations of
typical forest variables, like Lorey’s mean height
(HL), basal-area or stem volume. Many countries,
including Sweden, can now offer complete national
digital terrain models (DTMs) based on ALS, which
has enabled also other techniques for estimations of
this type of variables. One such technique is stereo
matching of very high resolution (VHR) satellite
images, which often possess resolutions below 1
meter. The matched result is a digital surface model
(DSM) from which the DTM can be subtracted to
obtain the forest canopy height.
In the current study, stereo matched WorldView-2
and Pléiades data have been compared for two
Swedish test sites (Krycklan, Lat. 64°16’N, Long.
19°46’E, and Remningstorp, Lat. 58°30’N, Long.
13°40’E). Systematic 10 m field plots were distributed
at the respective test sites, from which the field
variables in question have been computed, using
established allometric equations. The estimated
variables were HL, basal-area (BA) and stem volume
(VOL).
One main advantage of using stereo matched VHR
images over ALS data, is the significant lower price
($40/km2 compared to >$200/km2).
The conclusion is that stereo matching of VHR
satellite images is a promising method for estimating
forest variables, when a high-resolution DTM is
available. The results from this study, indicated
that the accuracies fore HL are similar to those
achieved from similar ALS based estimations, while
the estimations of BA and VOL are not sufficient by
using only stereo matched heights.
Studies from Germany and Finland have used stereo
matched WorldView-2 images to estimate forest
variables, and there are a few similar studies which
have considered Pléiades data for estimations of
other types of variables.
The results show that HL could on plot level be
estimated with 8% to 10% RMSE, the basal-area
with 24% to 29% RMSE and the stem volume with
30% to 33% RMSE. The results seem robust across
the different test sites and across different tree
species.
161
Universidad Mayor
Forest Mapping & Inventory
POSTER
The 2016 NASA AfriSAR campaign for tropical forest
structure and biomass measurements: design,
execution and first results
Lola Fatoyinbo1, Christy Hansen1, Naiara Pinto2, Michelle Hofton3, Bryan Blair1, Sassan
Saatchi2, Marc Simard2, Yunling Lou2, Ralph Dubayah3, Scott Hensley1,
Laura Duncanson1,4, Marco Lavalle2.
NASA Goddard Space Flight Center, Greenbelt, Maryland
NASA Jet Propulsion Laboratory, California Institute of Technology
3
University of Maryland, College Park, Maryland
4
NASA Postdoctoral Program/USRA, Columbia, Maryland
1
2
Keywords: Forest Canopy Height, Biomass, SAR, Lidar, Tropical Forests, Airborne Campaigns
Background
The AfriSAR campaign was a joint NASA and ESA
airborne campaign conducted in Gabon in support
of the upcoming ESA BIOMASS, NASA-ISRO
Synthetic Aperture Radar (NISAR) and NASA Global
Ecosystem Dynamics Initiative (GEDI) missions. The
aim of the campaign was to collect ground, airborne
SAR and airborne Lidar data for the development
and evaluation of forest structure and biomass
retrieval algorithms. The campaign consisted of
2 deployments, the first in 2015 with the ONERA
SETHI SAR system and the second in 2016 with the
NASA LVIS (Land Vegetation and Ice Sensor) Lidar,
the NASA L-band UAVSAR and the DLR F-SAR. In
addition, field teams from the Gabon ANPN (Agence
Nationale des Parcs Nationaux), University College
London and NASA were collecting ground data.
Here we focus on the 2016 NASA contributions to
campaign.
Aim
The objectives of the ESA/NASA AfriSAR
deployments were to: 1) measure forest canopy
height, canopy profiles and biomass density under
a variety of forest conditions such as primary and
degraded forest and a variety of forest types,
including tropical rainforest, mangroves, forested
freshwater wetlands and savannas 2) acquire
detailed measurements of airborne SAR data
and Lidar data for cross calibration of NASA and
ONERA/DLR instruments and for CAL/VAL support
162
of the BIOMASS, NISAR, and GEDI missions, and
3) conduct technology demonstrations of joint SAR
and Lidar applications.
Methods
The NASA AfriSAR campaign involved 2 aircrafts:
the LaRC (Langley Flight Research Center) B-200
carrying LVIS and the AFRC (Armstrong Flight
Research Center) C-20 carrying the L-band UAVSAR.
The aircrafts were based out of Libreville, Gabon
for 3 weeks in February 2016, jointly with the DLR
aircraft and SAR instrumentation.
Results
A central motivation for the AfriSAR deployment
was the common biomass requirement for the three
future spaceborne missions and the lack of sufficient
airborne and ground calibration data covering the
full range of biomass in tropical forest systems.
During the campaign, the NASA sensors collected
Tomographic SAR, Polarimetric InSAR, Polarimetric
SAR and Waveform Lidar data over six joint sites.
These sites were the Mondah Forest, Lope National
Park, Pongara National Park, Mabounie, Rabi, and
the Lower Ogooué River. Additional data was also
collected over additional sites and in long transects.
In total, over 70 hours of science data were collected.
In this presentation, we will present the design and
execution of the NASA AfriSAR campaign, and show
some of the first results and data products from the
campaign.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Uncertainty estimation of stand structural variables in
LiDAR-based forest inventories at different sample sizes.
Sastre L, Mauro F*, Pascual C, Gomez-Roux M, Manzanera JA, Garcia-Abril A.
Research Group for Sustainable Management SILVANET. FoReStLab. College of Forestry and Natural Environment,
Technical University of Madrid, Ciudad Universitaria, 28040 Madrid, Spain.
*: Oregon State University, Corvallis, OR, USA.
Corresponding author: J.A. Manzanera (joseantonio.manzanera@upm.es).
Keywords: EBLUP, forest inventory, LiDAR, Small Area Estimation, uncertainty estimation.
Abstract:
The aim of this work seeks to analyze an improvement in the quality of the forest attribute estimates
of a forest inventory when we include auxiliary data coming from a LiDAR (Light Detection And
Ranging) dataset. Multiple regression models were designed for the basal area, density, quadratic
mean diameter, dominant height, biomass and volume variables. The estimation errors for these
variables were analyzed using random sampling methods, and reducing the number of sample plots.
One of the main conclusions of this study is that the inclusion of LIDAR-derived metrics significantly
reduces the sampling effort and improves the precision of the estimation. The inclusion of modelbased and Empirical Best Linear Unbiased Predictor estimations further improves the accuracy at the
plot and stand levels.
Introduction
Forest structure is defined as the spatial distribution
of both living and dead vegetation size, age and
species, with special emphasis on the arboreal
component (Spies & Franklin, 1991). Deepening the
characterization of the structure affects the progress
of Global carbon cycle studies, forest productivity,
use of cover habitat by birds, arboreal mammals and
arthropods, interaction between forests and rivers,
and in the prediction of fire behavior (Means et al.,
2000).
The most relevant aspects of the study of the
structure in the forest stands are the distribution
of trees, the specific composition both in species
diversity and in their distribution in the stand and
the differentiation in diameter, height and size of
crowns, as well as the Different vertical strata (M. del
Río 2003). Forest attributes have traditionally been
mapped using passive airborne or satellite sensors
and statistical methods (McRoberts and Tomppo
2007). However, optical sensors have important
limitations in quantifying vegetation characteristics
because they only generate information in two
dimensions. In addition, in dense coverage areas, the
highly reflected energy tends to saturate the signal
captured by the sensor (Lefsky, Cohen, Harding, et
al., 2002), making it impossible to make estimates
of the different dasometric variables above a certain
threshold. In this context, LiDAR technology is able
to traverse the forest cover and provide information
of its three-dimensional dimension. Since it does
not present the problem of signal saturation, LIDAR
allows to evaluate the three-dimensional patterns
of the arboreal canopy and to estimate the vertical
structure of the plant communities (Lefsky, Cohen,
Harding, et al., 2002). The three-dimensional
information provided by the LiDAR sensor has been
used successfully in the estimation of different forest
parameters, such as canopy height, wood volume,
etc., automatically and with greater accuracy than
that achieved with techniques of traditional inventory
(Sithole and Vosselman 2005) or other approaches
with optical sensors (Maltamo et al., 2006).
163
Universidad Mayor
Forest Mapping & Inventory
Managers need up-to-date information at the
time of decision-making on the status, structure
and composition of the forest stands, provided
individually for each of the management units and
subunits. This goal can be achieved by different
statistical techniques, such as model-assisted,
model-based, indirect estimation fixed or mixed
effects estimation (Mauro Gutiérrez et al. 2013,
2015).
Materials and Methodology:
This analysis is carried out on the Peña del Águila,
mountainside, located in Fuenfria ́s Valley (Cercedilla,
Madrid). To this day, the LiDAR flight has been used
in 2011 and in two field inventories; one of them was
made in 2013/2014, covering 60 sampling plots, and
the other one made in 2015, covering 80 sampling
plots. Multiple regression models were designed for
the basal area, density, quadratic mean diameter,
dominant height, biomass and volume variables.
The errors committed in estimating these forest
attributes were analyzed from models and by using
random sampling methods, reducing the number of
sampled plots.
Results:
It was found that adding LiDAR ́s information
significantly reduces the forest inventory errors.
Secondly, stand, pixel and mount level estimates
were carried out using sampling design and model
based techniques and the errors committed
were compared. The first technique included the
simple random sampling and the model assisted
estimation; the second one embraced Small Area
Estimation (SAE) techniques. Moreover, the model
based estimation allowed to develop a pixel and
stand level error mapping.
Finally, outcomes showed that EBLUP estimators
(Empirical Best Linear Unbiased Predictor) brought
out more accurate estimates tan those coming from
sampling design based techniques, in population
subunits (stands).
Conclusions:
The main conclusion of this study is that the
inclusion of LIDAR-derived metrics significantly
reduces the sampling effort and improves the
164
precision of the estimation. The inclusion of modelbased and Empirical Best Linear Unbiased Predictor
estimations further improves the accuracy at the
plot and stand levels.
References.
Lefsky, Michael A, Warren B Cohen, David J Harding,
et al. 2002. “Lidar Remote Sensing of aboveGround Biomass in Three Biomes.” Remote
Sensing of Environment: 393–99.
Lefsky, Michael A, Warren B Cohen, Geoffrey G
Parker, and David J Harding. 2002. “Lidar Remote
Sensing for Ecosystem Studies.” BioScience
52(1): 19–30.
Maltamo, M et al. 2006. “Nonparametric Estimation
of Stem Volume Using Airborne Laser Scanning,
Aerial Photography, and Stand-Register Data.”
436: 426–36.
Mauro Gutiérrez, F., García García, D., García
Abril, A., Martín-Fernández, S., Núñez Martí,
M.V., Gonzalez García C., Ayuga Téllez E. 2013.
“Reducción Del Numero de Parcelas de Muestreo
Al Incorporar Información Auxiliar LiDAR En
La Estimación de Variables Dasométricas.” 6º
Congreso Forestal Español: 1–13.
Mauro Gutiérrez, Francisco. 2015. “Estimación de
Variables Dasométricas a Partir de Datos LiDAR
Y Obtención de Modelos de Referencia Para
Las Distribuciones de Alturas Y Diámetros Del
Arbolado.” PhD Thesis, UPM.
McRoberts, Ronald E., and Erkki O. Tomppo. 2007.
“Remote Sensing Support for National Forest
Inventories.” Remote Sensing of Environment
110(4): 412–19.
Means, Joseph E et al. 2000. “Predicting Forest
Stand Characteristics with Airborne Scanning
Lidar.” 66(11): 1367–71.
Río, M. del. 2003. “Índices de Diversidad Estructural
En Masas Forestales.” Forest Science (January).
Sithole, George, and George Vosselman. 2005.
“Filtering Of Airborne Laser Scanner Data
Abstract :” Photogrammetric Engineering and
Remote Sensing: 66–71.
Spies, Thomas A, and Jerry F Franklin. 1991. “The
Structure of Natural Young , Mature , and OldGrowth Douglas-Fir Forests in Oregon and
Washington.”
Forest Modelling
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Data assimilation of InSAR-based estimated
forest stand attributes
Lindgren, N.1, Persson, H.J.1, Nyström, M.1, Grafström, A.1, Nyström, K.1, Muszta, A.1,
Wallerman, J.1, Willén, E.2, Fransson, J.1, Ståhl, G.1, Olsson, H.1
Department of Forest Resource Management, Swedish University of Agricultural Sciences,
Umeå, Sweden, firstname.lastname@slu.se
2
The forestry research institute of Sweden, Uppsala, Sweden, firstname.lastname@skogforsk.se
1
Keywords: Data assimilation, TanDEM-X, InSAR, extended Kalman filter,
forest inventory, growth functions, time series
Up-to-date, cost-efficient and accurate information
of the state of the forest is important to manage
and monitor forests. Data assimilation can be
used to optimally combine forecasts of previous
estimates with new observations of the current
state, based on the uncertainties in the forecast and
the observations. In other fields of study, such as
robotics and meteorology, data assimilation has led
to better accuracies and reduced need for reference
data. Early studies have also shown a potential for
keeping forest stand registers up-to-date, using
simulation (Ehlers et al., 2013) as well as empirically
with a time series of canopy height data from digital
photogrammetry (Nyström et al., 2015).
In the present study, data assimilation through
extended Kalman filtering was used to assimilate
a series of stand attribute predictions made from
TanDEM-X InSAR data. In total a time series of 19
InSAR images acquired between 2011 and 2014 were
used in combination with a high-resolution (2x2 m)
digital terrain model. Forest variables were predicted
for each InSAR image, using empirical regression
estimates with forest data from field plots as
response variables. Each prediction was assimilated
with forecasts from the previous time point.
Both assimilated estimates and estimates based
on a single acquisition were evaluated using crossvalidation on a set of 137 sample plots with 10 m
radius inventoried twice, in 2010 and 2014. Linear
interpolation was used to create reference data for
the intermediate years, and evaluation was made
with the reference data matching the acquisition
date.
The results show that, at all assimilated time points,
data assimilation resulted in better estimates than
estimates based on the single acquisitions. Median
reduction in RMSE was 0.4 m for Lorey’s mean
height, 0.9 m2/ha for basal area and 15.3 m3/ha for
stem volume. The conclusion is that data assimilation
is a promising method for estimating forest stand
attributes when multiple data sets from different
acquisition time points are available, for example
the continuous flow of data from TanDEM-X.
167
Universidad Mayor
Forest Modelling
Delineation of forest structure patterns in the circumpolar
taiga-tundra ecotone
Paul M. Montesano, Christopher Neigh, Jon Ranson
ForestSat 2016
Interactions between broad-scale climate, local-scale
site factors and disturbance history have produced
forest patterns in the taiga-tundra ecotone (TTE)
that include a range of woody cover from forest cover
patches and clusters of shrubs to sparsely arranged
individuals. These biome boundary patterns influence
ecological processes and are evident at local scales,
but are often not apparent in moderate-resolution
imagery from most earth observing satellites. Such
scale problems have contributed to uncertainties in
the extent of the TTE, its structural characteristics,
and its net effect on current and future processes in
the high northern latitudes. Recent advances using
the global archive of Landsat data have provided
calibrated tree cover at 30-m resolution, which
may help improve TTE delineation. Furthermore,
168
U.S. federal access to sub-meter DigitalGlobe data
provides the means to improve estimates of the TTE
forest patterns with observations of individual trees
and structure from stereoscopic observations. A
supercomputer cluster with cloud storage facilitates
(1) image-based sampling of the entire TTE domain
at the scale of forest structure change, (2) examining
detailed forest structure patterns, and (3) exploring
novel techniques for capturing horizontal and vertical
forest structure. Site-scale (~1 m) maps of forest cover
and height in sample strips across the TTE domain
will be used to optimize the medium-scale (~30 m)
delineation of this biome boundary and examine the
spatial variability of detailed forest structure patterns.
The resulting products may refine input to ecosystem
models while providing insight into the variability
of recent structural changes. An overview and early
results of our current work will be presented.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Detecting the spread of invasive tree species in central Chile
with combined Landsat and Sentinel-2 data
Tobias Schmidt1, Michael Foerster1, Birgit Kleinschmit1, Julian Cabezas2, Fabian Fassnacht2
1
Institute for Landscape Architecture and Environmental Planning, Technical University of Berlin
2
Institute for Geography and Geoecology, Karlsruhe Institute of Technology (KIT)
Keywords: One-Class-Classifier, Species Distribution Modelling, Pinus radiata, Ulex europaeus, Acacia
dealbata, Sentinel-2, Landsat
Chile has a large number of endemic species due
to its isolated location and is therefore one of the
biodiversity hots-spots of the Planet. At the same
time a number of invasive species occurred over
the last decades and have shown to have a notable
negative effect on Chilean ecosystems. Between
2016 and 2018, the project SaMovar (Satellite-based
Monitoring of invasive species in central-Chile) will
investigate the past and recent spread of selected
invasive species.
Ulex europaeus, Acacia dealbata and Pinus radiata
have a large impact on local and regional biodiversity
in central-Chile. Two objectives of the project are
the mapping of the current state of spread and the
identification of the historical invasion dynamics of
the three species. Moreover, the future spread of
the invasive species will be spatially estimated to
produce risk maps for management measures. For
these objectives the application of multi-temporal
and multi-sensor satellite data (especially Sentinel-2
and Landsat) in combination with Unmanned Aerial
Vehicle (UAV) data will be pursued.
In a first step, the most recent distribution ranges
of the three target species will be detected using
the data of both sensors by the application of oneclass-classifier (OOC) techniques (e.g. Maxent).
Presence information of the target species will be
derived from spatial and spectral high resolution UAV
data. To increase the model accuracy, the different
phenological phases of the species can be considered.
Due to the high temporal resolution of both sensors, it
is possible to assess which combination of acquisition
dates of each sensor is leading to the result with the
significantly highest model accuracy. For this purpose
the McNemar test will be applied to find significant
differences in between multiple combinations of
images and the subsequently derived classifications
with a semi-exhaustive feature selection approach. As
a result we can detect optimal acquisition dates, which
will be used to compare the capability of Landsat and
Sentinel-2 to distinguish between native vegetation
and invasive species. Finally, the information on the
classification quality of the different acquisition dates
can be used for the identification of the historical
spread of the three target species using the Landsat
archive. For validating the retrospective classifications
historical aerial photos from the last 30 years will be
used. .
Together with the disturbance history of the region
and additional environmental features (e.g. digital
elevation model, climate data) this information will
be used as input layer to species distribution models
which will be developed to model the future spread
of the three target species assuming different
climate scenarios.
Preliminary results of the project are presented,
especially regarding the identification of Pinus
radiata in the natural forest remnants of the Maule
Region using the already collected UAV data and
first results of the classification using single Landsat
and Sentinel-2 images.
169
Universidad Mayor
Forest Modelling
How much forest area should be sampled to get accurate
biomass estimations at different scales?
Rico Fischer1, Jessica Hetzer1, Andreas Huth1
Department of Ecological Modelling, Helmholtz Centre for Environmental Research – UFZ, Permoserstr. 15, 04318
Leipzig, Germany
1
Keywords: Tropical Forest, Biomass, Sampling, Model
Abstract:
Tropical forests play an important role in the global carbon cycle. Field inventory plots were used for
understanding forest structure and dynamics at different scales under the assumption that these
plots accurately represent their surrounding landscape. Here, we tested at different scales whether
inventory plots meet this assumption for biomass estimations in tropical forests. In a first step, we
investigated this assumption on a local scale. Using a large 50-ha forest inventory plot from Panama we
analyzed how representative 1-ha subplots are for biomass estimations. Results showed that about 6
of these 1-ha subplots are needed to accurate estimate biomass for this region even at its most reliable
sampling strategy. Using a mathematical model we found out that for accurate biomass estimations
in less homogenous forest sites (such as disturbed forests or montane forests) the number of needed
field plots increased by 60-80%. In a second step, we used larger data sets to test different sampling
strategies for accurate estimates of forest biomass at regional scale. It turned out that the required
number of inventory plots is high. To overcome the limited number of inventory plots for accurate
tropical forest biomass estimations the use of inventory plots in combination with high-resolution
remote sensing products could be one promising solution.
170
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Lessons Learned – National Individual Tree Species Extent and
Parameter Modeling for Insect and Disease Risk Mapping
James Ellenwood [US Forest Service – R&D], Frank Krist [US Forest Service – S&PF/FHTET]
Keywords: Forest modeling, Remote sensing, Landsat, Tree Species Distributions, Insect and Disease Risk
Parameter datasets of presence [Ellenwood 2015
and others], basal area (BA), and stand density index
(SDI) were modeled for 289 individual tree species
using forest inventory data, Landsat imagery, and
several national datasets for the coterminous US
and Alaska at a 30-meter resolution to support
the 2013-2027 National Insect and Disease Risk
Map – NDRIM (Krist and others 2014). NIDRM is a
nationwide strategic assessment of potential tree
mortality due to insects and diseases. To facilitate
local and regional planning efforts and develop
models of forest parameters by tree species, USFS
Forest Inventory and Analysis (FIA) data were linked
to each national data layer and analyzed with See5
and Cubist data mining software (Quinlan 2012).
Model outputs were processed into a 30-meter
spatial dataset
A number of issues were identified in the completion
of this project that would be valuable for future
considerations in subsequent projects. Three issues
are presented: phenology representation, plot-pixel
mis-match, and exceptional areas.
This project utilized three-season Landsat 5/7
imagery to represent vegetation phenology.
Though somewhat useful for differentiating
conifer and hardwood species, the coarse temporal
representation is highly variable with respect to
plant phenology. Individual species have different
phenological responses and the subtle seasonal
changes are not decipherable from the threeseason dataset. Observed in the individual species
models, reflectance data was seldom selected as
a significant variable for most of the 289 species.
A possible explanation is the “noise” created by
different phenological stages not represented by
the three-season images. Newer techniques that
composite multiple images to achieve a cloud-free
representation may exacerbate this situation. Tools
which create time-integrated images as smoothed
seasonal representations show promise.
The plot clusters do not represent “whole” pixels.
This project attempted to mitigate this by utilizing
individual subplots within the cluster as separate
samples. Confidence images for each of the
parameter rasters indicated a very high degree of
variation. It is surmised that a substantial portion
of this variation is due to the plot-pixel size mismatch. Options to either increase plot size through
two-stage sampling or reducing pixel size through
incorporation of “pan” images are examined.
Though the plot design creates a representative
sample for the Nation, exceptional areas are difficult
to represent, yet the modeling process attempts to
apply to all areas. For this project, parameters were
limited to no greater than 10% extrapolation. This
may not be optimal for biological phenomenon.
Creating separate layers of extrapolated parameters
may yield insightful information for subsequent
product utility.
Citations:
Ellenwood JR, Krist FJ, Romero SA. National
Individual Tree Species Atlas. FHTET-15–01.
Fort Collins, Colorado: U.S. Department of
Agriculture, Forest Service, Forest Health
Technology Enterprise Team; 2015.
Krist FJ, Ellenwood JR, Woods M, McMahan A,
Cowardin J, Ryerson D, Sapio F, Zweifler M,
and Romero SA. 2013–2027 National Insect and
Disease Forest Risk Assessment. FHTET-14-01.
Fort Collins, Colorado: U.S. Department of
Agriculture, Forest Service, Forest Health
Technology Enterprise Team; 2014.
Quinlan R. Cubist/See5 [computer programs].
Rulequest, Inc. Available from: URL: http://www.
rulequest.com; 2012.
171
Universidad Mayor
Forest Modelling
Modeling forest structure and aboveground biomass
integrating airborne LiDAR and satellite Radar data
Mariano Garcia1,2, Sassan Saatchi1 and Heiko Balzter2,3
Jet Propulsion Laboratory (JPL), California Institute of Technology, Pasadena, CA 91109, USA.
University of Leicester, Centre for Landscape and Climate Research, Department of Geography, Leicester, LE1 7RH, UK.
3
National Centre for Earth Observation, University of Leicester, Leicester, LE1 7RH, UK.
1
2
Keywords: LiDAR, SAR, aboveground biomass, extrapolation.
Abstract
Improved spatially explicit information on forest structure and aboveground biomass (AGB) is required
to understanding the role of forests in the carbon cycle, for greenhouse gases inventories or for
sustainable forest management. This study aims at developing algorithms to allow the extrapolation
of LiDAR-based forest height estimates over larger regions using satellite Radar data. We selected two
study sites corresponding to a temperate broadleaf and mixed forest. The use of exponential models
showed a poor performance for both study sites. Improved results were obtained using a support
vector regression (SVR) algorithm. Inclusion of additional information from texture metrics as well as
elevation data from SRTM significantly improved the results (R2=0.76 and 0.8; RMSE= 2.63 and 1.72 m
for WLEF and Howland, respectively). Regardless of the modeling approach used, model performance
improved as the resolution decreased from 25 m to 100 m.
Finally, we evaluated the potential of calibrating a single model for both study sites and compared it to
the site-specific calibrated models. No significant differences between the site-specific and the biome
model were observed at WLEF with R2= 0.76 and 0.77 and RMSE= 2.63 and 2.62 m for the site-specific
and the biome models, respectively. For Howland, a decrease in R2 from 0.8 to 0.71 and an increase in
RMSE from 1.72 to 2.04 m was observed.
Introduction
Improved spatially explicit information on forest
structure and aboveground biomass (AGB) is
required to understanding the role of forests in the
carbon cycle, for greenhouse gases inventories or
for sustainable forest management. Field methods
are time consuming and limited in regards the
spatial coverage, making difficult capturing the
heterogeneity of forest structure. The high variability
of forest structure both, at local and landscape
scales, hampers extrapolation of forest structure
estimates to regional or continental scales (Xu et
al., 2016). Remotely sensed data, particularly LiDAR
and radar sensors, offers the opportunity to capture
this spatial variability in forest structure and improve
AGB estimates over larger regions.
172
Two space LiDAR missions are planned for launch
in the coming years, namely the ICESat-2/ATLAS
(Gwenzi and Lefsky, 2014) and the Global Ecosystem
Dynamics Investigation (GEDI; Dubayah,et al.,
2014). Likewise, two satellite radar missions will be
launched. The first is the BIOMASS mission, a P-band
synthetic aperture radar (SAR), which is part of the
ESA’s (European Space Agency) Earth Explorers
(http://www.esa.int/Our_Activities/Observing_
the_Earth/The_Living_Planet_Programme/
Earth_Explorers/Future_missions/Biomass).
The
second is the NASA-ISRO Synthetic Aperture Radar
(NISAR), an L-band InSAR mission resulting from the
partnership between NASA and the Indian Space
Research Organization (ISRO).
Universidad Mayor
The LiDAR sensors will be multi-beam profilers that
will provide a dense sampling of the Earth’s terrestrial
ecosystems; however, in order to provide wall-towall estimates of forest structure integration of
these data with other satellite borne sensors will be
required. This study aims at developing algorithms
to allow the extrapolation of LiDAR-based forest
height estimates over larger regions using satellite
Radar data. We address three research questions: 1)
can we extrapolate LiDAR height using L-band SAR
polarimetry data? 2) At what scale we obtain the best
results? And 3) can we develop a single model for an
entire biome or do we need site-specific models?
Methodology
Study sites
We selected two study sites corresponding to a
temperate broadleaf and mixed forest. The first
study site (WLEF) is located in Park Falls, Wisconsin,
USA. The area selected was approximately 13 x 9 km,
mainly covered by deciduous and mixed coniferous
forests as well as some wetland areas. The second
study area is the Howland forest, Maine, USA. This
a boreal transitional forests in Howland mixed
deciduous-coniferous forest, generally fragmented
due to both natural disturbance and logging or
management practices. This site covered an area of
8 x 5 km.
Datasets
ALOS-PALSAR
We used the high resolution terrain corrected
product developed by the Alaska Satellite Facility
from the Advanced Land Observing Satellite/Phased
Array L-band Synthetic Aperture Radar (PALSAR).
This product combines the radiometric and terrain
correction processes to provide gamma naught
power images at 12.5 m resolution (Gens, 2015).
We selected 4 images for WLEF and 3 images for
Howland from 2010, from May to October.
Airborne LiDAR
Our reference data to validate the height estimations
from SAR data was the 1 m spatial resolution canopy
height model (CHM) product developed by the
G-LiHT (Goddard’s LiDAR, Hyperspectral & Thermal
Imager) team.
Additional data
The additional data consisted of the SRTM V2.1
ForestSAT 2016 Abstracts Summary
product, with a spatial resolution of 30 m and a
Landsat OLI image.
Data processing
Environmental conditions, like vegetation and
soil moisture, have a significant impact in the
backscatter signal (Santoro et al., 2006; Saatchi
et al., 2011). In order to remove this effect, each
pixel value was averaged using all dates available.
Subsequently the images were resampled to 25 m,
50 m and 100 m. We also derived texture features
from the images (standard deviation and entropy)
using 5x5 and 9x9 window sizes. The same texture
metrics were derived from the SRTM, which had
been resampled to the same resolutions used for
the ALOS-PALSAR data. As for the Landsat data,
the spectral information was summarized into the
brightness, wetness and greenness components of
the tasseled cap transformation. The same texture
metrics as for the SAR data were derived for each TC
component.
Height modeling
We used two different approaches to estimate
height from the ALOS-PALSAR data. The first
approach used an exponential model fitted using
the HH, the HV bands or a combination of both. To
fit the models, the mean backscatter value obtained
at each 1 m height interval was computed to reduce
the noise of the backscatter data.
The second approach used a support vector
regression (SVR) machine learning algorithm to
estimate the height. We used the LS-SVMlab
Toolbox developed by De Brabanter et al. (2011).
A radial basis function was used as kernel and the
two parameters defining the bandwidth (h) and
the regularization parameter (g), which defines
contribution of the training data error in the loss
function to be minimized, were calculated using a
grid search with a ten-fold cross-validation.
We tested different models based on the ALOSPALSDAR data alone and consecutively adding
additional information from the SRTM and the
Landsat sensors. Approximately 10% of the data were
used for calibration and 90% for validation. Finally,
in order to evaluate the possibility of developing
a single model for both study sites, a model was
trained using sample data from both study sites and
compared to the site-specific models.
173
Forest Modelling
Results
For both study sites, the exponential model based on
HV band gave the best results due to the dominance
of volume scattering from the canopy. Regardless
of the band used, an increase in R2 and a decrease
of RMSE was observed as the spatial resolution
decreased from 25 m to 100 m (Figure 1).
SVR models trained using HV data improved R2
more than 10% (WLEF: 0.43-0.58; Howland: 0.490.59 at 100 m) and reduced the RMSE by almost 40%
(WLEF: 5.61-3.49 m; Howland: 3.71 – 2.42 m at 100 m).
As for the parametric models, the best results were
obtained at 100 m resolution. Inclusion of additional
data improve the results significantly at both sites.
For WLEF, the use of texture metrics derived using
a 9x9 window and the elevation information from
Universidad Mayor
the SRTM provided the best results with R2=0.76
and RMSE 2.63 m. Additional information including
texture from SRTM and spectral and textural
information from Landsat did not improve the
results. For Howland, texture information provided
important information for the height estimation
given the fragmented landscape of this study site.
Therefore, the best results were achieved using
ALOS bands and their texture information along
with the elevation and texture information derived
from the SRTM data. The R2 obtained for this model
at 100 m resolution was 0.8 and the RMSE was 1.72
m. Figure 2 shows the R2 and the RMSE achieved for
each of the models trained.
The calibration of a single model (biome model) for
both study sites showed good performance over
both sites. Thus, no significant differences between
Fig. 1: R2 (bars) and RMSE (lines) values obtained for the each parametric model and spatial resolution at
each study site.
Fig. 2: R2 (bars) and RMSE (lines) values obtained for the each SVR model and spatial resolution at each
study site.
174
Universidad Mayor
the site-specific and the biome model were observed
at WLEF with R2= 0.76 and 0.77 and RMSE= 2.63 and
2.62 m for the site-specific and the biome models,
respectively. For Howland, a decrease in R2 from 0.8
to 0.71 and an increase in RMSE from 1.72 to 2.04 m
was observed.
Conclusions
In this study we have demonstrated the LiDARbased height can be extrapolated to larger areas
using L-band SAR data in temperate broadleaf
and mixed forests. In order to reduce the effect of
environmental conditions on the backscatter signal,
multitemporal data are required. NISAR will provide
12 day global observations, which will allow for an
improved reduction of the effects of environmental
conditions. The use of SRTM data help improve
height estimates, particularly over areas with
rougher topography, whereas texture information
improve results over fragmented landscapes. Non
parametric algorithms (SVR) showed greater ability
to model canopy height than parametric models,
allowing to better capture complex relationships
between the backscatter signal and the LiDARbased height. In all cases, spatial resolution affected
significantly the estimations, with the best results
achieved at 100 m spatial resolution. Our results are
of significance for the integration of data provided
by future space missions like GEDI and NISAR.
References
De Brabanter, K., P. Karsmakers, F. Ojeda, C. Alzate,
J. De Brabanter, K. Pelckmans, B. De Moor, J.
Vandewalle, and J. A. K. Suykens (2011), LSSVMlab Toolbox User’s Guide. Version 1.8,
Katholieke Universiteit Leuven. Department
of Electrical Engineering, ESAT-SCD-SISTA.
Kasteelpark Arenberg 10, B-3001 LeuvenHeverlee, Belgium.
Dubayah, R., Goetz, S., Blair, J.B., Luthcke, S.,
Healey, S., Hansen, M., et al. (2014a). The Global
Ecosystem Dynamics Investigation (abstract).
ForestSAT 2016 Abstracts Summary
ForestSAT 2014 Conference, Nov. 4-7, 2014, Riva
de Garda, Italy.
Gens, R. (2015). ASF radiometric terrain corrected
products. Algorithm Theoretical Basis Document.
Gwenzi, D., Lefsky, M.A., Suchdeo, V. P. and
Harding, D.J, (2016). Prospects of the ICESat-2
laser altimetry mission for savanna ecosystem
structural studies based on airborne simulation
data. ISPRS Journal of Photogrammetry and
Remote Sensing, 118, 68-82.
Saatchi, S.S., Marlier, M., Chazdon, r. L., Clark, D.,
B. and Russell, A.E. (2011). Impact of spatial
variability of tropical forest structure on radar
estimation of aboveground biomass. Remote
Sensing of Environment, 115, 2836-2849.
Santoro, M. Eriksson, L., Askne, J. and Schmullius, C.
(2006). Assessment of stand-wise stem volume
retrieval in boreal forests from JERS-1 L-band
SAR backscatter. International Journal of Remote
Sensing, 27:16, 3425-3454.
Xu, L., Saatchi, S.S., Yang, Y., yu, Y. and White,
L. (2016). Performance of non-parametric
algorithms for spatial mapping of tropical forest
structure. Carbon Balance and Management, 1118.
Acknowledgements
Mariano Garcia is supported by a Marie Curie
International Outgoing Fellowship within the 7th
European Community Framework Programme
(ForeStMap - 3D Forest Structure Monitoring and
Mapping, Project Reference: 629376). The contents
on this paper reflect solely the authors’ views
and not the views of the European Commission.
Heiko Balzter was supported by the Royal Society
Wolfson Research Merit Award, 2011/R3 and the
NERC National Centre for Earth Observation. The
research was carried out at the Jet Propulsion
Laboratory, California Institute of Technology, under
a contract with the National Aeronautics and Space
Administration.
175
Universidad Mayor
Forest Modelling
POSTER
Modeling of ALS data statistics in tree-level – application to
single tree detection using Bayesian inference
Teemu Luostari1, Timo Lähivaara1, Petteri Packalen2, Aku Seppänen1
1
Department of Applied Physics, University of Eastern Finland, Finland
2
School of Forest Sciences, University of Eastern Finland, Finland
Keywords: remote sensing, forest inventory, Bayesian inference, uncertainty quantification,
modeling, single tree detection
Abstract
The analysis of Airborne Laser Scanning (ALS) data can be divided into two categories: area based
methods and single tree detection. While the area based methods aim to predict the plot-level statistics
directly from ALS data using regression-type methods, in the single tree detection, the plot-level
statistics are derived from individual tree information inferred from the ALS data. The latter approach
was taken, e.g., in (Andersen, Reutebuch, & Schreuder, 2002), (Lähivaara et al., 2014), (Micheas, Wikle,
& Larsen, 2014), where the detection of individual trees and estimation of their dimensions was based
on fitting a set of canopy height models (CHM) to the ALS data. In all these papers, the problem of single
tree detection was considered in the framework of Bayesian inference (Kaipio & Somersalo, 2005). The
Bayesian framework allows for incorporating prior information on the allometric relations of trees
into the ALS based estimates of tree crown shapes. Such information can improve the reliability of the
tree detection significantly: In (Lähivaara et al., 2014) the Bayesian approach resulted a success rate
of about 70 % in a case where the success rate of a conventional height-based filtering (HBF) method
was approximately 50 %.
In the CHM fitting proposed in (Lähivaara et al., 2014), the tree crowns are modeled as rotationally
symmetric objects. If the error caused by this idealization is not accounted for, it can result biased
estimates for the tree crown parameters and also lead to false detection of trees. In this paper, we
aim at improving the Bayesian single tree detection by statistical modeling of errors caused the
rotationally symmetric CHMs. The statistics of these modeling errors are computed using a training
data set consisting of ALS data and field measured tree locations and sizes. The results show that
accounting for the tree-level modeling error statistics improves the Bayesian estimates for the single
tree parameters and the success rate of the tree detection.
References
Andersen, H., Reutebuch, S. E., & Schreuder, G. F.
(2002). Bayesian object recognition for the analysis of complex forest scenes in airborne laser
scanner data. International Archives of Photogrammetry Remote Sensing and Spatial Information Sciences, 34(3/A), 35-41.
Kaipio, J., & Somersalo, E. (2005). Statistical and
computational inverse problems (1st ed.) Springer-Verlag New York.
176
Lähivaara, T., Seppänen, A., Kaipio, J. P., Vauhkonen,
J., Korhonen, L., Tokola, T., et al. (2014). Bayesian
approach to tree detection based on airborne laser scanning data. IEEE Transactions on Geoscience and Remote Sensing, 52(5), 2690-2699.
Micheas, A. C., Wikle, C. K., & Larsen, D. R. (2014).
Random set modelling of three-dimensional objects in a hierarchical bayesian context. Journal
of Statistical Computation and Simulation, 84(1),
107-123.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Modelling residual stand volume using unmanned aerial
vehicles and digital aerial photogrammetry
Tristan Goodbody1*, Nicholas C. Coops1, Piotr Tompalski1, Patrick Crawford2, Ken Day3
Faculty of Forestry, University of British Columbia, 2424 Main Mall, Vancouver, BC, V6T 1Z4, Canada. (604) 822 6452,
goodbody.t@gmail.com, nicholas.coops@ubc.ca, piotr.tompalski@gmail.com
2
Spire Aerobotics, 211-2386 East Mall Vancouver, BC, V6T 1Z3, Canada. (778) 928-4203, patrick@spireaerobotics.com
3
Alex Fraser Research Forest, 72 South 7th Avenue, Williams Lake, BC, V2G 4N5,
Canada. (250) 392-2207, ken.day@ubc.ca
Preference for Oral Presentation Format
1
Keywords: Unmanned Aerial Vehicles, Forest Inventory, Stereo Photogrammetry, Timber Volume,
Predictive Modelling
Abstract
To improve precision management and the cost effectiveness of forest planning and operations, we
look to build an enhanced forest inventory. To do so, we utilized an unmanned aerial vehicle (UAV) to
collect a digital aerial photogrammetry (DAP) pointcloud to model residual stand volume and identify
harvest locations. To do so, UAV collected aerial imagery and field measurements were acquired in
2015 in The Alex Fraser Research Forest near Williams Lake, BC. Tree height, diameter at breast height
(DBH), and species were measured from systematically implemented variable radius plots throughout
the study area at the time of imagery collection. DAP pointcloud metrics and field measurements
were used to create a DAP volume model using linear regression with an RMSE% and bias% of 18.50
and 1.16 respectively. Results achieved from the DAP volume model indicate strong potential for
UAV acquired DAP data to outline the location and estimate the quantity of residual stand volume.
The cost effectiveness, ease of deployment, and collection of very high resolution imagery makes
UAV collected DAP a strong candidate for incorporation into forest management and enhanced
forest inventory practices. Future analysis into the utilization of DAP to gauge the success of applied
harvesting regimes and monitoring residual stand behaviour post-harvest are discussed.
177
Universidad Mayor
Forest Modelling
Modelling the effect of environmental factors on the height
increment of stands with the use of repeated Airborne
Scanning data
Jarosław Socha*1, Radomir Bałazy2, Marcin Pierzchalski 1, Mariusz Ciesielski2
1
University of Agriculture in Krakow, Faculty of Forestry, Al. 29 Listopada, Kraków, 32-425 Poland
2
Forest Research Institute, Braci Leśnej 3, Sękocin Stary, 05 – 090 Poland
*
Speaker, e-mail: rlsocha@cyf-kr.edu.pl
Keywords: LiDAR, height increment, microsite variability, topography
Abstract
Dynamics of forest communities expressed by height increment may be subjected to environmental
gradients. Tree height increment may be subjected to temperature and wetness gradients, provides
a quantitative baseline for understanding patterns of resource use, spatial structure and the
physiological mechanism of tree reaction ecosystems to climate stress as the factors predisposing to
climate affected disturbances.
The dependency between the topography and various environment factors has been observed for
a long time (Hess, 1965, Maciaszek et al., 2000). The most commonly used topographic indices are
elevation, aspect, slope, slope position, longitude and latitude (Chen et al., 1998; Hägglund and
Lundmark, 1977; Seynave et al., 2005; Socha, 2008). Increasing availability of precise ALS based
DEM models enable more detailed description of local topography by the use of different indices
such as: TPI, SPTPI, TWI (R. Sørensen et al., 2006) or geomorphons (Jasiewicz and Stepinski, 2013). In
site evaluation latitude and elevation are used as indirect measures of regional climate, while slope
position, slope gradient and aspect are used as measures of local climate (Chen et al., 2002; Monserud
and Rehfeldt, 1990; Socha, 2008; Wang and Klinka, 1996). Topographic indices such as Topographic
Wetness Index (TWI) may be used as indirect measures of water status of given site (R. Sørensen et
al., 2006). Due to the difficulties connected with measurement of height increment on standing trees
analyses of height increment and its response to environmental gradients have received so far less
attention and were limited mainly to relatively small samples collected from cut down trees (Gamache
and Payette, 2004).
This study constitutes the practical application of change detection by airborne laser scanners for
wall-to-wall monitoring the effect of environmental factors on the tree height growth over time using
repeated ALS data.
The research objectives of this study on the whole elevational and aspect gradients in the Jizera
Mountains were to (1) detect the site-dependent variability of mean annual height increments
measured using change detection techniques, (2) explore the dominant environmental factors limiting
height increment using repeated ALS data. We hypothesized that the limiting effect of water stress
and temperature on height growth is affected by topography. Therefore we revealed that elevation,
aspect and topographic indices, which by influence on thermal and moisture conditions affect height
increment of Norway spruce may be used as indicator of susceptibility of given site to disturbances.
178
Universidad Mayor
ForestSAT 2016 Abstracts Summary
In the study, the data used were acquired: (a) in August and September 2007 via ALTM 3100 Optech
laser scanner. (b) in July and August 2012 via LMS-Q680i laser scanner. ALS data had average density
of around 6 points per m2, height error up to 15 cm and an X,Y location error below 20 cm. With the
interpolation algorithms implemented in TerraScan software based on point cloud, the following
models were generated: Digital Terrain Model (DTM), Digital Surface Model (DSM), and normalized
Digital Terrain Model (nDSM) with a spatial resolution of 0.5 m. Digital Terrain Model (DTM) developed
based on data from the airborne laser scanning was of spatial resolution equal to 0.5 m and covered
whole area. Based on ALS data, digital surface models were generated for the years 2007 and 2012.
Then, with through the raster difference (DSM – DTM) in map algebra Crowns Height Models (CHM)
were established.
In the analysis the relationships between height increment obtained by change detection techniques
and predictor variables were investigated using generalized additive models (GAM, Hastie &
Tibshirani1990). Optimal amount of smoothing was estimated based on cross-validation. Graphs
of smoothing spline functions of GAM models was used in order to illustrate the effect of individual
predictor variables on height increment. Model GAM describing height increment within the period
2007 - 2012 as a function of site and stand characteristics explain about 62,5 % of height increment
variability (R2adj=0,625). It was found that height increment was significantly affected by site
characteristics including: altitude, slope, transformed aspect and other DTM products such as TWI and
TPI.
179
Forest Modelling
Universidad Mayor
Nationwide airborne laser scanning based models for volume,
biomass and dominant height in Finland
Eetu Kotivuori, Lauri Korhonen & Petteri Packalen*
eetu.kotivuori@uef.fi, lauri.korhonen@uef.fi, petteri.packalen@uef.fi, School of Forest Sciences, Faculty of Science and
Forestry, University of Eastern Finland, P.O. Box 111, 80101 Joensuu, Finland. Petteri Packalen is the presenter.
Keywords: Airborne Laser Scanning; LIDAR; area-based approach; remote sensing; regression analysis;
calibration; mixed-effect models
The aim of this study was to examine how well
stem volume, above-ground biomass and dominant
height can be predicted using nationwide airborne
laser scanning (ALS) based regression models. The
study material consisted of nine ALS inventory
projects in different parts of Finland. We used field
sample plots and ALS data to create nationwide and
regional models for each response variable. The final
models contained one or two ALS predictors, which
were chosen based on the root mean square error
(RMSE), and cross-validated. Finally, we tested how
much predictions would improve if the nationwide
models were calibrated with a small number of local
sample plots. Although forest structures differ in
different parts of Finland, the nationwide volume
and biomass models performed quite well (leave-
180
inventory-area-out RMSE 22.3% to 33.8%, bias
-13.8% to 18.7%) compared with regional models
(leave-plot-out RMSE 20.2% to 26.8%, no bias).
However, the nationwide dominant height model
(RMSE 5.4% to 7.7%, bias -2.0% to 2.8%, with
the exception of one region) performed nearly as
well as the regional models (RMSE 5.2% to 6.7%).
The results show that the nationwide volume and
biomass model predictions are likely to be biased,
because forest structure and ALS device have a
considerable effect on the predictions. Large biases
appeared especially in northern Finland. Local
calibration decreased the bias and RMSE of volume
and biomass models. However, the nationwide
dominant height model did not benefit much from
calibration.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Predicting the aboveground biomass of individual trees using
remote sensing data and new allometric models: a case study
in Norway
Michele Dalponte1, Lorenzo Frizzera1, Hans Ole Ørka2, Tommaso Jucker3, Terje Gobakken2, Erik Næsset2, Damiano Gianelle1
1
Dept. of Sustainable Agro-ecosystems and Bioresources, Research and Innovation Centre, Fondazione E. Mach, Via E.
Mach 1, 38010 San Michele all’Adige (TN), Italy
2
Dept. of Ecology and Natural Resource Management, Norwegian University of Life Sciences, P.O. Box 5003,
NO-1432 Ås, Norway
3
Dept. of Plant Sciences, University of Cambridge, Downing Site, Cambridge CB2 3EA, United Kingdom
Keywords: individual tree crowns, aboveground biomass, airborne laser scanning, hyperspectral imagery
Introduction
Allometric models used in forestry to predict tree’s
aboveground biomass rely primarily on diameter
at breast height (DBH) and tree height as inputs. In
many forest types such models are species-specific,
which means that species must be recorded as well.
In particular DBH and species information play an
important role as tree height is often predicted as
function of these attributes. Nowadays airborne
laser scanning (ALS) data provide detailed height
measurements, in addition to which individual
tree crown (ITC) delineation methods provide a
means to determine (for each detected tree) crown
dimensions that subsequently can be related to DBH
(Hemery et al., 2005). Moreover, with hyperspectral
data detailed tree species mapping is possible
(Dalponte et al. 2012). As DBH cannot be directly
measured with airborne remote sensing (ARS) data,
it may be useful to develop new allometric models
based on tree characteristics that can be directly
measured from ARS data. Following this idea, Jucker
et al. (2016) recently developed allometric models
for predicting DBH and aboveground biomass
(AGB) based on height and crown diameter using a
worldwide dataset of more than a hundred thousand
individual trees measurements. The study of Jucker
et al. (2016) did not involve remote sensing data,
focusing exclusively on the development of the
allometries. Moreover it focused on developing
global and regional allometric models without
considering species-specific information. The
objective of this study is to develop species-specific
models for the prediction of DBH for three main
species groups in Norway using field measured height
and crown diameters, and to validate the models
when remotely sensed data (ALS and hyperspectral
data) are used to predict species, height and crown
diameter. To achieve this objective we developed
the models using data from one dataset and then we
validated them on an independent dataset.
Study areas
In this study two datasets collected in two different areas in southern Norway were used. The
first dataset collected in Aurskog-Høland municipality was used for the development of the
DBH allometric models for Scots pine, Norway
spruce and broadleaves species. The dataset
consists of 667 field measured trees for which
DBH, height, species and crown dimensions
were recorded. On this dataset a validation at
tree level was done.
The second dataset was used to validate the
models at plot level. It was collected in Våler
municipality. This dataset comprised 9376 trees
distributed on 153 plots of 400 m2. ALS and hyperspectral data have been acquired over the
plots in 2011. Inside each plot the position of all
the trees, the DBH and the species were recorded, while the height was measured only for few
sample trees in each plot.
181
Universidad Mayor
Forest Modelling
Methods
Using the Aurskog-Høland dataset, models that
relate DBH to field measured heights and crown
diameter were constructed following the methodology of Jucker et al. (2016). The models were
in the form:
(1)
where is the diameter at breast height in centimeters, the height in meters, and the crown diameter in meters. The model fitting was done by
using a data binning approach and a log normal
form of the models. To validate the models at
tree level, for each model, the root mean square
error (RMSE), average mean difference (D-bar),
and coefficient of determination (R2) were calculated on an independent validation dataset
corresponding to 10% of trees selected at random. This randomization procedure was repeated 100 times and for each metric the mean value
across the 100 iterations was calculated.
For the Våler dataset, individual tree crowns
(ITCs) were delineated using the ALS data and
the delineation algorithm of the R package
(itcSegment). Delineated ITCs were matched to
the field data and the matched ITCs were used
for species classification using hyperspectral
data. 1608 matched ITCs were used as training
and 1416 as validation. The hyperspectral pixels inside each ITC that had a NDVI value higher
than 0.5 were used in the classification process.
The species classification was performed using
a Support Vector Machine (SVM) classifier (R library kernlab) at pixel level, and then the labels
of the pixels inside each ITC were aggregated
using a majority rule. This approach was used to
predict the species of all the delineated ITCs.
DBH for each ITC was predicted using the model in equation 1. AGB was predicted for each
ITC using the models of Marklund (1988) with
height from ALS data as one predictor and DBH
from equation 1 and species from hyperspectral data as the two other predictors. AGB for
the field trees was predicted using the models
of Marklund (1988) and field measured height,
DBH, and species. Plot level field estimated AGB
was obtained summing the values of AGB of all
trees inside each plot. Similarly the plot level
ARS estimated AGB was obtained summing the
values of AGB of all ITCs inside each plot.
Results
In Table 1 the coefficients of the models in equations
1 for the three species considered are presented,
along with RMSE, D-bar and R2. The delineation
method used detected more than 90% of the
trees with DBH >20 cm, and only 56% of the trees
with DBH <20 cm (see Figure 1). The tree species
classification accuracy obtained on the validation set
were: i) overall accuracy: 87.7%; ii) kappa accuracy:
0.794; and ii) mean class accuracy: 83.6%. Figure 2
shows the results of the AGB estimation at plot level
for the Våler dataset.
As it is clear from these results (and by the ones
of Jucker et al. 2016) it is possible to reverse the
allometric models in order to use height and crown
diameters as input. The relation of these models
with ARS data needs to be considered more carefully
as it strongly depends on the forest characteristics.
Indeed the main limitation remains, namely that ITC
approaches do not detect all trees, and especially
not the suppressed ones, and thus there will always
be a fraction of the AGB missing (see Figure 2). The
tendency of missing trees will increase in multilayer
and dense forests. Moreover, dense canopies may
create high commission errors in the delineation,
resulting in an erroneous increase in the AGB estimate.
Table 1. Coefficients, RMSE, D-bar and R2 for equation 1 for all the species together,
and for the three main species groups.
. % 1
((
.03
%*
0+( 4 1
182
I?QIQ
I?RNK
J?LON
I?PRN
I?PRM
I?PMR
I?ORJ
I?PNM
D)E
L?O
K?O
L?O
K?L
B0DUE
BL?K
I?R
K?I
I?L
K
I?PR
I?QP
I?QN
I?NI
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Figure 1. Number of field measured trees (dark gray bars) and number of detected ITCs (light gray) for each
DBH class. The numbers at the top of the graph represent the percentage of detected trees.
Figure 2. Field estimated AGB and ARS estimated
AGB for each field plot in Våler dataset.
References
Dalponte, M., & Coomes, D. A. (2016). Tree-centric
mapping of forest carbon density from airborne
laser scanning and hyperspectral data. Methods
in Ecology and Evolution, doi: 10.1111/2041210X.12575.
Dalponte. M., Bruzzone, L., Gianelle, D. (2012) Tree
species classification in the Southern Alps based
on the fusion of very high geometrical resolution
multispectral/hyperspectral images and LiDAR
data. Remote Sensing of Environment, 123, 258–
270.
Hemery G.E., Savill P.S., & Pryor S.N. (2005)
Applications of the crown diameter–stem
diameter relationship for different species
of broadleaved trees. Forest Ecology and
Management, 215, 285–294.
Jucker, T., Caspersen, J. P., Chave, J., Antin, C.,
Barbier, N., Bongers, F., Dalponte, M., van Ewijk,
K., Forrester, D., Heani, M., Higgins, S., Holdaway,
R., Iida, Y., Lorimer, C., Marshall, P., Momo, S.,
Moncrieff, G., Ploton, P., Poorter, L., Rahman,
A., Schlund, M., Sonké, B., Sterck, F., Trugman,
A., Usoltsev, V., Vanderwel, M., Waldner, P.,
Wedeux, B., Wirth, C., Wöll, H., Woods, M.,
Xiang, W., Zimmermann, N., & Coomes, D. A.
(2016). Allometric equations for integrating
remote sensing imagery into forest monitoring
programs, Global Change Biology, submitted.
Marklund, L.G., (1988). Biomass functions for pine,
spruce and birch in Sweden (Report No. 45).
Swedish University of Agricultural Sciences,
Umeå.
183
Universidad Mayor
Forest Modelling
Prediction of alien species richness in two forest watershed of
South-Central Chile: a remote sensing synergistic approach
Mauricio Galleguillos1,2*, Andrés Ceballos1,2, Antonio Lara2,3
Universidad de Chile, Facultad de Ciencias Agronómicas, Departamento de Ciencias Ambientales y Recursos Naturales
Renovables.
2
Center for Climate and Resilience Research (CR2), Chile.
3
Universidad Austral de Chile, Facultad de Ciencias Forestales y Recursos Naturales, Instituto de Conservación,
Biodiversidad y Territorio.
*Corresponding author: mgalleguillos@renare.uchile.cl
1
Keywords: alien species richness, predictive model, spectral index, evaporative fraction,
spatio-temporal, textures, exotic forest plantations.
Abstract:
Central Chile is considered one of the biodiversity hotspot and is a highly threatened territory due to
land use changes, especially the establishment of exotic industrial plantations of pine and eucalyptus.
The intensive short time rotation of these monoculture provide increasingly fragmented landscape
configuration that favors the introduction and establishment of alien species. The aim of the research
was estimating the spatial distribution of alien species richness by the use of a synergistic approach
based on spatio-temporal remote sensing information. We developed a parsimonious and accurate
predictive model of alien species richness from in situ data taken in two watershed of South-central
Chilean coastal range, based on vegetation indexes and evaporative fraction derived from thermal
data.
Introduction
Ecosystems of the Chilean Mediterranean
biodiversity hotspot are under an increasingly
threatened condition mainly due to land use change
because of human use competition (Alaniz et al. 2016).
The landscapes of South-Central Chile are currently
configure by the dynamic of forest plantation
systems, and are constitute by a fragmented mosaic
of pine and eucalyptus monocultures, shrubs, bare
soil, agro-grassland and native forest remnants
with high endemism proportion (Miranda et al.,
2016). The high rates of change of land cover due
to the dynamic of short time rotation of industrial
plantations have generated modifications in the
floristic composition of plant species, leading to an
increase of alien species adapted to conditions of
constant change and variability.
The invasion of dangerous alien species can lead to
extinction of native species, modifying biodiversity,
affecting ecosystems services, modifying the
184
ecosystem metabolism and affecting the landscape
(Hobbs 2000). The monitoring of alien species is
paramount to prevent this effects, this is why remote
sensing has demonstrated to be a valuable tool in this
endeavor (Asner and Vitousek 2005). The generation
of especially explicit empirical models that allow to
identify where those species are grouped, can help to
improve the management decision at the landscape
level. The investigation present an original predictive
modelling methodology based on a synergistic
remote sensing approach that make use of spatiotemporal predictors based on vegetation index and
evaporative fraction information.
The study area comprise to subcatchments (Purapel
y Cauquenes) covered mainly by forest plantations
and are located in the coastal range mountains of
the Maule and Bio-bio regions (Fig. 1). Land cover
maps were computed with the Maximum Likelihood
supervised classification from Landsat data for
years 2001 and 2013 in order to determine the field
sample plots that represent the most representative
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Fig. 1: Study site and sample plots spatial distribution
ecosystem states within the catchments: Early
successional; adult pine plantation; adult eucalyptus
plantation; mixed-exotic-native; native forest (Fig
2). Field vegetation surveys were accomplished to
determine plant vascular species richness, including
herbs, shrubs and trees, in 39 plots of 225m2 each, by
using a nested plot design.
spatial heterogeneity of the landscape. A data mining
approach based on the Recursive Feature Elimination
(Guyon et al. 2002) algorithm was used to obtain
the combination of predictors that explain better
the response variable. Then, the selected predictors
were used to adjust a Generalized Linear Model with
a Poisson distribution (Lopatin et al., 2016).
Several kind of predictors were considered:
topographical, derived from a DEM SRTM; vigor of
vegetation derived from spectral indexes (NDVI,
GNDVI, NGRDI); and hydric status of vegetation
represented by evaporative fraction (EF) computed
with thermal data using the Simplified Surface
Energy Balance Index (S-SEBI) (Roerink et al. 2000)
following the parameterization of Galleguillos et al.
2011. Spectral and thermal predictors were obtained
from ASTER Level 2 products of surface reflectance
and land surface temperature, spatio-temporal
metrics were computed for the latter predictors,
and these were obtained from nine and eight scenes
spanning the years 2001 to 2014 for Cauquenes and
Purapel watersheds. Variables eco-hydrological
conditions were covered by the set of available
scenes according to the time lapse from the last
significant precipitation event. The predictors were
submitted to texture analysis with a gray-level cooccurrence matrix, which allowed to characterize the
A parsimonious model was obtained for alien
richness prediction (Fig. 3), with an R2 = 0.74, an
RMSE = 2.18 and a BIAS = 11.1% considering 4
predictors: the variance between 25 pixels (kernel
of 5*5) from a raster that represent the first quartile
of the temporal series of NGRDI (spectral-spatialtemporal predictor, NGRDIpc_Vx5), the quotient
of the absolute deviation from the median and the
median of temporal series of evaporative fraction
(temporal-energy flux predictor, EF_cb), the land
cover (structural predictor), and the watershed
where the alien species belongs (bioclimatic
predictor). Representation of the model: glm
(Richness ~ NGRDIpc_Vx5 + EF_cb + LandCover +
watershed). A bootstrapping validation assessment
shows the consistency of the model in the Fig. 3
a,b,c). The selected predictors were coherent with
the functional dynamic of alien species in those
kinds of ecosystems, and with the inner structural
compatibility of each land cover class.
185
Forest Modelling
Universidad Mayor
Figure 2. Ecosystems states identifies according to the land cover changes analysis of maps of 2001 and 2013.
Arrows represent expected transitions between classes and the width of the arrows represent the frequency
of the transition within the catchments.
Results and discussion
Fig. 3: a,b,c). Coefficient of determination, root-mean-square error and bias bootstraping validation
assessment respectively of the predictive model. d) Model representation (predicted alien species richness
by the model vs observed alien species richness).
Conclusions
The synergy generated by the contribution of
different source of satellite data and the spatio-
186
temporal analysis allowed to obtain an acceptable
accurate prediction of alien species richness in
forested ecosystems subjected to strong anthropic
pressure.
Universidad Mayor
References
Alaniz A., Galleguillos M., Pérez-Quezada J. 2016.
Assessment of quality of input data used to
classify ecosystems according to the IUCN Red
List methodology: The case of the central Chile
hotspot. Biological Conservation In press.
Asner G. and Vitousek P.M. 2006. Remote analysis of
biological invasion and biogeochemical change.
PNAS 102(12): 4383-4386.
Galleguillos M., Jacob F., Prévot L., French A.,
Lagacherie P., (2011). Comparison of two
temperature differencing methods to estimate
daily evapotranspiration over a Mediterranean
vineyard watershed from ASTER data. Remote
Sensing of Environment 115(6): 1326.
Guyon, I., Weston, J., Barnhill, S., (2002). Gene
selection for cancer classification using support
vector machines. Mach. Learn. 46, 389–422.
ForestSAT 2016 Abstracts Summary
Hobbs R.J. (2000) Land-Use changes and invasions.
In: Mooney H.A., Hobbs R.J. (eds.) Invasive
species in a changing world. Washington, DC.
Island Press. p. 31-54.
Lopatin J., Dolos K., Hernández J., Galleguillos M.,
Fassnacht F. (2016). Comparing generalized linear
models and random forest to model vascular
plant species richness using LiDAR data in a
natural forest in central Chile. Remote Sensing of
Environment 173:200–210.
Miranda, A. Altamirano, A. Cayuela, L. Lara, A.
González, M. 2016. Native forest loss in Chilean
biodiversity hotspot: revealing the evidence.
Regional Environmental Change: 1-13.
Roerink, G. J., Su, Z., & Menenti, M. (2000). S-SEBI: a
simple remote sensing algorithm to estimate the
surface energy balance. Physics and Chemistry
of the Earth. Part B: Hydrology, Oceans and
Atmosphere, 25(2), 147−157.
187
Universidad Mayor
Forest Modelling
Reconciling MODIS satellite with terrestrial forest inventory
data to assess forest productivity in Europe
Mathias Neumann1*, Adam Moreno1, Volker Mues2, Sanna Härkönen3, Matteo Mura4, Olivier Bouriaud5, Mait Lang6,
Giuseppe Cardellini7, Alain Thivolle-Cazat8, Karol Bronisz9, Jan Merganic10, Iciar Alberdi11, Rasmus Astrup12, Frits Mohren13,
Maosheng Zhao14, Hubert Hasenauer1
1
Institute of Silviculture, Department of Forest and Soil Sciences, University of Natural Resources and Life Sciences,
Vienna, 1190, Austria
2
University of Hamburg, Centre for Wood Science, World Forestry, Hamburg, 21031, Germany
3
University of Helsinki, Department of Forest Sciences, Helsinki, 00014, Finland
4
geoLAB - Laboratory of Forest Geomatics, Department of Agricultural, Food and Forestry Systems,
Università degli Studi di Firenze, Firenze, 50145, Italy
5
Universitatea Stefan del Mare, Suceava, 720229, Romania
6
Tartu Observatory, Tõravere, 61602, Estonia
7
KU Leuven – University of Leuven, Division Forest, Nature and Landscape, Leuven, 3001, Belgium
8
Technological Institute, Furniture, Environment, Economy, Primary processing and supply,
Champs sur Marne, 77420, France
9
Laboratory of Dendrometry and Forest Productivity, Faculty of Forestry, Warsaw University of Life Sciences,
Warsaw, 02-776, Poland
10
Czech University of Life Sciences, Faculty of Forestry and Wood Sciences, Prague, 16521, Czech Republic
11
INIA-CIFOR, Departamento de Selvicultura y Gestión de los Sistemas Forestales, Madrid, 28040, Spain
12
Norwegian Institute for Bioeconomy Research, Ås, 1431, Norway
13
Wageningen University, Forest Ecology and Forest Management Group, Wageningen, 6700, The Netherlands
14
Department of Geographical Sciences, University of Maryland, 20742, USA
* corresponding author mathias.neumann@boku.ac.at
Keywords: MODIS, NFI, Europe, trees, climate, modelling, carbon, biomass, bioeconomy
Forests provide important ecosystem services, such
as timber production for a growing bio-economy,
food, increasing demand biodiversity, water and
protection against natural hazards. Forests are
also increasingly important in mitigating climate
change effects by storing large amounts of carbon.
The physiology and productivity of forests are
strongly altered by environmental conditions
such as climate, soils and management practices.
Understanding the response and feedbacks of forest
towards historic, current and future environmental
change requires a consistent large scale model
framework incorporating the biogeochemical
processes between vegetation and the atmosphere.
Today large scale forest information is provided by
earth observing satellite systems which include the
Moderate Resolution Imaging Spectroradiometer
(MODIS). MODIS data can be used in conjunction
with the MOD17 algorithm to produce vegetative
productivity estimates. This algorithm combines
biogeochemical model principles with temporally-explicit vegetation data and daily climate data
188
to provide productivity measures worldwide on
a 1-km resolution. In this study we computed a
regional Net Primary Production (NPP) dataset
(“MODIS_EURO”) by rerunning MOD17 with local
European climate data. We next harmonized tree
carbon estimation methods from 13 European
countries and processed data from 196,434 plots
(more than 2 Mio. sample trees) from 13 National
Forest inventories to obtain terrestrial NPP. With
this reference dataset, we evaluated our previously
created remote sensing dataset “MODIS_EURO”
across scales and gradients. MODIS EURO shows
better agreement with the terrestrial reference NPP
from forest inventory data, than the global MODIS
NPP estimates that use a global climate data set,
highlighting the importance of accurate climate data
in remote sensing applications. MOD17 is climate
sensitive and local climate data better captures the
conditions in Europe. Discrepancies between MODIS
EURO and terrestrial data can be explained by
differences in forest stand density and the employed
tree carbon estimation methods. This new dataset
Universidad Mayor
- the key outcome of this study - provides wall-towall NPP estimates for Europe (EU-27 including
Norway and Switzerland) and allows spatial and
temporal analysis on the impact that environmental
change has on vegetative productivity. This study
highlights the advantages and disadvantages of two
conceptually different approaches for estimating
NPP and enhances our understanding of linking
terrestrial with remote sensing data.
References:
Moreno, A. & Hasenauer, H. (2015). Spatial
downscaling of European climate data.
International Journal of Climatology. http://doi.
org/10.1002/joc.4436
Moreno, A., Neumann, M., & Hasenauer, H. (2016).
Optimal resolution for linking remotely sensed
and forest inventory data in Europe. Remote
Sensing of Environment, 183, 109–119. http://
doi.org/10.1016/j.rse.2016.05.021
Neumann, M., Zhao, M., Kindermann, G., &
Hasenauer, H. (2015). Comparing MODIS Net
Primary Production Estimates with Terrestrial
ForestSAT 2016 Abstracts Summary
National Forest Inventory Data in Austria.
Remote Sensing, 7(4), 3878–3906. http://doi.
org/10.3390/rs70403878
Neumann, M., Moreno, A., Mues, V., Härkönen, S.,
Mura, M., Bouriaud, O., Lang, M., Achten, W.
M. J., Thivolle-Cazat, A., Bronisz, K., Merganic,
J., Decuyper, M., Alberdi, I., Astrup, R., Mohren,
F.,Hasenauer, H. (2016). Comparison of carbon
estimation methods for European forests. Forest
Ecology and Management, 361, 397–420. http://
doi.org/10.1016/j.foreco.2015.11.016
Neumann, M., Moreno, A., Thurnher, C., Mues, V.,
Härkönen, S., Mura, M., Bouriaud, O., Lang, M.,
Cardellini, G., Thivolle-Cazat, A., Bronisz, K.,
Merganic, J., Alberdi, I., Astrup, R., Mohren, F.,
Zhao, M., & Hasenauer, H. (2016). Creating a
Regional MODIS Satellite-Driven Net Primary
Production Dataset for European Forests.
Remote Sensing, 8(554), 1–18. http://doi.
org/10.3390/rs8070554
Zhao, M. & Running, S. W. (2010). Drought-induced
reduction in global terrestrial net primary
production from 2000 through 2009. Science,
329(5994),
940–3.
http://doi.org/10.1126/
science.1192666
189
Universidad Mayor
Forest Modelling
Relating forest height structure to virtual ground truth data
Nikolai Knapp1, Rico Fischer1, Andreas Huth1
1
Helmholtz-Centre for Environmental Research (UFZ), Leipzig, Germany
Key words: forest structure, dynamic modeling, Lidar simulation, calibration
Recent developments in remote sensing technology
have improved our abilities of forest monitoring
remarkably. Accurate height maps and even 3D
point clouds of forest canopies can be obtained from
light detection and ranging (Lidar) and increasingly
from synthetic aperture Radar (SAR) satellites
(e.g. TanDEM-X, Sentinel-1). However, it remains a
challenge to link metrics of height structure to the
main variables of interest, such as forest biomass,
productivity, carbon turnover, stem size distribution,
disturbance patterns or forest age. Usually, large
forest inventory datasets are required to establish
robust relationships between remote sensing
metrics and ground-based metrics. Such inventorybased calibration requires extensive field campaigns
and reaches its limits with variables which are
difficult to measure at plot scale, like carbon fluxes.
Here, we propose a new approach where we
complement real world ground-truth data with
virtual forest inventory data generated by forest
simulations. Dynamic forest models help us to
understand the links between ecological processes
and vegetation structure. Hence, joining the fields
of remote sensing and forest modeling will lead to
mutual benefits, by increasing structural realism of
190
forest models and facilitating the development and
calibration of remote sensing approaches.
We used the individual-based forest model
FORMIND to simulate the neotropical lowland
rainforest of Barro Colorado Island (BCI), Panama,
and derived virtual inventory data covering the
full range of possible successional states. For each
simulated stand we also conducted a virtual remote
sensing campaign using a Beer-Lambert-based Lidar
model. The simulated datasets were then used to
investigate relationships between remote sensing
metrics and intrinsic attributes of the forest. We
considered metrics of different spatial resolutions to
also mimic height data that could be obtained from
spaceborne SAR systems.
The presented methodology can serve to better
understand how carbon dynamics in forests relate
to the measurable structure of forests and how
this information can be used for remote sensing
based inventories and monitoring. Additionally, the
approach allows exploration of the potential of future
satellite missions (e.g. GEDI, BIOMASS, Tandem-L)
for estimating the different forest attributes.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Temporal and Angular Effects of the Spectral Signal on
Deciduous Forest Crown Components
Michael Foerster1, Ben Somers3, Kyle Pipkins1, Laurent Tits4, Karl Segl2, Maximilian Brell2, Birgit Kleinschmit1,
Angela Lausch5 and Anne Clasen1, 2
Geoinformation in Environmental Planning Lab, Technische Universität Berlin, Straße des 17. Juni 145,
D-10623 Berlin, Germany
2
Helmholz Centre Potsdam-GFZ German Research Centre for Geosciences, Telegrafenberg, 14473 Potsdam, Germany
3
Division of Forest, Nature and Landscape, Department of Earth and Environmental Sciences, KU Leuven,
Celestijnenlaan 200E, BE-3001 Leuven, Belgium
4
Geomatics Lab, Department of Biosystems, KU Leuven, Willem de Croylaan 34, BE-3001 Leuven, Belgium
5
Department of Compuational Landscape Ecology, Helmholtz Centre for Environmental Research-UFZ,
Permoserstrasse 15, D-04318 Leipzig, Germany
1
Keywords: Unmixing; MESMA; forest; bark; observation angle; phenology.
Abstract
Imaging spectroscopy has proven to be a promising tool to provide information on forest biochemical
and biophysical variables. The presented study is investigating crown component fractions in a forest
ecosystem on a sub-pixel-basis. In this context, the potential of spectral unmixing to derive information
on the distribution of the endmembers ‘leaf’, ‘bark’ and ‘soil’ within a pixel is analyzed.
Close range spectral measurements of the canopy and its components were taken with an ASD
FieldSpec instrument mounted to a crane measurement platform situated in a temperate deciduous
forest in North-East Germany. Reference fractional abundances of the components were precisely
determined from photographs taken simultaneously to the ASD measurements. Unlike most other
studies, which only consider two components, mainly ‘leaf’ and ‘soil’, this experimental setup facilitated
the inclusion of the additional component ‘bark’ for the unmixing of forest crown components.
Measurements from different stages of the phenological phases were used in this study. In the 2015
field campaign, data was collected on April 21st, June 5th, August 3rd and October 1st. Airborne HySpex
data was collected simultaneously on the last three of these dates and on July 7th.
Especially in summer, the dense foliation in this complex vegetation structure causes the problem of
saturation effects, when applying broadband vegetation indices. This study illustrates that multiple
endmember spectral mixture analysis (MESMA) can contribute to overcoming this challenge. The
results indicate that the inclusion of an additional ‘bark’ endmember clearly improves the accuracy of
the unmixing results. When only nadir viewing directions were taken into account, a mean absolute
error of 7.9% could be achieved for the fractional occurrence of the ‘leaf’ endmember and 5.9% for
the ‘bark’ endmember. In order to evaluate the results of this field-based study for remote sensing
applications, a transfer to airborne and satellite-based imagery was carried out. All sensors used in this
study were capable of unmixing crown components with a mean absolute error ranging between 3%
and 21% for the nadir measurements.
However, airborne and spaceborne sensors provide imagery that includes not only nadir viewing
directions. Off-nadir viewing zenith angles up to 10° are a common case for sensors with a large swath.
This needs to be considered in image processing and the derivation of biophysical parameters. BRDF
corrections account for illumination and viewing geometry as well as structural and optical properties
of the surface, but commonly, they do not account for variations in the actual visibility of the canopy
components at different observation angles. The crane measurement platform allowed for a very
detailed view on the changing visibility of the occurring canopy components at different angles.
191
Universidad Mayor
Forest Modelling
Use of Random Forest Modeling Techniques
to Predict and Detect
Shrub Locations Under Canopy using LiDAR Structure and
Topography Metric
Caileigh Shoot, Sean Jeronimo, Van Kane, Monika Moskal, Jonathan Kane
University of Washington, School of Environmental and Forest Sciences, Seattle, WA 98195;
Jim Lutz
Utah State University, S. J. & Jessie E. Quinney College of Natural Resources, 5230 Old
Main Hill, Logan, UT 84322-5230
Abstract
Shrubs are an important component of forested ecosystems around the globe. Fulfilling many
ecological niches, shrubs can be anything from a source of food and habitat, to an understory fuel in
wildfire. Shrubs add vital carbon and nutrients to forests, assist in water filtration and storage in forest
soils, and can act as a key indicator of overall forest vigor. Despite their importance, shrub distributions
are widely unknown across landscapes. In addition, mapping shrub locations in the field is arduous
and virtually impossible to implement across entire landscapes. Thus, in order to protect and manage
shrubs, it is important that a method for mapping and quantifying shrubs across the landscape is
developed. LiDAR has been proven to accurately measure upper canopy features when compared
with field measurements, but has not been proven to be a viable method of measuring below-canopy
features such as shrubs, despite it’s ability to penetrate the forest canopy. This study will create and
evaluate a method for detection and prediction of shrub locations across a landscape using Random
Forest, a classification and regression machine learning algorithm.
This study was performed on the 25.6 ha Yosemite Forest Dynamics Plot in Yosemite National Park.
Medium-resolution (averaging 30 points per m^2) discrete-return LiDAR data was acquired in 2010,
and in 2011 field crews mapped and identified the species of all shrub patches >2 m^2. The LiDAR
data was analyzed using FUSION, which generates 400+ metrics from the LiDAR data, including both
canopy structure and topographic metrics. These metrics were input as predictors into a Random
Forest model, and then used to detect and predict shrub locations. With this method, explanatory
model accuracies averaged 84% using only topography metrics, 75% using only structure metrics,
and 79% using both structure and topography metrics. Predictive model accuracies were shown to be
much lower, averaging 33% accuracies for all 3 predictor inputs. These lower accuracies appear to be
spatially correlated and may potentially be explained by variables not captured in previous field data
collection efforts. In order to better understand and explain these results, further investigation into
these errors is needed and will likely be performed in Fall 2016.
192
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Vegetation Chlorophyll estimated from multi-angle
MODIS and tower hyperspectral observations: A tool for
scaling ecosystem seasonality and leaf demography across
Amazonian evergreen forests
Thomas Hilker1,2 , Lênio Soares Galvão3, Luiz E. O. C. Aragão3, Yhasmin M. de Moura3, Cibele H. do Amaral4, Alexei I.
Lyapustin5, Jin Wu6, Loren P. Albert6, Marciel José Ferreira7, Liana O. Anderson8, Victor A. H. F. dos Santos9, Neill Prohaska6,
Edgard Tribuzy9, João Vitor Barbosa Ceron9, Scott R. Saleska6, Yujie Wang10, José Francisco de Carvalho Gonçalves9, João
Victor Figueiredo Cardoso Rodrigues7, Maquelle Neves Garcia9
Oregon State University, College of Forestry, Corvallis Oregon 97330, USA, 2 University of Southampton, Department
of Geography and Environment, Southampton SO17 1BJ, United Kingdom, 3 Instituto Nacional de Pesquisas Espaciais
(INPE), São José dos Campos - SP, 12227-010, Brazil, 4 Federal University of Viçosa, Viçosa - MG, 36570-900, Brazil, 5 NASA
Goddard Space Flight Center, Greenbelt, MD 20771, 6 University of Arizona, Department of Ecology and Evolutionary
Biology, Tucson, AZ 85721, 7 Federal University of Amazonas, Manaus - AM, 69077-000, Brazil, 8Centro Nacional de
Monitoramento e Alertas de Desastres Naturais, São José dos Campos - SP, 12247-016, Brazil, 9 Instituto Nacional de
Pesquisas da Amazônia (INPA), Manaus - AM, 69067- 375, Brazil, 10 Joint Center for Earth System Technology, University of
Maryland Baltimore County, Baltimore MD
1
Keywords: leaf pigments; phenology; photosynthetic capacity; MAIAC, tropical ecosystem
Understanding controls on seasonal and interannual variability in vegetation phenology is of
paramount importance for Earth system modeling.
Here, we introduce a new approach for scaling
tropical ecosystem phenology by using multi-angle
Moderate Resolution Imaging Spectroradiometer
(MODIS) observations to invert a fully coupled canopy
reflectance model (ProSAIL). Monthly estimates of
vegetation leaf pigments derived from the inverse
model showed strong seasonal variations across
two flux-tower sites in the central Amazon basin
with marked increases in chlorophyll concentrations
during the early dry season. Remotely sensed
chlorophyll concentrations were strongly correlated
to field measurements (r2=0.64 and r2=0.98). We also
utilized hyperspectral data from a camera installed
on a tower to scale modeled chlorophyll pigments to
MODIS observations (r2=0.73). Chlorophyll pigment
concentrations (ChlA+B) were correlated to changes in
the amount of young and mature leaf area per month
(0.59 r2 0.64). Increases in MODIS observed ChlA+B
were preceded by increased photosynthetically
active radiation (PAR) during the dry season (0.61
r2 0.62) and followed by changes in net carbon
uptake. This allows the conclusion that, at these two
sites, seasonality of plant productivity is controlled by
vegetation phenology rather than sunlight directly.
Remote sensing of leaf pigment concentrations
can help to improve our understanding of seasonal
changes the photosynthetic leaf area’s efficiency to
absorb sunlight.
193
Development of Methods
194
Universidad Mayor
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Forest Monitoring
195
Universidad Mayor
ForestSAT 2016 Abstracts Summary
A polyalgorithm for land cover trend and change detection
Rishu Saxena1, Layne Watson1, Randolph Wynne2,and ValerieThomas2
1
Department of Computer Science, Virginia Tech.
Department of Forest Resources and Environmental Conservation, Virginia Tech.
2
Keywords: Time series analysis, satellite images, landsat imagery, change detection
Abstract:
Forest monitoring and management can benefit immensely from satellite images. Time series analysis
of satellite imagery is one way of extracting the wealth of information held in these images. Several
algorithms to this end have been proposed by different groups in the remote sensing community
utilizing different approaches: analytics based, mathematical functions based, some purely temporal,
some spatio-temporal, and the like.
However, the design and selection of these algorithms, so far, appears to be predominantly context
specific, i.e., most of these proposed methods seem to perform well on the dataset that they are
designed for but their performance on randomly picked datasets from across the globe has not been
studied. In general, no single algorithm designed so far will work for all regions is highly likely. The
same is also indicated by our experiments.
In this talk, we will present a novel time series analysis algorithm, specifically a `polyalgorithm’, to
circumvent this gap. Our polyalgorithm consists of four different time series analysis algorithms along
with a set of metrics for analyzing the requisite results to produce the final result. The algorithms we
have chosen are fundamentally unique to each other in their design as well as in the phenomenon they
capture. We have experimented with 6 different path/rows of Landsat images, collectively covering a
gamut of ecosystems. For scalability and robustness, all the codes are written in Fortran.
Our algorithm will be useful in understanding and monitoring deforestation, urbanization, changes
in tree canopy covers and such other processes, both anthropogenic as well as natural. It can thus be
directly utilized by the policy makers for managing forests. It will further be of use in a much broader
spectrum of applications including but not limited to general Earth monitoring (eg., land usage, land
cover), health monitoring and urban transportation.
197
Universidad Mayor
Forest Monitoring
AFIS - Wildfire Visualisation and Multi-sensor detection
capabilities
Philip Frost, CSIR Meraka Institute, South Africa
Keywords: Wildfire, detection, fire danger, burned area, mobile application, dashboard, satellites,
camera
The Council for Scientific and Industrial Research
(CSIR) lead by die Meraka Institute and supported by
partners such as the South African National Space
Agency (SANSA), University of Maryland and University of Wisconsin Madison have been involved
in the developed of the Advanced Fire Information
System (AFIS) over the last 12 years. The original aim
of the system was to provide near real time fire information to fire fighters, farmers and forest managers
across Southern Africa based on Earth Observation
satellite products from Terra and Aqua MODIS and
the Geostationary MSG. With the launch of the system in 2004, Eskom (South Africa’ and Africa’s largest power utility) quickly became the biggest user
of the system where more than 300 line managers
and support staff all around the country would receive cell phone and email fire alert messages whenever a MODIS or MSG active fire was detected within
2 km of any of the 28 000km of Eskom’s transmission
lines. Today the system has evolved in to a comprehensive wildfire information system that provides
users with information and intelligence around the
prediction, detection, monitoring and assessment of
wildfires globally. EO satellite wildfire detections are
complemented with both crowdsource detections
from the AFIS Watchtower app as well as through
automated in situ camera detection systems capable
198
of detecting thin smoke plumes up to 20 miles away.
Fire Danger models such as the Canadian FWI and
Australian McArthur Fire Danger models have been
fully integrated in to the system and can be produced for any area globally. A new MODIS burned area
product has been produced by merging the standard
MCD45 MODIS with the MCD64 BA product from
Louis Giglo while a new Landsat 8 and Sentinel 2 algorithm has also been developed to provide higher
resolution mapping of new fire events.
The visualization of AFIS has evolved from an online
web viewer to a hybrid wildfire dashboard providing
situational awareness for any region of interest. The
NASA World View integration in to the dashboard
has dramatically improved the visualisation component within the dashboard and now provide users
with access to not only the Direct readout station
data but also all the NASA GIBS data sets.
The AFIS mobile app is available on both the ITunes
and Play stores and provides users with access to fire
detections, fire weather forecasts and fire history
through their mobile devices. The new AFIS Watchtower app allow fire managers to report the location
of a new fire through their phone camera and is currently been tested by fire managers in Portugal.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Application of high-resolution satellite data for monitoring
forest areas in changeable climatic conditions
Zbigniew Bochenek1*, Dariusz Ziolkowski1 , Maciej Bartold2,1, Karolina Orlowska2, Bogdan Zagajewski2
Institute of Geodesy and Cartography, Remote Sensing Centre, Modzelewskiego 27, 02-679 Warsaw, Poland, email:
zbigniew.bochenek@igik.edu.pl; dariusz.ziolkowski@igik.edu.pl; maciej.bartold@igik.edu.pl
2
Department of Geoinformatics, Cartography and Remote Sensing, Faculty of Geography and Regional Studies, University
of Warsaw, Poland, email: bogdan@uw.eu.pl; karolina.orlowska@uw.edu.pl
* Corresponding author
1
Keywords: climate change, forest monitoring, high-resolution data, vegetation index
Abstract
The main objectives of the research work were to determine usefulness of high-resolution optical
satellite data for monitoring various forest parameters and to assess impact of changeable climatic
conditions with the use of vegetation indices derived from remotely sensed data. Three types
of high-resolution satellite data were used in the study: Landsat TM/OLI, SPOT 5 and Sentinel 2
images. Three forest areas located in a temperate climatic zone in northeastern Poland, differing in
environmental conditions were taken into consideration. Five vegetation indices describing various
aspects of plant condition and vegetation structure - Normalized Difference Vegetation Index NDVI, Enhanced Vegetation Index - EVI, Triangular Vegetation Index - TVI, Normalized Difference
Infrared Index - NDII, Disease Water Stress Index - DSWI - were derived from satellite images. Their
values were analyzed in a temporal profile in four vegetation seasons: 2006, 2014, 2015 and 2016, in
conjunction with meteorological parameters – temperature and precipitation. The results of analyses
proved that dedicated vegetation indices characterizing water stress in plants – Disease Water Stress
Index (DSWI), Normalized Difference Infrared Index (NDII) and Triangular Vegetation Index (TVI) are
sensitive to changeable climatic conditions, especially detecting drought impact on forest canopies.
The indices are also useful for characterizing types of forest site, tree stand mixture and tree species.
In particular they give an opportunity to differentiate dry, fresh and humid forest sites, three levels of
mixing conifers and hardwoods, as well as some species within deciduous forests.
Introduction
Forest monitoring with the use of remote sensing
techniques is an important topic, both in global and
regional scale. As far a global scale is concerned,
there were many initiatives aimed at application
of low and medium-resolution satellite data for
deriving information on forest state and its changes
[Boyd and Danson, 2005; Blackard et al., 2008;
Soudani el al., 2008; Asner et al., 2009]. The most
important international projects dealing with
this topic were: TREES, REDD, Geoland Forest
Monitoring, GLOBBIOMASS and others [Gibbs et
al., 2007; Bicheron et al., 2008; Le Toan et al., 2011]
. But parallel to forest monitoring at a global scale
numerous projects have been conducted in order
to apply high-resolution satellite images for forest
mapping and for estimating forest parameters
[Puzzolo et al., 2003; Huang et al., 2010; Morton et
al., 2011; Miettinen et al., 2014]. Different types of
satellite data, both optical and microwave, have
been used for this purpose and various aspects of
forest monitoring were taken into account. Several
practical conclusions on applicability of remotely
sensed data for forest monitoring have been already
drawn, nevertheless due to complexity of forest
environment and its differentiation through various
climatic zones there is still a need to further develop
and improve the existing methods. The presented
work is a contribution to that development,
199
Universidad Mayor
Forest Monitoring
concentrating on the use of optical high-resolution
satellite images for characterizing various forest
parameters and for studying impact of climate
changes on forest behavior in a temperate zone.
Methodology
Three forest areas located in northeastern Poland
have been selected as pilot sites for the presented
research work. These are: Bialowieska Forest,
Knyszynska Forest and Borecka Forest. Northeastern
Poland, where the sites are located, is under
influence of continental climate, characterized
by impact of polar air masses, shorter vegetation
period comparing to the rest of Poland and quite
high temperature fluctuations. There are some
differences in climatic conditions between sites
– Borecka Forest is slightly influenced by oceanic
climate which has a certain impact on vegetation
development. The selected forest areas are also
different, as far as tree species and forest sites are
concerned.
Three types of high-resolution satellite images have
been used: Landsat TM/OLI images, SPOT 5 images
and Sentinel 2 images. Apart from satellite images
ground reference data have been collected for the
regions of interest. The detailed digital forest maps
prepared by Forest Service have been compiled;
they comprise comprehensive information on
forest, including species, stand mixture and type
of forest site. Moreover, ground measurements
characterizing real plant condition have been
conducted
(spectroradiometric,
fluorescence,
pigment content measurements).
In order to study, if specific vegetation indices derived
from high-resolution satellite data are sensitive
to variable climatic conditions characterized by
changes of meteorological parameters and if they
change due to some forest parameters – type of
forest site, tree species and tree stand mixture, a set
of indices was generated, namely:
Normalized Difference Vegetation Index – NDVI
characterizing general plant condition
NDVI = (NIR - RED) / (NIR + RED)
Normalized Difference Infrared Index – NDII
characterizing water content in plants
NDII = (NIR – SWIR 1) / (NIR + SWIR 1)
Disease Water Stress Index – DSWI characterizing
stress of plants due to water shortage and damage
DSWI = (NIR – GREEN) / (SWIR 1 + RED)
200
Triangular Vegetation Index – TVI characterizing
plant condition and chlorophyll content in red-edge
TVI = 0.5 * [120 * (R750 – R550) – 200 * (R670 – R550)]
All indices were analyzed in a temporal profile. In
order to quantify their changes and to relate them
to climatic conditions, simple index characterizing
decrease of vegetation index was produced using
the following formula:
Decrease index value = Max index value at peak season – Min index value at the end of season
Regarding tree species six species represented
within study areas were selected for the analysis:
two coniferous species – spruce and pine and
four deciduous species – birch, alder, oak and
hornbeam. Three types of forest site were taken
into consideration: fresh, humid and dry site, in
combination with various tree species – conifers and
hardwoods. Tree stand mixture was analyzed at three
levels of mixing: pure stands (90 – 100 %), 70 to 90 %
of dominant species, 50 to 70 % of dominant species.
In order to study relations between vegetation
indices derived from high-resolution satellite data
and climatic conditions meteorological information
was compiled for the regions of interest, using webavailable database http://en.tutiempo.net/climate/
poland.html. It comprises various meteorological
parameters e.g. mean daily temperature and
precipitation. These parameters at first stage of
the works were processed in order to produce socalled hydrothermal index, which combines both
temperature and precipitation information in
10-day’s cycle, characterizing drought conditions for
the study area.
Results
At first stage of the works thorough analysis of the
selected vegetation indices has been performed,
in order to find those which are more useful for
differentiating particular tree species and which
can respond to variable climatic conditions in a best
way. The analysis was done for indices derived from
Landsat, SPOT and Sentinel 2 data with the aim
to compare usefulness of these types of satellite
images for forest studies. Comparative analysis
of four vegetation indices – NDVI, NDII, DSWI
and TVI revealed that Disease Water Stress Index
(DSWI) derived from Landsat images and Triangular
Vegetation Index (TVI) derived from Sentinel 2 data
are the most sensitive to species differentiation,
Universidad Mayor
ForestSAT 2016 Abstracts Summary
analyzed through vegetation season. Both indices
demonstrate possibilities of differentiating some
tree species, e.g. pine and spruce from coniferous
group, as well as alder, birch and hornbeam from
deciduous group. Changes of indices for 2015 and
2016 are presented in figure 1.
of correlation analysis revealed, that there is a
quite strong relationship between vegetation index
derived from remotely sensed data and the index
characterizing climatic conditions - the correlation
coefficient r was equal to -0.723. The results are
presented in graphical form in figure 2.
As both analyzed indices – DSWI and NDII revealed
some decreasing trends at the second part of
vegetation period, it was decided to verify, if these
changes are related to meteorological situation. For
this purpose hydrothermal index (HT) which informs
on drought has been calculated and next DSWI
index based on Landsat data has been computed
for all vegetation seasons between 2000 and 2015.
Next differences between maximum and minimum
DSWI value were determined and correlated
with the mean hydrothermal index derived from
meteorological data for the years of interest, using
Statistica software package for this purpose. Results
The second part of the research work was devoted
to analysis of impact of type of forest site and stand
mixture on values of vegetation indices derived from
high-resolution data. The analyses conducted for
three types of forest site – dry, fresh and humid site,
as well as for three levels of stand mixture revealed
that both forest parameters have an influence on
values of DSWI and TVI indices, decreasing level of
index in case of dry site and increasing its value in
case of mixing coniferous forests with deciduous
species. These impacts should be taken into account
while analyzing relationships between vegetation
indices and climatic conditions.d
Fig. 1. Changes of vegetation indices - DSWI and TVI for 2015 vegetation season
Fig. 2. Results of correlation analysis between changes of DSWI index and hydrothermal index
201
Forest Monitoring
Conclusions
The main conclusion from the presented study
is that there are relations between changeable
meteorological conditions and forest condition,
expressed by remote sensing based indices, but
these relations vary depending on type of forest
site and degree of tree species mixture. It was
found that vegetation indices derived from highresolution satellite data, which include information
on spectral reflectance both in near infrared and
shortwave infrared bands, can be effective for
evaluating stress in forest stands due to drought
conditions. The second important conclusion is that
some basic characteristics of forest stands, such as
forest site type and degree of tree species mixture
have impact on remote sensing based parameters,
while applying adequate vegetation indices, such as
Disease Water Stress Index (DSWI) and Triangular
Vegetation Index (TVI), which modify the general
relations between climatic changes and forest
condition. While analyzing relations between EObased vegetation indices and forest characteristics it
was found that deciduous stands are more sensitive
to drought conditions, as expressed by changes of
DSWI index, than coniferous forests. Comparing
forest sites it was proved that coniferous stands
located on dry forest site are more resistant to longterm drought than those situated on fresh and humid
sites. These conclusions open new possibilities for
the detailed forest studies and assessments, but also
point out complexity of analysis of satellite images,
when applied for forest monitoring. So finally one
can conclude, that optical, high-resolution satellite
images are useful for assessing impact of climate
change and for analyzing main environmental
features of forests, but in order to derive more
information e.g. on forest structure synergistic use of
optical and microwave images should be considered.
References
Asner G.P., Knapp D.E., Balaji A., Paez-Acosta
G., 2009. Automated mapping of tropical
deforestation and forest degradation: CLASlite.
J. Appl. Remote Sens., 3 (2009) 033543, 24p.
http://dx.doi.org/10.1117/1.3223675
Bicheron, P., P. Defourny, C. Brockmann, L.
Schouten, C. Vancutsem, M. Huc, S. Bontemps,
M. Leroy, F. Achard, M. Herold, F. Ranera, and O.
Arino. 2008. GLOBCOVER Products Description
and Validation Report. MEDIAS-France, ftp://
202
Universidad Mayor
uranus.esrin.esa.int/pub/globcover_v2/global/
Blackard, J.A., Finco M.V., Helmer E.H., Holden G.R.,
Hoppus M.L., Jacobs D.M., Lister A.J., et al. 2008.
Mapping US forest biomass using nationwide
forest inventory data and moderate resolution
information. Remote Sensing of Environment112
(4): 1658–1677.
Boyd D.S, Danson F.M. 2005. Satellite remote
sensing of forest resources: three decades of
research development. Prog Phys Geogr 2005;
29(1): pp. 1-26. [doi:101191/0309133305pp432ra]
Gibbs, H.K., Brown S., Niles J.O. and Foley J.A.. 2007.
Monitoring and estimating tropical forest carbon
stocks: Making REDD a reality. Environmental
Research Letters 2(4): 045023.
Huang C., Goward S.N., Masek J.G., Thomas N.,
Zhu Z., Vogelmann J.E., 2010. An automated
approach for reconstructing recent forest
disturbance history using dense Landsat time
series stacks. Remote Sens. Environ., 114 (2010),
pp. 183–198
Le Toan T., Quegan S., Davidson M.W.J., Balzter H.,
Paillou P., Papathanassiou K., Plummer S., Rocca
F., Saatchi S.S., Shugart H., Ulander L., 2011.
The BIOMASS mission: mapping global forest
biomass to better understand the terrestrial
carbon cycle. Remote Sens. Environ., 115 (2011),
pp. 2850–2860
Miettinen J., Stibig H.S., Achard F., 2014. Remote
Sensing of forest degradation in Southeast Asia
– aiming for a regional view through 5-30 m
satellite data. Global Ecology and Conservation,
Vol. 2 December 2014, pp. 24-36
Morton D.C., DeFries R.S., Nagol J., Souza Jr C.M..,
Kasischke E.S., Hurtt G.C., Dubayah R., 2011.
Mapping canopy damage from understory fires
in Amazon forests using annual time series of
Landsat and MODIS data. Remote Sens. Environ.,
115 (2011), pp. 1706–1720
Puzzolo V, De Natale F, Giannetti F. 2003. Forest
species discrimination in an alpine mountain
area using a fuzzy classification of multitemporal SPOT (HRV) data. Geoscience and
Remote Sensing Symposium, 2003 IGARSS ‘03
Proceedings 2003 IEEE. Int 2003; 4: pp. 2538-40.
Soudani, K., le Maire G., Dufrene E., Francois C.,
Delpierre N., Ulrich E. and Cecchini S., 2008.
Evaluation of the onset of green-up in temperate
deciduous broadleaf forests
derived from
Moderate Resolution Imaging Spectroradiometer
(MODIS) data. Remote Sensing of Environment
112 (5): 2643–2655.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Assessing the cumulative climatic effects on regional forest
decline dynamics in coniferous forests
David M. Bell, USDA Forest Service, Pacific Northwest Research Station
Warren B. Cohen, USDA Forest Service, Pacific Northwest Research Station
Zhiqiang Yang, Oregon State University, College of Forestry
Keywords: coniferous forests; ecological drought; forest decline; Landsat; TimeSync
Abstract
Protracted, multi-year changes in forest vegetation have become the dominant form of disturbance
in the western United States and these forest declines are generally attributed to climate-mediated
stressors like drought, insects, and disease. However, climatic effects compound over time (e.g., multiyear droughts), implying that examination of forest declines must incorporate cumulative climatic
effects. Our research assesses the contribution of climate to remotely sensed forest decline (RSFD).
We leveraged a two-stage stratified random sample of coniferous forests in the western United States
and the TimeSync Landsat time series visualization tool to use 30-m multispectral imagery (Landsat
TM and ETM+; 1984-2012) and repeat photography (Google Earth) to derive independent plot-based
estimates of the presence or absence of RSFD at each sampling location.
Using MaxEnt, a machine learning algorithm for modeling presence data, we found that previous year’s
precipitation, current year’s maximum temperature, and trends in temperature and precipitation during
the previous five years all contributed to observed patterns of RSFD, implicating ecological drought
(decreasing precipitation and increasing temperature) as a driver of RSFD. Stochastic Antecedent
Modeling, a hierarchical Bayesian statistical approach to examining lagged, or cumulative, effects,
provided a more nuanced image of climatic influences on RSFD. Preliminary results for a subset of
spruce, fir, and mountain hemlock (SFMH) forest locations experiencing decline (n = 50) indicated that
maximum monthly spring temperatures (April – June) and total monthly summer precipitation (July
– September) exhibited cumulative effects on RSFD timing up to 8 years into the past. SFMH forest
decline was most sensitive to cumulative temperature and precipitation in colder climates, highlighting
sensitivity of high elevation forests to drought-induced decline. Additionally, RSFD associated with
high temperatures was associated with dense, closed canopy forests (i.e., high tasseled-cap wetness),
whereas it was associated with sparse forest canopies (i.e., low angle) for low precipitation. These
results indicate that forest vulnerability to drought-induced decline will vary regionally, that these
declines develop over the course of several years, and that local variation in forest structure will
mediate the impact of interannual variation in climate on forest health.
203
Forest Monitoring
Universidad Mayor
Assessing the predictions of high-resolution climate surfaces:
a statistical analysis in a Southern Hemisphere country
Christian Salas1, Luis Morales2, Renato Cifuentes3, Rodrigo Vargas1, Andres Fuentes-Ramirez1
Laboratorio de Biometría, Departamento de Ciencias Forestales, Universidad de La Frontera, PO
Box 54-D, Temuco, Chile
2
Laboratorio de Investigación en Ciencias Ambientales, Departamento de Ciencias Ambientales y
Recursos Naturales Renovables, Universidad de Chile, PO Box 1004, La Pintana, Santiago, Chile.
3
KU Leuven, Department of Biosystems, Division of Crop Biotechnics, Willem de Croylaan 34, BE3001 Leuven, Belgium
1
Keywords: temperature, precipitation, interpolation, uncertainty, hypothesis testing.
Climate variables are needed in several disciplines
ranging from engineering to natural sciences
for management, policy making, and scientific
purposes. In order to overcome the problems of
rather scatter available climate data in some places,
several interpolated climate surfaces (ICS) have
been produced worldwide. These climate variable
surfaces are spatial grids at high spatial resolution
(e.g., 1 km2), and are currently used for many
applications. The WorldClim project is one of the
most used ICS because it has been made widely open
and is freely available on the Internet. In forestry,
ecological, and geographical applications, a ICS is
not only used for imputing missing climate variables,
but also for mapping and modelling climate change
and its effects on related variables, such as, species
distribution and population changes. However,
neither its uncertainties nor its predictions had been
thoroughly analyzed. We assessed the monthly
predictions of WorldClim for four climate variables
(maximum, mean, and minimum temperature; and
precipitation) through all Chile (i.e., over an area of
more than 750000 km2, through an extension of 4270
km) based on a large set of weather station data,
204
to create concurrent observations. We assessed
the prediction errors, as well as the performance
of WorldClim by conducting statistical hypothesis
testing on the estimated parameters of an observedpredicted simple linear regression model. The
results indicate that only the maximum temperature
is correctly predicted during the entire year by
WorldClim. The precipitation is better predicted
during summer and spring seasons (i.e., September
– February) than during autumn and winter seasons
(i.e., March – August). However, there are statistical
differences for predicting both the minimum and
mean temperature for all months. We point out
that WorldClim predictions although have several
problems, are quite useful in practice. Nevertheless,
we recommend calibrating or validating the climate
predictions as much as possible in any study that
would heavily rely on climate data. This process will
provide insights into the geographic distribution of
uncertainties and will allow for improvements of
climate variable surfaces at locations of interest.
Finally, we demonstrate a statistical assessment
based on both prediction capabilities and hypothesis
testing.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Assessment of forest productivity from MODIS NPP data in
relation to forest management and optimal leaf area index
Mait Lang1,2, Tiit Nilson1, Mathias Neumann3, Adam Moreno3
2
1
Tartu Observatory, 61602 Tõravere, Tartumaa, Estonia; e-mail: lang@to.ee.
Institute of Forestry and Rural Engineering, Estonian University of Life Sciences, Kreutzwaldi 5, Tartu 51014, Estonia.
3
Institute of Silviculture, Department of Forest and Soil Sciences, University of Natural Resources and Life Sciences,
Vienna, Peter-Jordan-Str. 82, A-1190 Wien, Austria.
Keywords: hemi-boreal forest, optimum LAI, net primary production, Estonia.
Abstract
Global scale maps of terrestrial ecosystem net primary production (NPP) are routinely produced
based on coarse spatial resolution satellite data (Zhao, et al. 2005). The NPP is estimated by using
Monteith hypothesis which states that total photosynthesis product (GPP) is proportional to absorbed
photosynthetically active radiation and light use efficiency and limited by air temperature and water
vapor pressure deficit. NPP is obtained by subtracting growth and maintenance losses from GPP. It
can be expected that the NPP estimates are related to forest growth and disturbances.
Figure 1. Envisat MERIS data based
NPP, site fertility, % coniferous
and forest management induced
disturbances (1987-2013) for 1km2
pixels clustered according to
disturbance signatures.
We used three NPP datasets to assess the impact of forest management induced disturbances to
forest productivity in hemiboreal forests in Estonia. Landsat TM based regeneration felling maps
(U. Peterson, Tartu Obsevatory) covering years 1987-2013 with about 5-year interval were used to
calculate disturbance signatures for 1km2 forested pixels of NPP maps. Three independent NPP
estimates were used (Zhao, et al. 2005, Nilson et al. 2012, Neumann et al. 2016). The mean NPP
estimate was positively correlated with disturbances and site fertility and negatively correlated with
the share of coniferous (Figure 1).
The NPP relationship with site fertility can be expected to be positive and negative correlation of NPP
with the share of coniferous can be explained by the tree species preferences to sites and also forest
spectral signatures and the NPP model constants. However, the positive correlation between NPP
and disturbances can be induced by the optimum LAI phenomena in the NPP algorithms (Nilson et al.
2014) happening due to saturating relationship between fAPAR and LAI and increasing maintenance
losses with increasing LAI.
205
Forest Monitoring
References
Neumann, M., Moreno, M., Thurnher, C., Mues,
V., Härkönen, S., Mura, M., Bouriaud, O., Lang,
M., Cardellini, G., Thivolle-Cazat, A., Bronisz,
K., Merganic, J., Alberdi, I., Astrup, R., Mohren,
F., Zhao, M., Hasenauer, H. (2016). Creating a
regional MODIS satellite-driven Net Primary
Production dataset for European forests. Remote
Sensing, (In review).
Nilson, T., Rennel, M., Luhamaa, A., Hordo, M., Olesk,
A., Lang, M. (2012). MERIS GPP/NPP product for
Estonia: I. Algorithm and preliminary results of
simulation. Forestry Studies | Metsanduslikud
Uurimused 56, 56–78.
Nilson, T., Rennel, M., Lang, M. (2014). MERIS
GPP/NPP product for Estonia: II. Complex
meteorological limiting factor and optimum leaf
area index. Forestry Studies | Metsanduslikud
Uurimused 61, 5–26. Zhao, M., Heinsch, F.A.,
Nemani, R., Running, S. (2005). Improvements
of the MODIS terrestrial gross and net primary
production global dataset. Remote Sensing of
Environment 95, 164–176.
206
Universidad Mayor
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Automatic recognition of burned areas with the use of a
support vector machine (SVM) using VNIR spectral bands with
multiple satellite sensors.
Manuel Castro1, Patricio Acevedo1
1
Laboratorio de Teledetección Satelital, Departamento de Ciencias Físicas, Facultad de Ingeniería
y Ciencias, Universidad de La Frontera, Casilla 54-D, Temuco, Chile.
Keywords: SVM, burned areas, Automatic recognition, VNIR
Abstract
The increasing need to monitor and oversee burned areas with greater precision has resulted in a
series of studies which are focused on improving the accuracy and efficiency of mapping using satellite
platforms with high resolution spatial sensors, those which generally only incorporate spectral bands
in the visible and near-infrared range (VNIR). Traditionally for the study and detection of burned areas
the near-infrared (NIR) and mid-infrared (SWIR) spectral zones have been used, given their sensitivity
to this type of land cover, but only the medium and low spatial resolution sensors incorporate SWIR
bands, in oriented studies usually at the regional level. The present study seeks to classify burned
areas using visible and near-infrared (VNIR) spectral bands, using a support vector machine (SVM).
The analysis exclusively considers pixels of the VNIR bands from different satellite platforms such as
Fasat-C, SPOT, GeoEyes, WorldView-2, Landsat-8 and Sentinel-2A, as well as some spectral indices
derived from this zone of the spectrum. The results of the classification based on pixels from different
servers using an SVM indicate that the use of spectral indices combined with VNIR spectral bands,
instead of just the VNIR spectral bands, improves the accuracy and reliability of the recognition of
burned areas. A SVM is shown effective since using images from different sensors still does not have
major problems in detection. Nonetheless, volcanic slag zones tend to be classified as areas affected
by burning mainly because the spectral signature is very similar in the VNIR spectral region, but this is
solvable applying a mask of non-vegetational soil uses.
207
Universidad Mayor
Forest Monitoring
POSTER
Barren ground caribou (Rangifer tarandus groenlandicus)
behaviour after recent fire events; integrating caribou
telemetry data with Landsat fire detection techniques
Gregory J.M. Rickbeil1, Txomin Hermosilla1, Nicholas C. Coops1, Joanne C. White2, Michael A. Wulder2
1
Integrated Remote Sensing Studio, Faculty of Forestry, University of British Columbia, 2424 Main Mall, Vancouver, BC,
Canada, V6T 1Z4; phone: 604-827-4429; fax: 604 822 9106
2
Canadian Forest Service (Pacific Forestry Centre), Natural Resources Canada, 506 Burnside Road, Saanich, British
Columbia, Canada, V8Z 1M5
Keywords: arctic, disturbance, Landsat, mammal, movement, remote sensing, telemetry, ungulate
Abstract
Fire is a major factor affecting access to high quality forage such as terricholous lichens for barren
ground caribou (Rangifer tarandus groenlandicus). Herein, we characterize how the size and severity of
fires are changing across five barren ground caribou herd ranges occurring in the Northwest Territories
and Nunavut, Canada. Additionally, we demonstrate how time since fire, fire severity, and time of year
(season) result in complex changes in caribou behavioural metrics estimated using telemetry data. Fire
disturbances were identified using novel time series of gap-free Landsat surface reflectance composites
from 1985 to 2011 across all herd ranges. Burn severity was estimated using the differenced normalized
burn ratio (dNBR). Annual area burned and burn severity were assessed through time for each herd
and, using generalized additive mixed models, were related to two behavioural metrics: velocity and
relative turning angle. Neither annual area burned nor burn severity displayed any temporal trend
within the study period. However, certain herds, such as the Ahiak and Beverly, have more exposure
to fire than other herds (i.e. Cape Bathurst had a maximum forested area burned of less than 4 km2).
Time since fire and burn severity both significantly affected velocity and relative turning angles. Across
all seasons, fire virtually eliminated foraging-focused behaviour for all 26 years of analysis while more
severe fires saw a marked increase in movement-focused behaviour compared to unburnt patches.
Between seasons, caribou used burned areas as early as one-year post fire, demonstrating complex,
non-linear reactions to time since fire and fire severity, and these reactions changed depending on
season. In all cases movement-focused behaviour was detected post fire. We conclude that changes in
caribou behaviour immediately post-fire are primarily driven by changes in forest structure rather than
changes in terricholous lichen availability.
208
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Bringing Earth Observation Services for Monitoring Dynamic
Forest Disturbances to the Users – EOMonDis
F.Enßle, T. Haeusler, S. Gomez, C. Storch, M. Pape, H. Ott, T. Wagner, G. Ramminger
GAF AG
Munich, Germany
Keywords: Forest Monitoring, REDD+, Zero Deforestation, Dense Time Series, Tropical Dry Forests
Abstract
The geographical extent and limited accessibility of tropical forest ecosystems hinders accurate
estimation and spatial explicit monitoring of forest resources by purely field-based inventories. The
spatial explicit representation of forest ecosystem status and changes thereof is a key requirement to
protect natural forests, enable effective forest law enforcement and support sustainable management
of forest resources. This paper introduces the background of the EOMonDis project, mapping service
requirements and the methodological concept for algorithm development and product verification.
Furthermore, the processing chain and results for optical based dense time series analysis for one out
of four test sites are described.
Introduction
During the last decade, remote sensing has become
an integral part of national and international
forest policy programmes. An important forest
management policy is embedded in the United
Nations Framework Convention on Climate Change
(UNFCCC) programme Reducing Emissions from
Deforestation and Forest Degradation and the
role of conservation, sustainable management of
forests and enhancement of forest carbon stocks in
developing countries (REDD+), which was started in
2005 by an initiative of a few countries. The initial
effort was further developed such that the REDD+
mechanism is based on a three phased approach, of
which the final phase relates to payments for verified
emission reductions. These financial incentives
should preserve carbon stored in natural forest
ecosystems and reduce emissions from forested
lands. An essential component of such a system is
a reliable monitoring system of forest resources
and changes thereof. At the same time, the private
sector companies of the commodity sector want
to establish deforestation free production chains.
Within this context, a methodological approach was
developed to identify areas for potential expansion
and identify areas of High Carbon Stocks (HCS)
that shall be preserved [1]. Thus, for the successful
implementation of forest policies, forest resources
have to be mapped spatially and subsequently be
monitored to track changes on a regular basis.
Mapping of activity data (AD) furthermore provides
a valuable information source for a sustainable forest
management. With the launch of first satellites of
the Sentinel constellation, optical and radar data
of high spatial and temporal resolution are freely
available. In the near future, the temporal coverage
will be even increased by the launch of other
satellites complementing the constellation. With
the high repetition rate of the Sentinel-2 satellites,
complemented by Sentinel-1 data and Landsat
sensors, fast response monitoring systems can be
implemented with low Earth Observation (EO) data
costs. Up-to date information on activity data can
thereby be provided to realise an effective forest
management.
“Bringing Earth Observation Services for Monitoring
Dynamic Forest Disturbances to the Users” project
(EOMonDis) is a project supported by the European
Commission’s Horizon 2020 programme and is being
implemented from 2016-2019. The overall aim is to
209
Forest Monitoring
improve tropical forest monitoring products that are
based on satellite borne high-resolution sensor data
and of high temporal resolution. The project focuses
especially on developing state-of-the-art processing
chains based on dense time series and suitable
for large area mapping by integration of available
optical and radar data with particular emphasis on
Sentinel 1 and 2 data.
This paper will present in the next Section the main
forest monitoring requirements for the UNFCCC
REDD+ policy as well as for the Zero Deforestation
(ZD) programmes. Section 3 will summarise the
Study Sites that will be used in the current project.
The following Section 4 will then present the planned
methodological approaches and the paper will end
with a concluding section.
Forest Monitoring Requirements
Countries that are willing and able to reduce
emissions from deforestation and forest
degradation are recommended by the United
Nations Framework Convention on Climate Change
(UNFCCC) Conference of Parties (COP) to establish
robust and transparent forest monitoring systems
to account for anthropogenic forest-related
greenhouse gas emissions by sources and removals
by sinks. The methodological approach proposed by
Intergovernmental Panel on Climate Change (IPCC)
for carbon emission accounting requires two basic
data: (1) activity data (area extent of the activity)
and (2) emission factor (carbon stock per unit area).
Consequently, potential REDD+ countries need to
estimate emissions and changes in forest carbon
stocks from deforestation or forest degradation and
have a means to establish reference emission levels,
against which future emissions can be compared, as
well as to address the displacement of emissions.
Additionally the IPCC Good Practice Guidance
(GPG) for Land Use, Land Use Change and Forestry
(LULUCF) defines 3 Tiers or levels of accuracy for
reporting which have to be considered for the REDD
services. “Moving from Tier 1 to Tier 3 increases the
accuracy and precision of the estimates, but also
increases the complexity and the costs of monitoring”
[2]. Furthermore, in the case of deforestation the
GPG presents three approaches to obtaining areal
extents: “(i) only identifying the total area for each
land category (approach 1); (ii) tracking of land-use
210
Universidad Mayor
changes between categories (approach 2); and (iii)
tracking land-use changes using sampling or wall-towall mapping techniques (approach 3). Approach 3 is
the only approach that tracks forest and other land
conversions on an explicit spatial basis, including
gross deforestation and gross change in other land
cover classes” (FCCC/TP/2009/1).
For the establishment of baseline or reference
emission levels, there is a need for robust and costeffective methodologies to estimate and monitor
changes in forest cover and associated carbon
stocks and greenhouse gas emissions. Thus, more
specifically countries will need to know:
● The aerial extent of deforestation and forest
degradation (hectares),
● For degradation, the proportion of forest biomass
lost (percentage),
● Where the deforestation or forest degradation
occurred (which forest type),
● The carbon content of each forest type (tons of
carbon per hectare), and
● The process of forest loss that affects the rate
and timing of emissions.
The Zero Deforestation program is a further
development of the High Conservation Value
(HCV) approach, addressing concerns that the HCV
approach may allow for the conversion of forests,
which do not fall under a HCV category, but storing
at the same time high amounts of carbon. The ZD
program and the effective evaluation of companies’
commitments require two different kind of forest
monitoring products. For the development of new
plantations companies that are committing to ZD
have to restrict their expansions to locations of low
carbon stocks and preserve areas of high value.
With the knowledge of the spatial explicit location
of high carbon stocks within the concessions the
evolvement of these forests can be monitored over
time. An efficient monitoring of forest resources is
therefore required to help companies in confirming
their ZD commitments. The two main forest
monitoring products can be summarised as:
● The aerial extent and amount of forest carbon
stocks (tons of carbon per hectare),
● The process of forest loss in areas under
protection (hectares).
Universidad Mayor
Study Sites
Within EOMonDis project, study sites in four
countries have been established to adequately cover
a wide range of tropical forest compositions. The
study sites are located in Gabon, Cameroon, Malawi
and Peru.
Cameroon
Cameroon has a South to North extent from around
latitude 2° to 12° N. The transition from North to
South is represented by various climatic conditions
and land cover types. These can be roughly divided
in three major zones. The northern part is mainly
covered by croplands and open woody vegetation.
The middle part of Cameroon has deciduous
woodlands with some areas of mosaic forest and
savanna. The southern part has large areas of closed
evergreen forests with a transition zone of deciduous
woodlands in the west and Mosaic Forest in the east.
ForestSAT 2016 Abstracts Summary
southern part of the country. The mean annual
temperatures in these areas are about 13° C, whereas
in the lower regions annual temperatures are up
to 25° C. The highlands have at the same time high
annual precipitation of more than 1600mm and in
the cold dry season precipitation is provided by mist,
resulting in a consistent cloud cover in the northern
mountain areas near the lake.
A.
The various forest cover types in the Centre province
qualified this region for method development in
the current project. To demonstrate the feasibility
of developed algorithms on different forest cover
types, two sites were established. The northern
demonstration site has an area of 15,000 km² and
is mainly covered by savanna. The second site has
an area extend of around 10,000 km² and covers a
highly fragmented and divers landscape with the city
Yaoundé in the centre of the Area of Interest (AOI).
Gabon
Gabon is located on the western side of Central
Africa and is intersected by the equatorial line. The
vegetation is affected by a typical equatorial climate,
characterised by high annual precipitation of more
than 2000mm and average daily temperatures of
about 24°C - 27°C with very low annual variations.
The selected test site (~38,000 km²) is located at
the western side of Gabon. As most of Gabon,
the selected site is covered by evergreen dense
equatorial forest as well as some part of savanna. The
site also includes some of the recent agro-industrial
development that have recently been taking place in
Gabon leading to substantial deforestation.
B.
Malawi
Situated between latitude 9° to 17° S and longitude
30° to 36° E, Malawi has a total area of ~ 120.000
km². Several highlands between 1300 masl and
3000 masl are located in the northern, middle and
C.
The main part of the country consists of flat plateaux
with elevations around 750 -1300 masl. Towards the
southern part, the elevation declines at around 180
masl. The main forest type is Miombo (Brachystegia
woodlands) generally found between 600 - 1500
masl with an annual rainfall of 510 mm to 1530
mm. “In Malawi, Miombo woodlands constitute
92.4% of the country’s total forested area, and
are mainly located in forest and game reserves
established for water catchment as well as for soil
and biodiversity conservation” [4]. Miombo forest
represent an important forest composition with a
wide distribution in several African countries with a
characteristic leafless period during the dry season.
In the sub-Saharan region, Miombo forests cover
approximately 17.3 million km² [5] and have large
distribution areas in Southern America. Up to now
the most Earth Observation (EO) based mapping
approaches do focus on humid tropical forest and
only few research results for tropical dry forest have
been presented.
The Malawian study site of dry tropical forest has an
area of ~32.000 km² with highlands in the northern
part, the Kaningina Forest Reserve in the middle,
Kasungu National Park in the south-west and
Nkotakhota Wildlife Reserve in the eastern part.
D. Peru
Situated in the north-western part of South America,
Peru covers an area of ~1.285.000km². The Andes
significantly affect the climate and ecosystems of
Peru. From the coastline on the West the elevation
rises to the peak of the Andes and declines towards
the Amazon Basin on the Eastern side. Eastern of
the Andes, the main forest types can be summarised
as lowland evergreen broad leaf rain forest, lower
montane forest and upper montane forest. The
western part of Peru has low precipitation, resulting
in Sclerophyllous dry forest and an overall sparse tree
cover.
211
Universidad Mayor
Forest Monitoring
The area size of the Peruvian sites is ~35.000km² and
mainly covers forests in the provinces San Martin
and Ucayali; both provinces are known as areas
where rapid deforestation has occurred. Recent
agro-industrial development in relation to the palm
oil and cocoa industry led to significant deforestation
events in these sites.
Forest Cover Mapping
EOMonDis developments are based on experience
from previous projects in the domain of tropical
forest ecosystem monitoring and will undergo a
permanent improvement process. Especially the
method for dense time series analysis, which has
been developed within the European Space Agency
(ESA) funded Global Monitoring for Environment
and Security Service Element for Forest Monitoring
(GSE FM) project, builds the starting point for the
processing approach. The developed method [6]
focuses on tropical dry forest with leaf-on and leafoff periods. Identification of forest is much more
challenging in these ecosystems due to scattered
distribution of trees, which causes mixed spectral
signatures in the high-resolution EO data domain,
hampering precise estimation of canopy cover per
area unit. Another drawback is the coincidence of
the vegetation period with the season of heaviest
cloud cover during the rainy season. Conditions for
optical image acquisition are much better during the
dry season, but the leaf-off state of deciduous trees
hinders an accurate determination of forested areas.
The entire project is realised by a two-phased
approach; the first phase comprises a requirement
assessment and improved algorithm development,
the second phase is dedicated for product
enhancements and inclusion of product feedback
that is gathered at the end of the first phase. With
the knowledge and information gained from
the user requirement assessment, the technical
developments are targeted towards the product
output specifications. These output specifications
define the EO input data that can be used to meet
pre-defined product requirements. The processing
chain therefore has to be adaptable to input data
of different type and capable of processing data
composites of varying spectral characteristics
within the dense time series (e.g. phenology, leafon/leaf-off). At the current stage of the project,
the automated processing chain is proven for the
212
handling of optical EO data, mainly Sentinel-2
and Landsat data. Having said this the subsequent
description applies to forest cover and change
mapping based on optical data. The monitoring
system is built to start with the baseline year 2010
and provides a continuous monitoring of forest cover
and forest cover changes (including IPCC relevant
Land Use and Land Use Change classes) from 2015
onwards using multi-sensoral and -temporal satellite
data, which are validated with VHR optical data.
E. EO Data Processing
Before entering into the detailed time series
analysis, common pre-processing procedures are
applied to all input data, namely a) Cloud detection
and masking b) Radiometric calibration in-between
scenes c) Calculation of Top of Atmosphere (ToA)
reflectance d) Geometric validation and e) Tasselled
Cap Transformation [7]. Initially, the data will be
segmented into image objects with similar spectral
properties to reduce data volume and counteract
alignment problems in this multi-temporal, multisource data scenario. In order to derive objects that
reflect all different states of the land cover over
the entire timeframe under consideration, each
individual scene will be investigated separately in a
sequential (sub-) segmentation procedure, starting
from a master image. The resulting segments
then function as a basis for the subsequent image
analysis.
Classification of segmented satellite images is
performed using a supervised approach with
K-Nearest Neighbour algorithm, independently for
each image of the time series.Training data is collected
by visual interpretation of very high-resolution
(VHR) imagery and high-resolution (HR) imagery. A
special focus is given to areas with persistent cloud
clover to guarantee a sufficient amount of reference
data for each scene. Suitable reference locations
are these ones, which do not undergo any changes
during the period under consideration. These stable
locations are used to extract spectral information for
each scene individually at object level to train the
classificator, image by image, to perform a sequence
of land cover classifications. The information content
of the chronologically ordered satellite images is
then used to a) improve accuracy of each individual
classification, b) replace NoData values at a specific
point in time by applying logical rule sets and c)
determine the location and date of change events.
Universidad Mayor
Dense Time Series Analyses
The methodology is based on rating independent
classification results in their temporal context,
taking account of quality aspects of the specific
EO data source and predicted, standardized landcover-specific dynamics. Prior to the time series
analysis, all input scenes are rated according to the
time of acquisition (leaf-on or leaf-off-season) and
atmospheric conditions (clouds/shadows/haze).
Both factors do influence the accuracy of the forest
cover classification. Thus, results from a classification
performed on a satellite image acquired during
leaf-on stage and under clear sky are assumed to
be more reliable than results from an image taken
under interfered conditions. Contrary to the a priori
reliability, the continuity of classification results of
a specific object is used to calculate context driven
reliability. A moving window, that respects five
successive classification results, is used to compare
each class assignment with the four adjacent ones in
the timeline. The level of agreement is calculated for
each segment. The value, denoted as class reliability,
considers the consistency of classification results
along the time axes. Additionally, a plausibility
rating is introduced to consider natural behaviour of
forest ecosystems such as time needed until canopy
cover is again detectable in satellite imagery and the
phenology of trees.
F.
Based on these indicators, the algorithm iteratively
optimises the individual classification results by
maximising the plausibility of the entire code
sequence. The procedure results in logically
consistent code sequences that correspond to an
expected behaviour of – or transition between - the
land cover under consideration.
Product Verification Approach
Mapping results have to be validated with statistical
sound methods in order to fulfil the stakeholder
requirements and enable an unbiased judgement
of product quality. There is no doubt, that accuracy
assessments shall be always performed with an
independent data set. Generation of the reference
data to evaluate the quality of single map products
can be produced either by field data collection or
by visual interpretation of very high-resolution
imagery. Satellite and airborne imagery have
the great advantage of covering large areas in a
continuous manner. Field based observation are
most often sample based with a limited extend of
G.
ForestSAT 2016 Abstracts Summary
coverage. Especially for remote areas, field based
reference data collection is not feasible regarding
time, cost and efficiency. Recommended land use
classes of the IPCC can be reliably interpreted
on VHR data. Within EOMonDis, the concept of
stratified random sampling with primary sampling
units (PSU) and secondary sampling units (SSU) is
adopted. Combining a random sampling approach
with a systematic grid ensures that the whole area
is sampled [8]. The method can be used to generate
reference samples for classification and reference
samples for accuracy assessments. For this purpose,
VHR data should be divided in a northern and
southern part to guarantee independent data sets
for classification and validation.
Figure 1 : Example of the sampling design for the
selection of primary sampling units (PSU) and
secondary sampling units (SSU). A stratification
across the PSU can be performed to cover different
intensities of deforestation. For each stratum, one
PSU per grid cell would be used.
Figure 1 shows the conceptional design of this
sampling procedure for a grid with 20km, an area
of 25 km² for the PSU and random points within a
selected PSU, whereas the points are the SSUs. The
design is flexible and adaptable to local conditions.
In order to accurately estimate the accuracy of
mapping results, a stratification of PSU locations
is performed according to their risk of change. The
stratification is performed by incorporating global
forest change (GFC) data, which contains the yearly
change of forests at a spatial resolution of 30m
[9]. It is most likely, that new changes will occur in
the proximity of former deforestation events. The
overall area of changes is calculated for each grid
cell and for each PSU. Afterwards the PSUs of each
grid cell are equally divided into PSU of high risk of
change and PSU of low risk of change
213
Universidad Mayor
Forest Monitoring
H. Precusor Study Malawi
Within the Malawian test site, a precursor study
on the rule based classification methodology was
conducted by incorporating Landsat and SPOT EO
data in a dense time series analysis with the former
described methods. Details of the approach for land
cover and land cover change mapping are described
in [6]. The main results are presented within this
section to illustrate the applicability of the method,
which was tested for tropical dry forest. A 10%
threshold for canopy cover, a minimum mapping
unit of 0.5 hectare and maturity height of 5 m was
chosen as forest definition, which is in line with the
Food and Agriculture Organisation (FAO) definition
of dry forest.
EO data for classification was comprised of 376
Landsat and 28 SPOT5 scenes, covering a time
period from the year 2000 until 2014. All EO data
was individually pre-processed and classified.
Forest losses were classified into one of the
Intergovernmental Panel on Climate Change (IPCC)
compliant land cover classes: Cropland, Grassland,
Settlement, Wetland, Water and Other Land Use.
The accuracy assessment was performed with 542
sample points, which were generated by a posteriori
stratification and selection of 12 PSUs. Thus, the
sample selection is unbiased and independent from
reference data that was used for classification.
214
Figure 2 shows a subset of the result from change
date mapping. The classified dense time series that
covers the period in between the reference dates
for the change map enable the determination of
narrow time windows for each specific change
event, an important requirement to specify yearly
deforestation rates. Beyond that, the existence of a
continuous, consistent classified time series allows
for filling most data gaps – e.g. related to clouds – in
case of stable land cover around a map’s reference
dates, testified by classification results of adjacent
scenes in the timeline. With an overall accuracy of
89.11% for the year 2000 and 86.16% for the year
2014, the entire processing chain is suitable to meet
required standards. The low accuracy of change
detection (62.28%) can be explained by the use of
IPCC categories, a binary evaluation of changes
would result in higher accuracies.
I. Conclusion
The EOMonDis project is addressing the need for
user tailored, precise and validated forest monitoring
products. The basis for method development are
the stakeholder requirements from the climate
policy, private sector companies in the domain of
deforestation free supply chains and governmental
institutions at country level that need reliable
information on their forest resources and changes
thereof. The integration of Sentinel-2 and Sentinel-1
Figure 2 Subset of the Malawian test site.
Upper row: Reference data for change detection (Landsat5, RGB = NIR, SWIR, RED).
Historic Image with cloud/shadow data gaps
(left) and reference date 2015 (right) Lower
left: Completed forest change map between
the reference dates (upper row), red colour
indicates deforestation. Lower right: Deforestation dates derived from EO time series 2000
- 2014 (legend below).
Universidad Mayor
data into dense time series builds the solid basis
of the monitoring concept. The variety of forest
ecosystems, the knowledge based classification
procedure with adaptable rule sets and the validation
approach will demonstrate the capabilities of EO
data for tropical forest monitoring, accompanied by
documented quality assurance.
ForestSAT 2016 Abstracts Summary
[5]
[6]
References
[1]
[2]
[3]
[4]
HCS Approach Steering Group, 2015, The HCS
approach Toolkit – The High Carbon Stock
Approach: No Deforestation in Practice. Version
1.0, Kuala Lumpur. HCS Approach Steering
Group.
GOFC-GOLD, 2015, A sourcebook of methods
and procedures for monitoring and reporting
anthropogenic greenhouse gas emissions and
removals associated with deforestation, gains
and losses of carbon stocks in forests remaining
forests, and forestation. GOFC-GOLD Report
version COP21-1, (GOFC-GOLD Land Cover
Project Office, Wageningen University, The
Netherlands).
UNFCCC, 2009, Cost of implementing
methodologies and monitoring systems relating
to estimates of emissions from deforestation and
forest degradation, the assessment of carbon
stocks and greenhouse gas emissions from
changes in forest cover, and the enhancement
of forest carbon stocks, Technical Paper FCCC/
TP/2009/1. UN Framework Convention on
Climate Change.
Kachamba, D.J.; Eid, T.; Gobakken, T., 2016,
Above- and Belowground Biomass Models for
Trees in the Miombo Woodlands of Malawi.
Forests 2016, 7, 38.
[7]
[8]
[9]
Chidumayo, E. & Marunda, C., 2010, Defining Dry
Forests and Woodlands of Sub-Saharan Africa.
In: Chidumayo, E.N. and Gumbo, D.J. (eds) The
dry forests and woodlands of Africa: managing
for products and services, p. 1. Earthscan Library,
London. ISBN 978-1-84971-131-9
Storch, C., Wagner, T., Ramminger, G., Pape, M.,
Ott, H.,Häusler, T., Gomez S., 2016, Automatic
derivation of forest cover and forest cover change
using dense multi-temporal time series data
from Landsat and SPOT5TAKE5. Proceedings of
ESA Living Planet Symposium SP-740, 09-13 May
2016, Prague Czech Republic.
Kauth, R.J. & Thomas, G.S., 1976, The Tasselled
Cap -- A Graphic Description of the SpectralTemporal Development of Agricultural Crops
as Seen by LANDSAT. Procs. Symposium
on Machine Processing of Remotely Sensed
Data, West Lafayette, Indiana, 29 June -1 July
1976 (West Lafayette, Indiana: LARS, Purdue
University), 41-51.
Sannier C., McRoberts R.E., Fichet L-V, Massard K.
Makaga E., 2014, Using the regression estimator
with Landsat data to estimate proportion forest
cover and net proportion deforestation in Gabon,
Remote Sensing of Environment, Volume 151,
Pages 138-148.
Hansen M. C., Potapov, P. V., Moore, R., Hancher,
M., Turubanova, S. A.,Tyukavina, A., Thau, D.,
Stehman, S. V., Goetz, S. J., Loveland, T. R.,
Kommareddy, A., Egorov, A., Chini, L., Justice,
C. O., Townshend J. R., 2013, High-resolution
global maps of 21st-century forest cover
change. Science 342, 850–853. doi:10.1126/
science.1244693.
215
Universidad Mayor
Forest Monitoring
POSTER
Change Detection in Multitemporal SAR Orthoimages
Rafael A. S. Rosa1,2, David Fernandes2, João B. Nogueira Jr.3 , Karlus A. C. Macedo1, Juliano Lázaro1, Dieter Lübeck1
1
Bradar Indústria S.A.
Instituto Tecnológico de Aeronáutica – ITA
3
Santo Antônio Energia S.A.
2
Keywords: change detection, multitemporal images, synthetic aperture radar, SAR.
Forest monitoring is a major concern today due
to climate changes, conservation of fauna and
flora and to the lack of water. Therefore, several
environmental monitoring techniques have been
developed and used to detect changes in the scenes.
However, the main techniques have limitations
related to weather conditions and does not have the
expected effect. The use of SAR (Synthetic Aperture
Radar) seems appropriate to detect changes due
to its independence of atmospheric and lighting
conditions. The SAR change detection is a process
that uses SAR images acquired in the same geometric
conditions but in different moment (multitemporal)
to identify changes in the surface that occurred
between two acquisitions. This work presents a new
method of change detection in multitemporal SAR
orthoimages using the difference between them.
Experimental tests were conducted using real SAR
data obtained by the airborne sensor OrbiSAR-2
from Bradar in the Amazon Forest and the results
showed a higher accuracy than those found in the
literature.
One of the most important topics today is the
climate change, and one of its main possible cause
is the deforestation of rainforests around the world.
Just talking about the Amazon Forest, the last years
deforestation average has been 5,000 km²/year.
This scenario makes more and more organizations
and institutions search surveillance methods for
the reduction and prevention of deforestation. The
most important Brazilian programs for this purpose
are PRODES (Amazon Deforestation Monitoring
Program) – the world’s biggest forest monitoring
program – and DETER (Real Time Deforestation
Detection), both from INPE – National Institute
for Space Research, doing the monitoring
through satellite imagery. These programs have
216
demonstrated their relevance, however they have
two limitations due to their sensors: the resolution
(15m) and the inability to map regions covered by
clouds. At the beginning of these programs, there
was a reduction of the Amazon deforestation rate,
however after their limitations become known
by the deforesters this rate has grown by using
invisible techniques to that sensors: selective
logging (isolated and scattered lumbering in order
to hinder the identification); and the deforestation
in rainy periods, when optical imaging sensors
have restrictions due to clouds. That is why
change detection by radar imaging, mainly SAR
(Synthetic Aperture Radar), seems appropriate to
equatorial and tropical forest monitoring due to its
independence of weather conditions, without losing
the high resolution. The purpose of this study was to
develop an algorithm able to automatically detect
changes in the surface in a time interval using SAR
orthoimages from temporally spaced acquisitions,
however with identical geometry (multitemporal).
The goal was to create an efficient tool able to
automatically identify regions where there was
some kind of change, such as appearance of gaps in
vegetation areas, trails, changes at forests borders,
selective logging, growth of pastures, plantations
and other dynamic land use. The following figures
show examples of SAR orthoimages from the Bradar
OrbiSAR-2 airborne sensor, obtained in the region
of Porto Velho (Brazil) in 2013, provided by the
Santo Antônio Energia S/A, with change indications
through the RGB composition. Figure 1 shows three
images: (a) and (b) are SAR orthoimages obtained
with 1-month interval between them; and (c) shows
an RGB composition of these SAR orthoimages,
wherein the R-band is composed of the older SAR
orthoimage (a) and the G- and B-bands are equal
and have the latest SAR orthoimage (b), in this way,
Universidad Mayor
ForestSAT 2016 Abstracts Summary
the red regions are areas where there was less return
of the radar signal in the second than in the first
acquisition, in other words, they are areas where
the targets were “disappeared” (cut trees), the cyan
regions are areas where the opposite occurred,
i.e., they are areas in which the return of the radar
signal increased due to the “appearance” of targets
(growth of pasture) and the gray regions are areas
where there has been no significant change.
Figure 2 shows the SAR ability to detect the selective
logging: this figure is also an RGB composition
of two SAR orthoimages acquired with 1-month
interval between them, generated with the same
principle of the image (c) of Figure 1, this means
that, the red regions represents the cut trees. In
this image, the cyan regions are trees that were
hidden in the first imaging and were exposed in
the second. Figure 3 (a) shows an RGB composition
(a)
(c)
(b)
Figure 1. SAR orthoimages obtained by OrbiSAR-2 sensor at 2012/09/23 (a) and 2012/10/20 (b); and the
respective RGB composition (c).
Figure 2. RGB composition of multitemporal SAR orthoimages showing the selective logging in red and the
appearance of targets in cyan.
217
Forest Monitoring
Universidad Mayor
(a)
(b)
Figure 3. RGB composition of multitemporal SAR orthoimages showing a deforestation in the region
delimited by the yellow contour (a) and corresponding aerial photograph of this deforested area (b).
of two multitemporal acquisitions identifying a
deforesting on the Madeira River banks and (b) is
an aerial photography of this region obtained after
the second of these acquisitions, confirming the
detection.
218
The objective of this work was to create an algorithm
that, from multitemporal SAR orthoimages, is
able to identify changes between them and to
produce binary masks that can be used for creating
cartographic vectors.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Changing northern vegetation conditions are influencing
barren ground caribou (Rangifer tarandus groenlandicus)
behavior
1st: Gregory J.M. Rickbeil1 2nd: Txomin Hermosilla1 3rd: Nicholas C. Coops1 4th: Joanne C. White2
5th: Michael A. Wulder2 6th: Trevor C. Lantz3
2
1
Faculty of Forestry, University of British Columbia, Vancouver, BC, Canada, V6T 1Z4
Canadian Forest Service (Pacific Forestry Centre), Natural Resources Canada, Victoria, British Columbia,
Canada, V8Z 1M5
3
School of Environmental Studies, University of Victoria, BC, Canada, V8W 2Y2
Keywords: Arctic, EVI, GPS, Landsat, mammal, movement, productivity, telemetry
Abstract
Aim: To quantify changes in vegetation productivity over the past three decades across five barren
ground caribou (Rangifer tarandus groenlandicus) herd ranges and assess how these changes are
influencing caribou behavior.
Location: Northwest Territories and Nunavut, Canada
Methods: As an indicator of vegetation productivity, the enhanced vegetation index (EVI) was
calculated on newly-developed cloud free, gap free, Landsat surface reflectance image composites
representing 1984 to 2012. Changes in EVI were assessed on a pixel basis using Theil-Sen’s nonparametric regression and compared across herd ranges and land cover types using generalized
least squares regression. Animal movement velocity and turning angle were calculated from caribou
telemetry data and generalized additive mixed models were used to link these behavior metrics with
changes in vegetation productivity.
Results: Vegetation productivity increased across the five caribou herd ranges examined. The largest
productivity increase occurred over the ranges of the most western herds, generally in tundra and
shrub habitats. Results indicate that caribou tended to move more slowly and turn at larger angles
in tundra habitats with large increases in productivity. Caribou tended to move more slowly through
shrub habitats and more quickly through forest habitats with increased levels of productivity; however,
caribou did not respond to productivity changes in shrub and forest habitats in terms of turning angles.
Main Conclusions: Over the 3 decades of collected data, barren ground caribou habitats have become
more productive, which is consistent with other studies that have documented increases in Arctic
vegetation productivity. In particular, increasing levels of productivity in tundra habitats may lead to
improved foraging opportunities across barren ground caribou calving and summer ranges. However,
shrub habitats which experienced large gains in productivity were associated with decreased foraging
behavior, indicating that shrub proliferation may offset the potential benefits of increasing tundra
productivity.
219
Universidad Mayor
Forest Monitoring
POSTER
Characterization of the wildland-urban interface using
LiDAR data and OBIA as a tool for fire risk prevention and
management at a local scale
Miguel A. Rodríguez-Garrido
José Carlos García-López
Flor Alvarez-Taboada*
*corresponding author. flor.alvarez@unileon.es
Affiliation of the authors:
Miguel A. Rodríguez-Garrido and Flor Alvarez-Taboada : GEOINCA research group. Department of Mining Technology,
Topography and Structures, University of León, Ponferrada Campus, Avenida de Astorga s/n, 24400, Ponferrada, León,
Spain.
José Carlos García López: Head of the Section for Nature Protection (León). Regional Government of Castilla y León, Avd.
Peregrinos s/n, 24008, León, Spain.
Key words: LiDAR, woodlands, buildings, forest fire, OBIA
Abstract
The wildland-urban interface (WUI) is the area where buildings are in contact with the forest. Nowadays
it is one of the biggest problems regarding wildfire suppression. Thus, establishing measures and
proposals to reduce the fire risk in those areas is needed, so it is required to know its location and
extension. The main goals in this study are: (i) to identify and map the intervention areas to establish
preventive measures against wildfires fires in the WUI, (ii) the classification and mapping of the
vegetated areas according to the foreseeable fire behaviour, based on LiDAR data and (iii) to classify
the constructions in the WUI depending on the damage risk in case of a wildfire.
The methodology was structured in 3 parts: (i) processing of the LiDAR data, (ii) classification and
mapping of the vegetated areas according to the foreseeable fire behaviour in case of a wildfire, and
(iii) characterization of the WUI for fire risk prevention and management at a local scale, where the
buildings and infrastructures were classified regarding their damage risk.
The use of the normalized digital surface model derived from the LiDAR data and the RGB
orthophotographs in the object based image analysis was suitable to identify fuel models defined
by the canopy cover fraction and the woody vegetation height, as it stated the overall accuracy of
87.92% (kappa: 0.8554), obtained for the classification of the vegetation by height. It was possible to
identify and quantify the extension of the three levels of risk of damage due to a wildfire for the three
preventive intervention areas. It was concluded that LiDAR, orthophotographs and the techniques
used in this study are appropriate for the characterization of the WUI, critical for fire risk prevention
and management at a local scale.
Introduction
The wildland-urban interface (WUI) is defined as
the area where buildings are in contact with the
forest. Nowadays it is one of the biggest problems
220
regarding wildfire suppression. Thus, establishing
measures and proposals to reduce the fire risk in
those areas is needed, so it is required to know its
location and extension. Galicia (NW Spain) was the
Spanish region most affected by wildfires from 2001-
Universidad Mayor
2010, with 72,423 fires, 68 of them were classified
as large forest fires (more than 500 ha). In addition
to the frequent wildfires in this area, the scattered
distribution of the settlements and the large number
of buildings located in the wildland-urban interface,
brings to light the need to identify the buildings
which could be affected if there was a wildfire and
to quantify their risk of damage in regards to their
location.
The main goals in this study are: (i) to identify and
map the intervention areas to establish preventive
measures against wildfires fires in the WUI, (ii) the
classification and mapping of the vegetated areas
according to the foreseeable fire behaviour, based
on LiDAR data and (iii) to classify the constructions
in the WUI depending on the damage risk in case of
a wildfire.
Methodology
The study area covers 16 km2 in the surroundings of
Pedrafita do Cebreiro (Lugo, NW Spain). Data from
the Aerial Orthophotograph National Plan (PNOA)
(LiDAR and RGB orthophotographs), cadastral
cartography and ground truth points for the
validation of the filtering process and for the digital
terrain model (DTM) interpolation were used in this
project. The LiDAR dataset consisted of 0.5 point/
m2 data from two different flights: 2009 (Region of
Galicia) and 2010 (Region of Castilla y León), while
the RGB orthophotographs where gathered in 2009.
The methodology was structured in 3 parts (Figure
1): (i) processing of the LiDAR data, (ii) classification
and mapping of the vegetated areas according to
the foreseeable fire behaviour in case of a wildfire,
and (iii) characterization of the WUI for fire risk
prevention and management at a local scale, where
the buildings and infrastructures were classified
regarding their damage risk.
LiDAR data Processing
This first phase includes all the processes required
to obtain a normalized digital surface model
(nDSM). It involved a visual analysis for outlier
removal; a filtering phase to remove all the points
which do not belong to the terrain (using a linear
prediction algorithm (Kraus & Pfeifer, 1998)); and
the interpolation of these filtered data using a 2 m
x 2 m cell size to develop a DTM. The results of the
filtering process where were validated using ground
ForestSAT 2016 Abstracts Summary
truth data verified on the images, trying to minimize
the omission and commission errors, and trying to
keep the same filtering parameters for both flights.
The DTM was validated using data measured in
the field by GNSS receivers. The criteria to choose
the most suitable combination of parameters for
the interpolation were: (i) qualitative analysis, (ii)
minimizing the 95% error quantile, (iii) minimizing
the 68.3% error quantile and (iv) minimizig RMSE.
All these steps were carried out by using FUSION
tools (McGaughey, 2015) and R. Finally, a 2 m x 2
m nDSM was derived, and it was validated with a
random sample of vertical profiles.
Classification and mapping of the vegetated
areas according to the foreseeable fire
behaviour in case of a wildfire.
The second phase required the definition of 5 fire
behaviour fuel models, based on the height and the
canopy cover of the woody vegetation (Table 1).
The risk of damage for buildings in case of a forest
fire was defined for each fire behaviour fuel models
(based on the intensity and height of the flames and
the propagation speed).
Table 1. Fire behaviour fuel models, based on the
height and the canopy cover of the woody vegetation
(shrubs and trees).
Canopy cover < 2 m
< 20 %
Model 1
20 % - 60 % Model 2
> 60 %
Model 4
Height
2-8m
Model 1
Model 2
Model 4
>8m
Model 1
Model 3
Model 5
The classification was performed by following an
object based image analysis, with two levels of
segmentation and a non-parametric classification
of the vegetation based on its height. The input
data were the RGB images, the green ratio and the
nDMS. The process is showed in detail in Figure
1. To validate the results of this classification, an
accuracy assessment was performed using data
from a random sampling and the global accuracy,
omission error, commission error and kappa were
calculated. Subsequently, the canopy cover fraction
was calculated and the vegetation was classified
according to 5 fire behaviour fuel models, which also
221
Universidad Mayor
Forest Monitoring
"$ ! $ "!#
" ##
!' &""$
,20+
509-
"
$
$
$# &2
###
30
&1 $#
!' &""$
,;509-
#!
$$
300
!' &""$
,:209-
!%$$
!' &"
"$
&2 $#
$$
$$
,:0)4-
##$
!%$$
!' &""$
'
&$$
##$
'&$$
,0)4+
2-
'&$$
,2+
6-
'&$$
,;6-
"$"($
"& %"
% #
"$"($ $
%#*
#$"%$ #
#%$#
Figure 1. General workflow.
allowed determining the risk of damage due to a
wildfire.
Characterization of the WUI for fire risk
prevention and management at a local scale.
In this phase, each building and infrastructure was
classified regarding their damage risk, taking into
account Table 2. The risk assigned was the highest
for the vegetation in a radius of 25 m. Moreover,
and taking into account different laws, the risk in
a radius of 100 m and 400 m were also computed.
Those three areas were designated as preventive
intervention areas. Once the the risk in the preventive
intervention areas was identified and analyzed, then
the extension of each risk level (Low, High and Very
high) was quantified at two different scales: for the
study area and for every building/construction.
Results
The results showed that the most accurate 2 m x
2m DTM was generated using the LiDAR flight data
of Castilla y León, removing the outliers, with the
filtering parameter values of w=2 and g=-1.5 and
the switch ‘minimum’ in the interpolation process.
This DTM reached an accuracy of 1.36 m for the 95%
quantile, a root-mean-square error (RMSE) of 0.699
m and a precision= 1.378 m.
The use of the normalized digital surface model
derived from the LiDAR data and the RGB
orthophotographs in the object based image analysis
was suitable to identify fuel models defined by the
canopy cover fraction and the woody vegetation
height, as it stated the overall accuracy of 87.92%
Table 2. Area covered by each fuel model.
Fuel Model
1
2
3
4
5
222
Area (ha)
Area with woody
vegetation (%)
Total área (%)
Risk of damage due
to a wildfire
15.38
77.97
0.97
822.28
193.27
1.39
7.03
0.09
74.09
17.41
0.96
4.87
0.06
51.39
12.08
Low
High
Low
Very high
Low
Universidad Mayor
(83.70%-91.18%) (kappa: 0.8554), obtained for the
classification of the vegetation by. The five fuel
models were directly associated to three risk levels
in the wildland-urban interface, which allowed to
identify and locate the areas affected by the different
levels of risk due to fire damage (Table 2). The fuel
model number 4 (scrubland and dense young forest)
was the one that covered the largest area (882.28
ha), which means that the high risk of fire damage in
constructions affects 51.39% of the study area.
It was possible to identify and quantify the extension
of the three levels of risk of damage due to a wildfire
for the three preventive intervention areas. For the
preventive intervention area of 25 m, 3.68 ha were
identified and mapped as high risk areas and 9.10
ha as very high risk areas (1.33% of the study area).
Related with the characterization of the WUI for fire
risk prevention and management at a local scale,
more than 99% of the constructions (262) were at
high risk or very high risk. To decrease this risk in the
preventive intervention area of 25 m, 12.79 ha were
ForestSAT 2016 Abstracts Summary
identified and mapped and in that area the canopy
cover fraction is advised to be reduced by using
preventive silviculture, in order to reach a canopy
cover fraction lower than 20%, with the aim of
obtaining vertical and horizontal fuel discontinuity.
Conclusions
It was concluded that LiDAR, orthophotographs and
the techniques used in this study are appropriate for
the characterization of the WUI, critical for fire risk
prevention and management at a local scale.
References
[1] McGAUGHEY, R.J. FUSION/LDV: Software for
LiDAR data analysis and visualitation, 2015.
United States: Department of Agriculture.
[2] KRAUS, K., PFEIFER, N. Determination of terrain
models in wooded areas with airbone laser scanner
data. En: ISPRS Journal of Photogrammetry and
Remote Sensing, 1998, vol 53, pag 193-203.
223
Universidad Mayor
Forest Monitoring
Combining Sentinel-1, Sentinel-2 and Landsat 8 images for
near-real time forest change detection
Ruth Sonnenschein1, Claudia Notarnicola1, Carlo Marin1 & Malcolm Davidson2
EURAC, Institute for Applied Remote Sensing, Bozen/Bolzano, Italy
2
European Space Agency, ESTEC, Nordwijk, The Netherlands
1
Keywords: Forest change detection, Sentinel-1, Sentinel-2, Landsat-8, near-real time monitoring,
continuous change detection
Abstract
Earth observation has become the most important instrument for monitoring forest cover dynamics at
high spatial resolution and across large areas. A large variety of change detection approaches has been
developed which have been moved from bi-temporal change detection to annual change detection
and continuous change detection approaches (Zhu et al., 2012, Zhu and Woodcock, 2014, Zhu et al.,
2015) mainly driven by the availability of free and consistent optical Landsat images. Despite the need
of timely alerts about illegal logging activities or assessments of the spatial extent and impacts of
disturbances caused by natural hazards, near-real time monitoring of forest disturbances has been
restricted by the relatively low temporal resolution of the Landsat sensor and the presence of cloud
cover. Fusion concepts combining Landsat and the daily MODIS images overcome the low temporal
resolution but may miss small-scaled disturbances and furthermore, face similar limitations in areas
with persistent cloud cover. Radar images allow forest change mapping in such situations but data
access has been mainly restricted until now.
With the launch of ESAs Sentinel missions, earth observation monitoring capabilities have drastically
increased by a series of satellite constellations offering complementary technologies and a free and
open data access policy. Especially the Sentinel-1 C-band radar sensor and the Sentinel-2 multi-spectral
optical sensor with a high spatial and temporal resolution are designed to provide detailed information
on land changes over time. Yet, it remains unclear how the near-real time monitoring of forest
disturbances benefits from the two Sentinel satellites, the synergies between them and the synergies
with the Landsat-8 sensor. Our overall objective was to assess how accurate and timely forest changes
can be detected by using a multi-source concept. Based on previous research activities, we further
developed the continuous change detection approach (Zhu et al., 2012) and present a methodology
that allows detecting forest changes using each Sentinel-1 or Sentinel-2/Landsat-8 image. We fitted
band-wise time-series models to each pixel and data source through time considering data availability
and temporal variability of the forest ecosystem. We then predicted spectral and backscattering
values for each new acquisition date and flagged those pixels which considerable differed from the
actual values using the modeling period as reference. By integrating flagged pixels over time and
across Sentinel-1 and Sentinel-2 sensors, we finally labeled pixels as forest change. We validated our
results by simultaneously interpreting temporal profiles of spectral and backscattering values and
image chips following the concept of TimeSync (Cohen et al., 2010). We present a case study to show
the benefits and limitations of our multi-source approach together with estimates on how reliable the
single flags and their integrations over time are for accurate change alerting and change detection.
224
Universidad Mayor
References:
Cohen, W. B., Yang Z., Kennedy, R. (2010): Detecting
trends in forest disturbance and recovery using
yearly Landsat time series: 2. TimeSync — Tools
for calibration and validation. Remote Sensing of
Environment, 114, 2911-2924.
Zhu, Z., Woodcock, C.E. and Olofsson, P. (2012):
Continuous monitoring of forest disturbance
using all available Landsat imagery. Remote
ForestSAT 2016 Abstracts Summary
Sensing of Environment, 122, 75-91.
Zhu, Z. and Woodcock, C.E. (2014): Continuous
change detection and classification of land cover
using all available Landsat data. Remote Sensing
of Environment, 144, 152-171.
Zhu, Z., Woodcock, C.E., Holden, C. and Zhiqiang,
Y. (2015): Generating synthetic Landsat images
based on all available Landsat data: Predicting
Landsat surface reflectance at any time. Remote
Sensing of Environment, 162, 67-83.
225
Universidad Mayor
Forest Monitoring
Developing a U.S. national land use and land cover reference
data set to support inter-agency mapping, validation and
statistical estimation needs
Todd A. Schroeder1, Tom R. Loveland2, Bruce Pengra3, Warren B. Cohen4, Sean P. Healey5,
Mark Finco6, Steve V. Stehman7, and Zhiqiang Yang8
1
USDA Forest Service, Southern Research Station, Knoxville, TN USA
U.S. Geological Survey (USGS), Earth Resources Observation and Science (EROS) Center, Sioux Falls, SD USA
3
Stinger Ghaffarian Technologies/USGS Earth Resources Observation and Science (EROS) Center, Sioux Falls, SD USA
4
USDA Forest Service, Pacific Northwest Research Station, Corvallis, OR USA
5
USDA Forest Service, Rocky Mountain Research Station, Ogden, UT USA
6
RedCastle Resources/USDA Forest Service, Geospatial Technology Applications Center (GTAC), Salt Lake City, UT USA
7
College of Environmental Science and Forestry, State University of New York (SUNY-ESF), Syracuse, NY USA
8
Department of Forest Ecosystems & Society, Oregon State University, Corvallis, OR USA
2
Keywords: TimeSync visualization tool, Landsat time series, Reference data, Land use and land cover
change, Forest disturbance, Design-based statistical estimation, Map validation
Given the recent influx of moderate resolution
imagery from satellites such as Landsat and
Sentinel-2, remote sensing has become an
effective and affordable tool for monitoring forest
dynamics and land use and land cover change
over large geographic areas. With data constraints
now minimized, many government agencies are
turning to new remote sensing approaches to
help meet specific mission objectives, such as to
produce and deliver maps which resolve current
and historical characteristics of landscape condition
and change. Although the products developed
by different agencies often vary in scope and
attributes of interest, one commonality is the need
for high quality reference data which can support
image classification, map validation and statistical
estimation activities. With increased focus on using
the full temporal depth of the available satellite
archive, the analyst interpretation approach has
surfaced as one of the best and only means of
collecting reference data which spans the full
range (20-40 years) and interval (annual) of most
time series applications. Visualization tools such as
TimeSync (Cohen et al., 2010) have facilitated the
collection of analyst interpretations of disturbance
and other land use and land cover variables over
space and time, yet the costs associated with
measuring a large number of plots (~25,000) remains
high. In an effort to mitigate costs and reduce
redundancy, the U.S. Geological Survey (USGS) and
226
the U.S. Forest Service (USFS) have recently formed
a collaborative partnership to mutually collect a
U.S.-wide, nationally consistent reference data set
which can meet both individual and shared agency
objectives. In this talk we will discuss various facets
of this operational data collection effort, including
development of a TimeSync-based, joint response
design which meets overlapping, but divergent
programmatic needs, as well as implementation of a
flexible sampling design, which promotes statistical
estimation of land use and land cover variables, while
also allowing for future sample intensification and
addition of new partnering agencies. To shed light
on the consistency of data collected by different
agencies and interpreters, as well as to examine
the utility of the collected variables for assessing
map accuracy, results from two initial pilot studies
conducted in the Pacific Northwest and Great Lakes
regions will be presented. In conclusion, we will
assess the future benefits and challenges associated
with expanding inter-agency collaboration in the era
of rapidly evolving remote sensing technologies.
References
Cohen, W.B., Z. Yang, and R.E. Kennedy. 2010.
Detecting trends in forest disturbance and
recovery using yearly
Landsat time series: 2. TimeSync - Tools for
calibration and validation, Remote Sensing of
Environment 114:2911-2924.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Development of a UAV based platform for monitoring
simulated disease expression using time-series airborne laser
scanning and high resolution multi-spectral imagery
Jonathan P Dash1*, Marie Heaphy1, Toby Stovold1, Michael S Watt2, Heidi Dungey1.
Scion, PO Box 3020, Rotorua, New Zealand
Scion, PO Box 29237, Fendalton, Christchurch, New Zealand
1
2
Keywords: UAV, Pinus radiata, ALS, multispectral imagery, time-series
Forest products are a vital component of the
global economy and play a vital role in climate
regulation and the provision of ecosystem services.
In New Zealand the plantation forest industry is the
country’s third largest export earner with exports
worth more than $5 billion per annum. Plantation
forestry is dominated by the fast-growing conifer
species Pinus radiata which occupies 90 % of the total
plantation area. Such reliance on a single species
means that the industry is particularly vulnerable to
biosecurity incursions and plant pathogens. In New
Zealand Pinus radiata is affected by a number of
pathogens including Dothistroma pini, Cyclaneusma
minus and Phytophtora pluvialis and these result in
a significant loss in forest productivity. In light of
this a range of remote sensing technologies are
required for monitoring the spread of endemic
infections and detecting outbreaks so that
control measures can be implemented. Previous
research has shown that large-scale outbreaks
of certain foliar pathogens can be detected from
high resolution satellite imagery and imagery
from manned aircraft. However, detection in this
manner may be inadequate for early detection of
outbreaks, identification of small outbreaks, and
may be too expensive for assessment on a regular
basis. Unmanned aerial vehicles (UAV) potentially
provide a platform that provides very high
resolution data acquisition with a regular return
frequency in a cost-effective manner. With the
development of appropriate analytics and suitable
data acquisition settings this can potentially play
an important role in supporting the forest industry
and forest research.
Disease outbreaks in a biological system are
challenging to predict and so acquiring a timeseries dataset that encompasses the early stages of
disease expression are difficult to obtain. To address
this a trial was initiated where a disease outbreak
was simulated through tree poisoning. Poisoning
was intended to invoke changes in needle colour
and retention that might be consistent with disease
expression. Within a 9 ha trial plots were assigned
to 5 different treatments, including a control, and
for each treatment there were 5 replicates. The
treatment defined the number of trees poisoned
within each plot and all trees were poisoned in the
same manner. The trial was regularly assessed from
the ground using conventional tree health scoring
by an experienced technician and from a UAV
using high density airborne laser scanning (ALS)
and 5 narrow band multi-spectral imagery with a 6
cm ground surface distance (GSD). Multi-spectral
images were converted from digital numbers to
reflectance values using a reflectance panel with
known spectral properties and time-series images
were converted to like-values through identification
of spectrally invariant targets in consecutive
images and adjustment using a technique based on
robust regression. A series of spectral and textural
indices were calculated from the imagery and were
available as indicators of progressive discoloration.
High density ALS data were processed to provide a
point cloud representing the forest canopy and then
described using a number of metrics including height
percentiles, distribution of height percentiles, ratio
metrics and penetration metrics. Metrics describing
the point cloud were designed to detect changes in
227
Forest Monitoring
defoliation. Rasters detailing the remotely sensed
data for the trial were resampled to a range of
resolutions to provide insight into the utility of lower
resolution data to detect the simulated disease
expression.
Using data from the highest resolution datasets
a random forest model was fitted with the ground
assessment as the response variable and all
candidate explanatory variables derived from the
UAV data. This was used to identify predictors that
were most important for describing discolouration
and defoliation. Using the rasters of the important
predictors a rule set was developed using object based
image analysis (OBIA) software to segment healthy
and poisoned trees in each image. This allowed the
time elapsed since poisoning for detection to be
quantified using UAV data and data from traditional
ground assessment. A comparison of detection
rates across the different sized clusters allowed the
detection thresholds to be identified. A comparison
of rasters at various resolutions provided insight into
the relationship between image resolution and the
size of the outbreak that could be detected using the
UAV data.
228
Universidad Mayor
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Dominant tree species dynamics informed by 30 years of
Landsat time series in mountain areas of Northern Spain
Cristina Gómez, Isabel Aulló, Fernando Montes
INIA, Forest Research Centre, Dpt. of Silviculture and Forest Management, Crta. La Coruña km 7,5, 28040 Madrid, Spain
Keywords: annual land cover, Landsat, SVM, species dynamics, colonization, densification, Spain
Abstract
Dominant tree species may change gradually, as result of natural succession or encouraged by
fluctuations of average temperatures and precipitation. Species dominance may also change under
anthropogenic influence with forest management or use. Landsat data from the TM, ETM+, and OLI
sensors, spanning the period from 1984 to 2015 were used in this work to characterize the dominant
tree species dynamics in a 300000 ha mountain area englobing Ordesa National Park (NP) in Northern
Spain. Land cover types and land cover changes were retrospectively analysed with an almost annual
series built with images from the United States Geological Survey (USGS) and the European Space
Agency (ESA) archives. A twelve-class land cover classification scheme was specifically designed to
identify dominant tree species, accounting the species successional stages and phenological cycles in
this region. A Support Vector Machine (SVM) classification rule was trained with reference data from
field plots measured in 2013 with ForeStereo and contemporaneous aerial photography, and applied
to 28 summery images radiometrically normalized with the Iteratively Re-weighted Multivariate
Alteration Detection (IR MAD). The input variables to the classifier included reflectance, Tasseled Cap
Transformation (TCT) derived indices, texture features, and features derived from a Digital Elevation
Model (DEM). For stabilization of the temporal land cover product (28 land cover maps in a 32 year
period) a filter of ecologically feasible annual land cover transitions was applied. The final classification
was validated over a reference image (2013) with overall accuracy 75%. Temporal trajectories of
the Tasseled Cap Angle (TCA), an index positively related with density variables, and its temporal
derivative, the Process Indicator (PI) were analysed to characterize the change and rates of change
in forest density. Inference of the tree species changing dominance in mixed forests was based on
trends of change in the temporal trajectories of the classification probabilities deployed by the SVM
classifier at each annual land cover map. Our results indicate that some areas at high altitudes are
progressively being colonized by natural forests of pine. A generalized densification tendency has
been observed, particularly since the 1990s, in forest stands without human intervention and those
where grazing practices have been reduced or abandoned. Actual conditions seem to favour a growing
dominance of Fagus sylvatica L. and Abies alba Mill. in mixed forests. The dynamics of tree species in
the NP and adjacent areas indicate a similar pattern of change, suggesting changes are both resulting
from successional processes and changes in human influence. Frequent land cover products derived
from a series of Landsat data and supported by ecological knowledge provide valuable insights of
diversity and species dynamics.
229
Universidad Mayor
Forest Monitoring
POSTER
EO-1 SWIR band detection capability and comparison
with Landsat
Shannon Franks1, Christopher Neigh2
1
Stinger Ghaffarian Technologies (SGT), NASA/GSFC, Greenbelt Maryland 20771
2
NASA/Goddard Space Flight Center, Code 618, Greenbelt Maryland 20771
Keywords: EO-1, Advanced Land Imager, ALI, Shortwave infrared, SWIR, Landsat, Spectral Band Pass
Abstract:
The Earth Observing One (EO-1) satellite has been in orbit for 15 years. The mission was started to
test new technologies that could perhaps be incorporated in future earth observation satellites like
Landsat. In this study, we test one of those technologies for its usefulness in detecting Tamarack
(Larix laricina) tree mortality due to attack of Eastern Larch Beetle (Dendroctonus simplex) around
Park Falls, Wisconsin. The sensitivity of the extra shortwave infrared (SWIR) band (1.2 – 1.3 μm)
onboard the Advanced Land Imager (ALI) was studied in comparison to the traditional SWIR bands
that are collected from the Landsat satellites. Time series ALI and Landsat imagery calibrated to
surface reflectance was used in the study. The locations of the damage was provided by the USFS
Forest Health Technology Enterprise Team and was validated with high resolution WorldView 2 data
showing the individually attacked trees. Differences in the spectral sensitivity to tree mortality were
found in the various SWIR bands and that knowledge could be useful when detecting insect damage.
Our results show that the shorter wavelength SWIR band and the longer wavelength band are more
sensitive to the damage than the one that is traditionally used in satellite remote sensing.
**** I would prefer a poster presentation
230
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Estimación y monitoreo de cobertura de malezas a través de
imágenes satelitales
Rodrigo Saavedra 1, Rodrigo Burgos 1, Jorge Requena 2, Mauricio Reyes 3, Cecilia Muñoz4
Bioforest S.A, Unidad Bioinformática y División de Silvicultura. Forestal Arauco S.A.
2
Forestal Arauco Zona Norte, Mensura y Cartografía.
3
Forestal Arauco Zona Centro, Planificación Silvícola.
4
Forestal Arauco Zona Sur, Vivero y Planificación Silvícola.
1
Palabras claves: cobertura de malezas, monitoreo, imágenes satelitales, uso operacional.
Resumen
La presencia de malezas en las plantaciones forestales si bien tiene algunos efectos favorables como la
protección del suelo, estas afectan negativamente la sobrevivencia y el crecimiento de las plantaciones.
La cobertura de malezas es monitoreada a través de las visitas que realizan los jefes de áreas a cada
uno de los predios, sin embargo dado los grandes volúmenes de superficie que manejan anualmente,
estos monitoreos pueden ser tardíos para un control oportuno de la vegetación competidora. Es por
esto que se propone el desarrollo de un sistema como herramienta de apoyo para uso operacional, que
permita la estimación y monitoreo de forma oportuna y masiva de la cobertura de malezas a través de
imágenes satelitales.
Este estudio tiene como objetivos 1) Desarrollar metodología para estimar la cobertura de malezas a
través de imágenes satelitales y 2) Desarrollar una herramienta de uso operacional como apoyo a la
planificación del control de malezas. Para esto se utilizaron imágenes del satélite Landsat 8 (NASA
- USGS). Posteriormente se procedió a la evaluación de los predios donde se realizó un muestreo
compuesto generándose cuadrantes para la estimación de cobertura de malezas en terreno. Se
observó que existe una correlación significativa entre la cobertura de malezas observada en terreno
y el índice espectral de vegetación NDVI de 0.71, generándose un modelo que permita la estimación
de la cobertura de malezas a nivel patrimonial y con el fin de poder implementar un sistema de apoyo
para uso operacional.
231
Universidad Mayor
Forest Monitoring
POSTER
Exploring remote sensing potential in land use, land use
change and forestry (LULUCF) inventories in Aragón
Eva Sevillano Marco1, Eduardo Notivol Paíno2,
1
Regional de Blumenau - Regional University of Blumenau (FURB), Department of Forestry. Rua São Paulo, 3250, 89030000 Blumenau-Santa Catarina, Brazil evasevillano@yahoo.es.
2
Agro-food technology and research Centre of Aragón (CITA). Forest Resources and Agriculture Applications Avda.
Montañana 930, 50059 Zaragoza, Spain enotivol@aragon.es. Keywords: Landsat, Classification, Land Use, Change,
Forestry, Biomass, Kyoto Protocols
Abstract
In the Green-house Effect Inventories reports, surfaces estimation and annual changes in land uses
could be more accurately calculated by means of the application of remote sensing than with the
commonly used methods used by the Regional Government in Aragón (Spain) based on available static
maps, including national land cover or forest maps. The objective of the study is the evaluation of the
potential application of remote sensing techniques of land cover/land use yearly monitoring. LULUCF
categories maps and change detection (i.e., between two years, 2002 and 2009) were obtained from
a maximum likelihood supervised classification of Landsat multitemporal scenes (phenologically
representative dates throughout the year). Confusion matrices provided global accuracy values of
respectively 89.24% and 83.69% in 2002 and 2009.
Introduction
The Kyoto Protocol (IPCC 2000) establishes the
requirement of annual inventories submission of
all anthropogenic greenhouse gas emissions from
sources and removals from sinks. Surfaces and
annual changes in land uses could be more accurately
estimated than with commonly used methods in the
Green-house Effect Inventories reports, based on
static maps (i.e., CORINE Land Cover, Information
system of land cover in Spain SIOSE and National
Forest Map NFM). In this context, Volden et al. (2003)
and Romero et al. (2004) mention five European
countries –Finland, Italy, Netherlands, Norway,
Switzerland- that examined the potential of remote
sensing for the evaluating land use and land-use
change for the Kyoto Protocol reporting. Userrequirements are defined for determining land use,
assessing forest areas and estimating aboveground
biomass at a minimum spatial resolution for
mapping 0.5 ha (Kyoto Protocol requirements).
Transparency, reliability and periodicity of data
are highlighted features linked to remote sensing
methods. Also, a high degree of automation without
terrestrial control is desirable Romero et al. (2004).
The Landsat series was identified in several test
232
areas as one the most useful generally speaking
considering worldwide availability of products,
homogenization procedures across inventories
and compromise between the medium spatial
resolution it provides and the Kyoto requirements. It
is advisable to complement the imagery with aerial
ortophotographies, climatic and topographic maps.
Among the identified weaknesses, the first trials
showed overestimation in forestlands surfaces as
well as expected issues in small and sparse forest
patches. From the LULUCF categories surfaces, a
further step in the carbon stocks assessment is the
aboveground biomass quantification. Certainly
although not a specific requirement in the Kyoto
Protocol, it is a key element for the inventories
(e.g., land use change in forestalands suchlike
afforestation and deforestation), and can be
approached by means of remote sensing techniques
as well. Therefore, the objective of the present
study is the evaluation of the potential application
of remote sensing techniques in the methodology
definition at regional scale of land cover/land use
yearly monitoring for the region of Aragón, in Spain.
A first exploration of model fitting techniques is
undergoing starting from this study.
Universidad Mayor
Methodology
Aragon is one of the 17 autonomous regions of Spain
and covers an area of approximately 47719 km2. It is
located at NE of Iberian peninsula between 39ºN and
43ºN. The orography of the region is characterized
by the Ebro Valley depression (altitude ranging from
150 to 300 m above sea level) and two mountain
ranges surrounding the valley: the Pyrenees at north
with peaks over 3000 m and the Iberian System
range with maximum heights over 2000 m. Aragon’s
climate can be defined as continental moderate.
Temperatures are determined mainly by altitude,
ranging from cold or very cold in winter and cool in
summer in the mountains to the north (Pyrenees)
and to the south and west (Iberian range), to mild
in winter and hot in summer in the central lowlands.
Rainfall is also very variable, with very low mean
values in the central areas and increasingly higher
values in mountain areas, especially in the high
Pyrenees. In the middle of Aragon, which is only
200 metres above sea level, the annual average
temperature is around 14 °C To the north and south
of the Ebro valley, where the elevation rises to 500
m above sea level, the temperature drops by two
degrees. In the mountains, between 600 m and
1000 m observed temperatures are between 11 and
12 °C. The vegetation follows the oscillations of the
extreme topography and climate. There is a great
variety, both in wild vegetation and crops. In the
highlands forests (pines, firs, beeches, oaks) can be
found, as well as bushes and meadows, while in the
lower altitude areas of the Ebro valley green oak and
sabina are the most abundant trees, apart from the
agricultural lands.
Groundtruth
data
was
derived
from
photointerpretation and available products like
SIOSE 2005, CORINE 1990, 2000, 2006 and Global
Land Cover 2000, and National Forest Map (i.e.,
Coordination of information on the environment
CORINE Land Cover, Information system of land
cover in Spain SIOSE and National Forest Map
NFM). A compromise between capturing the
variability inherent to the thematic category and
its characteristic features is needed. The selection
criterion has been to prioritize the extraction of pure
areas, avoiding usage mixtures and ecotones that
could introduce noise to the classification stage.
Unchanged areas between the whole time series
have been chosen and erosion has been applied
ForestSAT 2016 Abstracts Summary
therefore minimizing border areas, displacement
and topology errors. The sample was split on a
50% basis to obtain the training areas as input
for the supervised classification and test areas for
the evaluation and validation processes. Avoiding
cloud scenes when possible, scenes were selected
in phenologically representative dates throughout
the year. Between three and six dates per year for
each scene are needed. Ideally, considering the
regional environmental variables: one scene in
between March-April, two between May-June, one
between August-September, one between OctoberDecember and one between January-February. All
the databases have been projected to ETRS 1989
UTM Zone 30N (EPSG 3042) and subset to the area of
study. Cloud, shadows and snow have been masked in
order to minimize spectral distortions in radiometric
corrections (Chávez 1996 method employed, using
a 20 m pixel digital elevation model DEM together
with the required inputs of Julian day, solar elevation
angle, and azimut angle, irradiances, transmissivity,
slope-gain and offset-bias). Spectral vegetation
índices were derived: Normalized Difference
Vegetation Index NDVI, Ratio Vegetation Index RVI,
Soil Adjusted Vegetation Index SAVI, and Tasseled
Cap. Additionally, from the DEM, slope and aspect
layers were generated to be used in the classification
process. Finally, to account for the units and ranges
diversity of all the scenes, indices and topographic
layers, previous to the classification, standardization
is required (Spiegel 1991), calculating the mean and
standard deviation for each variable.
The legend of six categories follows the classes
defined by the United Nations Framework
Convention on Climatic Change UNFCCC for LULUCF.
After the standardization process abovementioned,
6 spectral bands (optical bands 1,2,3,4,5 and 7 of
Landsat 5TM and Landsat 7 ETM+) from 3-6 dates
per year, vegetation indices, tasselled cap, slopes,
elevations and aspect were stacked to be used in the
classification. A supervised classification method
(maximum likelihood) generates the corresponding
land use thematic cartography for 2002 and 2009.
Accordingly, LULUCF categories maps and also the
change detection (i.e., between two years, 2002
and 2009) were obtained from classification of
Landsat multitemporal scenes. Confusion matrices
and signature separability were the evaluation
techniques used. A workflow diagram is shown in
Figure 1.
233
Universidad Mayor
Forest Monitoring
Results
The changes between categories between 2002 and
2009 are shown in Figure 2 and Table 1.
At detailed zooms, and legend containing each of
the transitions, many other features relevant for
the regional management and LULUCF monitoring
become intelligible.
Figure 2. Map of Aragón transitions in LULUCF
classes between 2002-2009.
Confusion matrices provided global accuracy values
of 89.24% in 2002 and 83.69% in 2009. User and
producer accuracies are depicted in Table 2.
Most vegetation covers surfaces remain unchanged,
as expected in vegetation dynamics in such period.
Figure 1. Workflow
Table 1. LULUCF Land use surfaces discriminated by remote sensing and their changes between the
considered period (2002-2009)
LULUCF Categories surfaces (km2) and changes
Forest
Grassland
Cropland
Soil
Urban
Water
km2
12000
878,939
5859,25
11400
475,379
309,146
To cropland
1249,37
To grassland
To water
To urban
To forest
To soil
Total changes
1313,05
16,81
2946,66
1896,86
6454,08
13876,83
Total unchanged
30900,00
Table 2. Kappa indices, producer accuracies and user accuracies for the LULUCF UNFCCC categories in 2002
and 2009 in Aragón (Spain).
g ( p )
2002 /2009
Grassland
Cropland
Soil
Urban
Water
Forest
234
Producer accuracy
User accuracy
Kappa index
62.9% / 67.8%
95.4% / 87.8%
0.95 / 0.86
93.7% / 79.4%
98.6% / 80.5%
76.0% / 75.6%
97.3% / 89.0%
96.9% / 86.7%
91.1% / 86.3%
65.6% / 54.6%
95.4% / 89.9%
100.0% / 95.4%
94.7% / 98.0%
0.89 / 0.83
0.61 / 0.48
0.95 / 0.72
1.00 / 0.94
0.93 / 0.97
Universidad Mayor
Likewise, the urban growth is clearly evidenced
locally. Tuning of methods is subject to availability
of updated coherent ground truth and field data
in accordance to requirements of classification
method and the subsequent definition of categories.
After the achievements verified in these first
trials, it is strongly recommended that future aims
address the automation of processes for a seamless
implementation at regional and homogenously at
national scales, in order to compile accurate land
cover temporal series for annual periods from the
reference year 1990 onwards (i.e., IPCC compliant
methods state reference year 1990). Such maps
can be complemented with more detailed legends,
for instance embedding Forest Maps, in order to
enable biomass stocks estimation by means of
geospatial modelling: e.g., relating stands data to
spectral variables or derived indices, resulting from
the combination of field data with remote sensing
variables. This line of work was initiated by fitting
aboveground biomass models from field data (CITA,
2008) and employing the LULUCF spectral variables
and derived products as predictors.
Conclusions
This exploration was intended to tailor research to
territorial management and provide the regional
government with fit-for-purpose resources in the
implementation of policies suchlike the Protocol of
Kyoto: in particular for the LULUFC sector accounting
methods in the Green-house effect gases inventory.
In the climatic change study, the continuity of
forerunner remote sensing missions, ever increasing
availability of products, with enhanced capabilities
(spatial and spectral resolutions, temporal coverage),
and development of sensors fusion techniques, are
relevant in the LULUCF sector accounting methods,
allowing for improved accuracy performance,
biomass modelling and back-monitoring as tested in
this pilot case.
Acknowledgements
This project was co-funded by the European
Regional Development Fund (ERDF) and the
Farming, Livestock, and Environment Department
of the Regional Government in Aragón (Spain),
2007-2013 Objective Regional Competitiveness
ForestSAT 2016 Abstracts Summary
and Employment, Axis 2: Environment and
Risk Prevention, Category 49: Climatic Change
Adaptation and Mitigation. The research was
assigned by the Regional Government to the Agrofood technology and research Centre of Aragón
(CITA), by the 182/2013 Decree 19th November 2013,
published in the official regional bulletin 234, 27th
November 2013. Research project: Green-house
effect gases inventory, emissions and sinks in the
Land Use, Land Use Change and Forestry –LULUCFsector).
References
Chávez P S, 1996. Image-based atmospheric
corrections.
Revisited
and
improved.
Photogrammetric Engineering & Remote
Sensing 62.9 1025-1036
Centro
de
Investigación
y
Tecnología
Agroalimentaria de Aragón (CITA), 2008. Estudio
sobre la funcionalidad de la vegetación leñosa
de Aragón como sumidero de CO2: existencias
y potencialidad (estimación cuantitativa y
predicciones de fijación). Centro de Investigación
y Tecnología Agroalimentaria de Aragón,
Zaragoza, Spain. Retrieved from: http://
www.aragon.es/estaticos/GobiernoAragon/
Departamentos/MedioAmbiente/Areas/03_
Cambio_climatico/06_Proyectos_actuaciones_
Emisiones_GEI/estudio.pdf
Intergovernmental Panel on Climate Change (IPCC).
2000. Watson R, Noble I R, Bolin B, Ravindranath,
N H, Verardo D J, Dokken D J (Eds) Land use,
Land-use Change, and Forestry: A Special Report.
Cambridge University Press. Cambridge, UK.
Romero J, Volz R, Giamboni M, Rüsch W 2004.
The role of the forest in the Kyoto Protocol Estimation of carbon reserve using satellite
data). Schweizerische Zeitschrift für Forstwesen
155 (5), pp. 125-133.
Spiegel M R 1991 Statistics. McGraw Hill 556 p.
Volden E, Bruzzone l, Romero J, Lumicisi A, Renda
O, La Fortezza R, Tomppo E, Aalde H 2003. Forest
environmental reporting services. In: Learning
from Earth’s Shapes and Sizes. Proceedings
of the International Geoscience and Remote
Sensing Symposium (IGARSS ‘03), Vol. 7, July 2125, 2003, Toulouse, France, pp. 4582-4584.
235
Universidad Mayor
Forest Monitoring
Fire behavior simulation from global fuel and climatic
information
M. Lucrecia Pettinari1,2, E. Chuvieco1
1.
Department of Geology, Geography and Environment, University of Alcala, Alcalá de Henares, Spain.
2.
E-mail: mlucrecia.pettinari@uah.es
This study presents a global fire behavior simulation
based on a global fuelbed dataset and climatic
and topographic information. The simulation was
executed using the Fuel Characteristic Classification
System (FCCS). The climatic information, extracted
from the ERA-Interim Global Reanalysis covered the
period 1980-2010, and daily weather parameters
were used to calculate the mean monthly fuel
moisture content (FMC) and wind speed for the
early afternoon period. Also, as the most severe
fires occurs with extreme environmental conditions,
a worst condition scenario was created using the
mean value of the 30 days with lowest FMC during
each month of the study period.
The FMC, wind speed and slope information was
grouped into classes, and FCCS was used to simulate
the reaction intensity, rate of spread and flame
length of the fuelbeds in the different environmental
conditions. These results were then mapped,
showing the variation in surface fire behavior during
the different months of the year throughout the
world, both due to the climatic conditions and the
characteristics of the fuels. The surface fire behavior
parameters identified the fuels and environmental
conditions that could cause more severe fire events,
and could be used as a mean to assess fire danger.
Keywords: global fuel map, FCCS, fuel moisture
content, fire behavior
Introduction:
Wildfires occur due to a combination of available
fuels, sources of ignition, and environmental
conditions (mostly weather), that allow the fire to
be sustained and spread. Most regional fire behavior
modeling systems consider weather variables to
estimate critical parameters such as rate of spread,
flame length or the energy released (Rothermel
1983), but only a few consider explicitly the
236
differences in fire behavior due to fuel properties.
The objective of this study was to combine fuel
and weather information to obtain fire behavior
parameters at a global scale, using a global fuel map
previously developed (Pettinari and Chuvieco 2016)
Methodology:
To calculate fire behavior, we used the Fuel
Characteristics Classification System (Ottmar et
al. 2007). The FCCS was designed to represent the
structural and geographic diversity in wildland fuels
and combines the fuel properties into “fuelbeds”,
which include the physical and chemical variables
used to model fire behavior and fuel consumption
and predict emissions (Riccardi et al. 2007). Based
on input environmental variables, the FCCS also
predicts surface fire behavior parameters using a
reformulation of the Rothermel (1972) fire behavior
model (Sandberg et al. 2007).
The fuel distribution and their physical and chemical
variables related to fire behavior were extracted from
the Global Fuel Dataset developed by Pettinari and
Chuvieco (2016). This dataset was developed from
different spatial variables, both based on satellite
Earth observation products and fuel databases, and
is comprised by two products: a global fuelbed map
(at a resolution of approximately 300m), developed
using land cover and biomes information, and a
database that includes the parameters of each
fuelbed that are related to fire behavior and effects.
To calculate the fire behavior parameters, FCCS
requires information on fuel moisture content
(FMC), wind speed and slope. Percentage slope was
calculated using the Global 30 Arc-Second Elevation
(GTOPO30) product and the different percentages
were grouped in three slope classes: 0-5%, 5-45%
and >45% slope.
Universidad Mayor
The FMC and wind speed was calculated using the
climatic data of the ERA-Interim Global Reanalysis
(Dee et al. 2011) of the European Centre for MediumRange Weather Forecast. The weather information
covered a 30-year period, with daily data between
1981 and 2010, with a spatial resolution of 0.5
degrees.
To obtain the FMC, the 10-h fuel moisture content
was calculated using the equations of the National
Fire Danger Rating System (NFDRS, Cohen and
Deeming 1985), and using the following weather
variables from the ERA-Interim database: 2 meter
temperature, 2 meter dew point, total precipitation,
total cloud cover and snow depth. The 10-h FMC
obtained was converted to Fuel Moisture Scenarios
(FMS, Scott and Burgan 2005) to be input in FCCS.
The daily values of FMS obtained from the 19802010 climatic data were aggregated in a monthly
basis, to obtain the mean FMS for each month in
each 0.5° cell, as a proxy to the mean conditions that
could be found in that region during that month of
the year. However, since severe weather conditions
(low FMC and high wind speed) usually cause the
most important fire events and largest fires, a “worst
conditions scenario” was created, calculating the
mean FMS of the 30 days with lowest FMC for the
entire period and for each month.
Wind speed was calculated from the 10m U and V
wind components of ERA-Interim, and converted
to midflame wind speed using a wind adjustment
factor of 0.348. The daily values of wind speed were
aggregated by month to obtain the mean midflame
wind speed for each month in each 0.5º cell, and the
resulting wind speeds were grouped in three classes:
0-1 m.s-1, 1-2.5 m.s-1 and 2.5-5 m.s-1. To account for
the worst conditions scenario, the midflame wind
speed corresponding to the 30 worst-FMC-days
was obtained for each month, and the values were
averaged and then converted to wind class for those
conditions.
The possible FMSs, wind classes and slope classes
created 36 combinations of possible environmental
scenarios. Each fuelbed from the Global Fuelbed
Dataset was run in FCCS version 3.03.203 module
inside the Fuel and Fire Tools (http://www.fs.fed.
us/pnw/fera/fft/index.shtml, accessed November
2016) in all the scenarios, to obtain all possible rate
of spread (ROS), flame length (FL) and reaction
ForestSAT 2016 Abstracts Summary
intensity (RI) values for each fuelbed. To create
the fire behavior maps, each fuelbed pixel in the
fuelbed map was assigned the fire behavior results
corresponding to the environmental scenario of
their location and month that was being evaluated,
both for the mean and worst conditions.
Results:
Figure 1 shows, as an example of the results obtained:
the mean reaction intensity for the months of January
(a) and July (b). The highest values of RI obtained
for January were 4664 kJ.m-2.s-1, both in mean
and worst conditions and located in the croplandshrubland belt of Northern Africa. But during worst
conditions the geographical distribution of RI values
higher than 2000 kJ.m-2.s-1 extended also to the
Brazilian Cerrado and South-American shrublands
in Argentina, Bolivia and Uruguay, along with the
savannas in Southern Africa. Regarding July, the
highest RI values obtained were 3835 kJ.m-2.s-1
for the mean conditions, located in Zambia, and
4435 kJ.m-2.s-1 for the worst conditions, located
in Argentina. The geographical distribution of the
cells with RI > 2000 kJ.m-2.s-1 during July’s mean
conditions include the Southern-African Savannas
and the Brazilian Cerrado, but extend to Northern
Africa and South-American shrublands when the
worst conditions are considered.
As with the RI, the ROS and FL also showed changes in
value according to seasonal weather conditions and
fuelbeds. The highest values of ROS were obtained
for grassland fuelbeds located mostly in the African
continent, and also in the North-American Deserts,
along with rice croplands in southern Kazakhstan,
reaching to values up to 1.013 m.s-1 for the July’s
worst conditions. Regarding the FL, the highest
values obtained (>6 m) were all located in Africa.
It should be noted that the coarse resolution of
the weather information allowed obtaining only a
generalized and estimated result of the fire behavior
outputs. Wind speed, for example, is highly variable
in time and space and affected by sheltering and
terrain, which could not be considered at the
resolution of the data (0.5º) or even the fuelbed map
(approx. 300m). As such, the values used for the
analysis are not expected to be realistic in local fire
events, but to introduce a coarse variable showing
areas and months in which fire could have a more
237
Forest Monitoring
Universidad Mayor
(a)
(b)
Figure 1: Mean Reaction intensity (RI) for January (a) and July (b) for the study period.
severe behavior due to general wind speeds being
higher than in other regions or months.
Still, fire behavior parameters are important because
they allow understanding the fire itself, and not only
classifying the fire as a binary event (fire – no fire).
Fire events are extremely variable, and they produce
different effects in the ecosystem. For this reason,
the study of the characteristics of the fire events,
by means of estimating fire behavior parameters,
can help identifying land cover and environmental
conditions combinations that could cause severe
fires with relevant environmental impacts, thus
posing a higher danger.
Conclusions:
The maps showed that the highest values of
reaction intensity were found in shrubland fuelbeds
238
in tropical and sub-tropical dry regions, the highest
rates of spread were obtained for grasslands in
those locations and shrublands in desertic areas,
and the highest flame lengths occurred in African
savannas. All parameters were highly affected by the
environmental conditions, increasing their values up
to an order of magnitude with changes in the fuel
moisture content of the fuels.
The coarse resolution of the climatic data, though
provided information on regional weather conditions,
cannot allow predicting accurate local fire behavior,
because local conditions of fuel moisture and wind
can rapidly change and modify the fire behavior.
Still, the results showed the importance of including
detailed fuel information into fire risk assessment
systems based on weather parameters, as it could
help to better estimate the expected fire behavior
and effects.
Universidad Mayor
References:
Cohen J. D. and Deeming J. E. (1985). “The National
Fire-Danger Rating System: basic equations.”
General Technical Report PSW-82. Berkeley, CA,
USDA Forest Service, Pacific Southwest Forest
and Range Experiment Station. 23 pp.
Dee D. P., Uppala S. M., Simmons A. J., Berrisford P.,
Poli P., Kobayashi S., Andrae U., Balmaseda M. A.,
Balsamo G., Bauer P., Bechtold P., Beljaars A. C.
M., van de Berg L., Bidlot J., Bormann N., Delsol
C., Dragani R., Fuentes M., Geer A. J., Haimberger
L., Healy S. B., Hersbach H., Hólm E. V., Isaksen
L., Kallberg P., Köhler M., Matricardi M., McNally
A. P., Monge-Sanz B. M., Morcrette J.-J., Park B.K., Peubey C., de Rosnay P., Tavolato C., Thépaut
J.-N. and Vitart F. (2011). “The ERA-Interim
reanalysis: configuration and performance of the
data assimilation system.” Quarterly Journal of
the Royal Meteorological Society 137: 553-597.
Ottmar R. D., Sandberg D. V., Riccardi C. L. and
Prichard S. J. (2007). “An overview of the
Fuel Characteristic Classification System Quantifying, classifying, and creating fuelbeds
for resource planning.” Canadian Journal of
Forest Research 37(12): 2383-2393.
Pettinari M. L. and Chuvieco E. (2016). “Generation of
a global fuel data set using the Fuel Characteristic
ForestSAT 2016 Abstracts Summary
Classification System.” Biogeosciences 13(7):
2061-2076.
Riccardi C. L., Ottmar R. D., Sandberg D. V., Andreu
A., Elman E., Kopper K. and Long J. (2007). “The
fuelbed: a key element of the Fuel Characteristic
Classification System.” Canadian Journal of
Forest Research 37(12): 2394-2412.
Rothermel R. C. (1972). “A mathematical model for
predicting fire spread in wildland fuels.” Research
Paper INT-115. Odgen, UT., USDA Forest Service,
Intermountain Forest and Range Experiment
Station. 40 pp.
Rothermel R. C. (1983). “How to predict the spread
and intensity of forest and range fires.” INT-143.
Boise, ID., National Wildfire Coordinating Group,
USDA Forest Service Intermountain Research
Station. 166 pp.
Sandberg D. V., Riccardi C. L. and Schaaf M. D. (2007).
“Reformulation of Rothermel’s wildland fire
behaviour model for heterogeneous fuelbeds.”
Canadian Journal of Forest Research 37(12):
2438-2455.
Scott J. H. and Burgan R. E. (2005). “Standard fire
behavior fuel models: a comprehensive set for
use with Rothermel’s Surface Fire Spread Model.”
RMRS-GTR-153. Fort Collins, CO., USDA Forest
Service, Rocky Mountain Research Station. 80
pp.
239
Universidad Mayor
Forest Monitoring
POSTER
Forest area changes and impact of forest boundary
delineation on change detection in forested
landscapes in Eastern Europe
Urmas Peterson1,2 and Jaan Liira 3
1
Tartu Observatory, Tõravere 61602, Tartumaa, Estonia, Tel: +3727410152, Fax: +3727410205, E-mail: urpe@aai.ee, webpage: http://www.to.ee
2
Institute of Forestry and Rural Engineering, Estonian University of Life Sciences, Kreutzwaldi 5, Tartu 51014, Estonia
3
Institute of Ecology and Earth Sciences, University of Tartu, Lai 40, Tartu 51005, Estonia,
Tel and Fax: +3727376222, E-mail: jaan.liira@ut.ee, web-page: http://www.botany.ut.ee
Keywords: Forest mapping, forest change detection, gradual changes, winter images
Boreal and northern temperate forests cover
substantial part of European land area. Fine-scale
and spatially explicit data on changes of the forested
area in Eastern Europe are sketchy. Within the last
decades the change of forested area in this region
is expressed mostly in a gradual and slow expansion
of forested patches on the former agricultural land.
Forests at northern latitudes are characterized with
winters during which snow covers several months
the ground. Winter in boreal and hemi-boreal
latitudes is the season with the greatest target to
background contrast on predominantly two-class
images composed of forest and non-forest classes.
In winter, forest patches are surround on all sides by
open areas with bright snow cover.
We used imagery of Landsat sensors from the
United States Geological Survey (USGS) archive for
three time periods: mid to late 1980s, early 2000s
and 2010s (years 2011 to 2016). Our study area: three
Baltic states, Byelorussia, Ukraine and European
part of Russia were covered with winter images on
these years in the USGS archive. Results obtained
from moderate resolution Landsat images were
locally supported with high resolution winter images
for error estimates or with national basic map data
if available.
The images were classified into “forest” and “nonforest” classes by thresholding pixel radiance values.
240
The optimal edge threshold was defined by looking
for a maximum radiance contrast of neighbouring
pixels in a forest boundary area. Forest boundary
segments were assigned an attribute of cardinal
direction according to their relative position at the
edge of a forest patch.
Forest boundary locations on images taken in different
solar illumination conditions were compared. The
aim was to distinguish real changes at forest edges
(i.e. forest patch expansion) from those changes in
image classification resulting from different solar
elevation and atmospheric haze conditions. We
found that shade and forest structure affect edge
detection on medium resolution satellite images.
Regression analysis was used to relate the possible
drivers of forest area increase in Eastern Europe.
We found locally extensive increase of forest patch
areas on abandoned farmland in the studied region.
However, we also found strong differences in the
rates and spatial patterns of forest area change
among the countries and administrative regions
in our study area. The role of local agricultural and
forest management activities, distance to local
and regional centres, relative forested area and
configuration of forested patches are discussed as
possible driving factors of forested area change.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Forest monitoring using remote sensing time-series: The case
of Colombian Andes Protected Areas
1
Murillo-Sandoval Paulo J.1,2 , Hilker Thomas1
Oregon State University, 2 Universidad del Tolima
Keywords: Protected Area, NDVI, Trends, MODIS, BFAST, Landsat, time-series
Protected Areas (PA’s) are an essential tool
for conservation of biodiversity and species
maintenance. The tropical Andes have been classified
as a “hyper” hotspot for endemism and conservation
support, and as a result, monitoring their PAs is a
vital requirement for conservation biology. One way
to evaluate the ecological integrity of PA’s is through
monitoring vegetation phenology, degradation and
disturbances. However, regular field monitoring
of remote areas such as the Andes Mountains is
often not feasible due limited accessibility and high
costs. Remote sensing therefore provides a key
tool for assessing landscape changes over time, but
prevailing cloud cover and lack of frequent, highresolution datasets provides significant challenges.
Recently, advances in atmospheric correction and
cloud screening have opened new opportunities for
remote observations of otherwise challenging areas,
such as the Andes. The Multi-angle Implementation
of Atmospheric Correction algorithm (MAIAC) is
a novel cloud screening technique that can help
to increases the quantity of available cloud free
observations while reducing noise in land surface
reflectance at 1km resolution. These observations,
while often too coarse to observe small scale
disturbances, such as caused by human intervention,
may be combined with complementary medium
resolution sensors to allow accurate monitoring of
landscape degradation and disturbance in space
and time. Here, we combine a time-series (20002014) of satellite observations derived from MODISMAIAC with 30 meter resolution Landsat imagery.
Trends in landscape patterns were evaluated using
the Breaks For Additive Seasonal and Trend (BFAST)
algorithm in order to compare 15-year vegetation
trends inside and outside the “Picachos” Protected
area of the Colombian Andes. Our results show a
progressive upslope logging inside “Picachos”. The
combination of low and medium resolution sensors
processed by BFAST is a useful technique to obtain
reliable information about the contribution of shortterm and longer-term shifts on Protected Areas and
to detect progressive deforestation patterns at fine
spatial scale.
241
Forest Monitoring
Universidad Mayor
Global fire impacts assessment from long term analysis of
burned area products
Emilio Chuvieco1*, Lucrecia Pettinari1, Itziar Alonso-Canas1 , Marc Padilla2, Kevin Tansey2, Chao Yue3,4, Philippe Ciais4,
Angelika Heil5 , Johannes Kaiser5, Florent Mouillot6, Jose Miguel Pereira7, Duarte Oom7, Aitor Bastarrika8, Ekhi Roteta8,
Guido van der Werf9, Thomas Storm10, Jose Gomez-Dans11 and Philip Lewis11
Environmental Remote Sensing Research Group, Universidad de Alcalá, Spain. mlucrecia.pettinari@uah.es,
itziar.alonsoc@uah.es,
2
Department of Geography, University of Leicester, Leicester, UK kjt7@leicester.ac.uk mp489@leicester.ac.uk
3
Laboratoire de Glaciologie et Géophysique de l’Environnement, UJF, CNRS, Saint Martin d’Hères CEDEX, France.
chaoyuejoy@gmail.com
4
Laboratoire des Sciences du Climat et de l’Environnement, LSCE CEA CNRS UVSQ, 91191 Gif- Sur-Yvette, France,
philippe.ciais@lsce.ipsl.fr
5
Max Planck Institute for Chemistry, Mainz, Germany. a.heil@mpic.de, j.kaiser@mpic.de
6
UMR CEFE 5175, CNRS/Université de Montpellier/Université Paul-Valéry Montpellier/ EPHE/IRD, France.
florent.mouillot@cefe.cnrs.fr
7
Centro de Estudos Florestais, Instituto Superior de Agronomia, Universidade de Lisboa. Lisboa, Portugal.
jmocpereira@gmail.com
8
University of the Basque Country, Vitoria, Spain. aitor.bastarrika@ehu.es; ekhi.roteta@gmail.com
9
VUA - Stichting VU-VUmc, Netherlands, guido.vander.werf@vu.nl
10
Brockmann Consult, Germany, thomas.storm@brockmann-consult.de
11
University College London, United Kingdom, {j.gomez-dans, p.lewis}@ucl.ac.uk
(*) Corresponding author: Emilio Chuvieco, Environmental Remote Sensing Research Unit, Universidad de Alcalá,
Telf: 918854438 / Fax: 918854439; e-mail: emilio.chuvieco@uah.es
1
Abstract
Biomass burnings (including forest, grassland, peatland and agricultural fires) have important
impacts on global terrestrial and atmospheric systems, affecting land cover, surface albedo, and the
atmospheric concentration of greenhouse gases, chemically reactive species and aerosols. Several
products have been generated in the last years to estimate total burned area, but uncertainties remain,
particularly those associated to small and low intensity fires. Impact of climate and societal changes
modify traditional fire regimes, extending fire seasons, increasing fire severity or introducing fire in
sensitive areas.
The Fire_cci project of the European Space Agency Climate Change Initiative aims to generate
consistent time series of burned area products to assess the extents of biomass burnings, as well as
their spatial and temporal characteristics. Fire impacts on atmospheric and terrestrial processes are
assessed, with particular attention to CO2, CO and CH4 emissions, modifications of vegetation patterns
and biomass availability.
The global burned area products of the Fire_cci program are being developed from MERIS and
MODIS sensors (300 m and 250m of spatial resolution respectively), complemented with a small-fire
databased generated from medium resolution sensors (Sentinel-2’s MSI, 10-20 m of spatial resolution,
and Proba-V, 100-300 m resolution) for the African continent. BA algorithms for new Sentinel-3 sensors
(OLCI and SLSTR) will also be developed. Validation of the global products is based on a statistical
sampling design of Landsat frames. For these sites, fire reference perimeters are generated based on
multitemporal analysis of TM/ETM+/OLI images.
The paper will present the current status of the project and will provide analysis of the first global
burned area product based on the full-time series of high resolution MERIS sensor data (at 300 m).
242
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Improving forest change detection in the UK using
LandTrendr and TimeSync Landsat analysis tools
Jacqueline Rosette1, Iain Bye1, Zhiqiang Yang2,3, Warren Cohen3, Dirk Pflugmacher4, Juan Suárez5 and Helen McKay5
2
1
Swansea University, Singleton Park, Swansea SA2 8PP, UK
Forest Ecosystems and Society, Oregon State University, Corvallis, OR 97331, USA
3
US Forest Service, Pacific Northwest Research Station, Corvallis, OR 97331, USA
4
Humboldt-Universität zu Berlin, 10099 Berlin, Germany
5
Forest Research, Northern Research Station, Roslin EH25 9SY, UK
Key words: Landsat, disturbance, change detection, forest management, forest condition
This research explores the innovative application of
remote sensing to enhance forestry management
and monitoring methods in the UK.
In partnership with the US Forest Service and
Oregon State University, the project utilises
the unique archive of Landsat imagery, and the
advanced LandTrendr and TimeSync time series
analysis tools, to investigate changes in forest
cover, to identify trends of decline and recovery,
and attribution of those trends. Results are
verified using subcompartment-level management
information, records of tree health and action taken,
and sequential lidar data showing changes in growth
trajectories.
containing veteran trees. The timber rights of this
private forest are leased to the Forestry Commission,
meaning that this area contains a combination of
managed stands and ancient woodland.
Figure 1 shows an example of management and
condition history for a stand in the Cowal and
Trossachs Forest District, detected using LandTrendr
time series analysis.
The project focuses on three study areas which
have experienced different predominant forms of
disturbance and forest cover.
The Loch Lomond and Trossachs National Park in
Scotland has suffered a series of severe storms in
recent years which have caused widespread loss
from windthrow, and subsequent progression of
wind damage through adjacent areas.
Japanese larch plantations in South Wales have
been affected by Phytophthora ramorum which
expanded from southwest England, and which
has been contained through severe management
intervention.
The Savernake Forest in England has been
designated a Site of Special Scientific Interest, and
is derived from ancient wood pasture management
Figure 1. Blue markers represent observed Tasselled
Cap Wetness (TCW) values of vertices (identified
points of inflection); red points and connecting red
lines show fitted trajectories (LandTrendr algorithm). Note that the trend is inverted such that an
increase in TCW (y axis values) indicates loss.
Automatic algorithm detection of stability (19852000), felling (2003-4), rapid recovery (to 2010) followed by slower decline in condition due to Dothistroma Needle Blight (>2010).
243
Forest Monitoring
The nature of these sites and their disturbance
histories provide the opportunity to investigate
the application of time series analysis using the
LandTrendr and TimeSync Landsat analysis tools
in the UK context, where cloud, seasonality and
fragmented forest cover create challenging
conditions. The sensitivity of the algorithms to
244
Universidad Mayor
accurately detect forest change of varying intensity
and duration (e.g. clear felling or sudden wind
loss versus decline and recovery relating to forest
condition) is investigated. This is of direct benefit
the British Forestry Commission for the accuracy
and currency of information regarding the national
forest resource.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Interpreted high resolution imagery for rapid assessment of
land use and land cover changes in the United States - The
USDA Forest Service’s Image-based Change Estimation
(ICE) project
Mark Finco 1, Abigail Schaaf 1, Kevin Megown2, Paul Patterson3, Tracey Frescino3, James Blehm4
1
2
RedCastle Resources, Inc. under contract to the USDA Forest Service, Salt Lake City, Utah USA
Geospatial Technology and Applications Center, USDA Forest Service, Salt Lake City, Utah USA
3
Forest Inventory and Analysis, USDA Forest Service, Ogden, Utah USA
4
Forest Inventory and Analysis, USDA Forest Service, St. Paul, Minnesota USA
Keywords: Change Estimation, Land Cover, Land Use, Interpretation, NAIP
Abstract:
The USDA Forest Service’s Forest Inventory and Analysis (FIA) program provides a long-standing,
rigorous, field plot-based national forest inventory for the United States. The cycle time for
remeasurement of field plots, however, is between 5 and 10 years, and national leadership identified a
need for more quickly available land cover and land use change information, even if the level of detail
is less than possible with field measurements. The FIA Image-based Change Estimation (ICE) protocols
resulted from this need to rapidly create change information.
Fortunately, the United States collects nationwide, high resolution imagery every two to three years
through the National Agricultural Imagery Program (NAIP) that is well suited for interpreting broad
land cover and land use classes. In collaboration with the Forest Service’s Remote Sensing Applications
Center (RSAC), FIA developed a two-step interpretation of NAIP over FIA plot locations to quantify
the amount and type of land cover and land use changes. The ICE project has been producing data for
nearly two years, and this talk will present experiences with the interpretation protocol and preliminary
change estimation results.
245
Universidad Mayor
Forest Monitoring
Landsat reveals the impact of disturbance on carbon storage
in the United States National Forest System
Sean Healey1, Alex Hernandez2, Chris Garrard2, James McCarter3, Crystal Raymond4
1
US Forest Service, 2 Utah State University, 3 North Carolina State University, 4 City of Seattle
Keywords: Landsat, disturbance, carbon
As part of a national strategy to address climate
change in the management of US national forests,
the US Forest Service has been directed to assess
the degree to which disturbances such as fire and
harvest have affected the amount of carbon stored in
each management unit. Because these assessments
are to be used in local forest planning, it is desirable
that they be based upon monitoring data that are
spatially and temporally specific; model results over
hypothetical or generalized landscapes are less
useful. The Forest Carbon Management Framework
(ForCaMF) was developed to meet this assessment
need by combining carbon dynamics built into an
empirically calibrated growth model with remotely
sensed (Landsat-based) records of historical
forest conditions and subsequent disturbance
history. Each of these data sources is calibrated
by and/or aligned with national forest inventory
data. ForCaMF outputs show the degree to which
disturbances since 1990 have altered current carbon
stocks on each national forest. A novel Monte Carlo
error simulation approach integrates uncertainties
associated with both model terms and remote
246
sensing inputs to provide confidence boundaries on
national forest-level assessments.
ForCaMF results reveal a country facing diverse
disturbance threats in different regions. On a perhectare basis, insects in the Rocky Mountains of
Colorado have had the largest local effects on
carbon storage, although fires in places as diverse
as California, Florida, and Montana have also had
a dramatic impact. Root disease is the dominant
disturbance process, carbon-wise, in some parts
of the Rocky Mountains, while harvest is the most
important disturbance factor in many southeastern
national forests. ForCaMF allows these impacts
to be compared both within and across individual
management units. Landsat has unique properties as
a monitoring tool that add irreplaceable temporal and
spatial context to assessments of forest dynamics.
Leveraging this context in highly scrutinized
applications such as these carbon assessments
will require the kind of uncertainty evaluation and
integration with other established monitoring
resources that have been built into ForCaMF.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Landsat time series analysis – The impact of forest ecosystem
history on biodiversity
Wanda Graf, Paul Magdon, Christoph Kleinn
Keywords: Landsat, times series, forest inventory, forest structure, biodiversity, functional diversity, scales
Forest ecosystems exhibit considerable dynamics
which manifest themselves in man-made or natural
changes which are either long-term or short term
in character (such as windfall or tree felling). The
current state of structure, composition and function
of a particular forest is the result of its ecosystem,
management and conservation history. These
dynamics are also mirrored in changes of biodiversity
patterns (Pickett & White 1985, Turner 2010).
We investigated the influence of ecosystem history
on species richness and functional diversity.
Intermediate disturbance regimes are expected
to cause highest species richness and functional
diversity according to the intermediate disturbance
hypothesis (Connell 1978). We hypothesize,
significant trends of remote sensing derived
vegetation indices point to changes in species
diversity and composition as well as functional
diversity.
This relationship of ecosystem history and
biodiversity is researched within the framework
of the “Biodiversity Exploratories” – one of the
leading long-term and large-scale projects. The
study is carried out at three sites (Schwäbische Alb,
Hainich-Dün, Schorheide-Chorin in Germany) in the
temperate mixed forest comprising 150 experimental
plots of 100 m by 100 m which feature a land use
intensity gradient from conservation to managed
forest ecosystems. For all experimental plots data
of a full census inventory of woody perennials and
herbaceous plants are available among other taxa.
Validation is done using an independent forest
inventory dataset collection, a systematic grid
covering the entire study area.
For exploring the ecosystem history of the
experimental plots we used a time series of Landsat
5, 7 and 8 from 1984 to 2016 of over 3000 images. We
combined Landsat imageries from the archives of the
United States Geological Survey and European Space
Agency to gain dense and continuous time series.
We analyze changes in trend (systematic change
over time) and disturbance events for the time
series of the Normalized Difference Vegetation
Index of Landsat surface reflectance imageries for
each experimental plot. Surface reflectance of the
raw data is calculated with the LEDAPS software
(Masek et al. 2013). Among others, we use the MannKendall-Trend test (Mann 1945) for trend analysis
and segmentation algorithms to detect disturbance
events. First analyses show that disturbance events
and trends can be detected in the time series by
using these methods. Moreover, long and dense
time series as used in this study are required due to
the low management intensity and stable nature of
these temperate forests. It showed that ecosystem
history is related to the observed biodiversity
patterns of the experimental plots.
This work helps to better understand the temporal
development of forest ecosystem structure,
composition and function as well as its impact on
species and functional diversity. Thus, this project
delivers valuable information for the development
of forest management and conservation plans.
References
Connell, H. J., 1978. Diversity in Tropical Rain Forests
and Coral Reefs. Science, 199 (1335), 1302-1310.
Mann, H. B., 1945. Nonparametric tests against
trend. Econometrica, 13, 245–259.
Masek, J. G., Vermote E.F., Saleous N., Wolfe R.,
Hall F. G., Huemmrich F., Gao F., Kutler J., Lim,
T. K., 2013. LEDAPS Calibration, Reflectance,
Atmospheric Correction Preprocessing Code,
Version 2. Model product. Available on-line
[http://daac.ornl.gov] from Oak Ridge National
Laboratory Distributed Active Archive Center,
247
Forest Monitoring
Oak Ridge, Tennessee, U.S.A. http://dx.doi.
org/10.3334/ORNLDAAC/1146
Pickett, S. T. A., White, P. S. (Eds.), 1985. The Ecology
of Natural Disturbances and Patch Dynamics.
Academic press, New York.
Turner, M. G., 2010. Disturbance and landscape
dynamics in a changing world. Ecology, 91, 2833–
2849.
248
Universidad Mayor
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Monitoring of forests to determine different levels of change
K. Sepp1, R.G.H. Bunce1, M. Lang1, M.Villoslada1, S. Mucher2
Estonian University of Life Sciences, 1 Kreutzwaldi St, 51014 Tartu, Estonia
2
Alterra, PO Box 47 6700AA Wageningen, The Netherlands
1
Keywords: Forest habitats, Stratification, LIDAR
Since the first satellite images became available in
the 1990’s maps of land cover have been produced
at varying levels of detail. However, whilst the
sophistication of the imagery has increased the
problem of detailed evaluation of the map units has
remained.
At a strategic level satellite imagery is now routinely
used to track changes in tree cover by forest
agencies and National parks. There is no doubt that
the most recent satellites especially in conjunction
with LIDAR, can monitor regeneration, regrowth,
encroachment and gap dynamics outside canopies
but are unlikely to be able to determine changes
in ground vegetation and changes in species. The
latter are needed in the assessment of biodiversity
and the methodology that will be described in the
paper provides a solution to the problem.
The methodology that has been developed over
the last 30 years is that of classification, using
the multivariate statistical analysis technique
ISOCLUSTER on environmental data, mainly climate
and topography, usually recorded from one kilometer
squares. The squares are classified into relatively
homogeneous strata at national or international
levels. The strata can then be used to describe the
range of variation within the domain and can also be
used to select dispersed random samples for in situ
surveys of habitats, vegetation and other ecological
parameters. The initial surveillance can then be
repeated to monitor change. A long term monitoring
example is the Countryside Survey of Britain which
was originally based on 32 strata which have been
increased to 40 so that separate figures can be
produced for England, Wales and Scotland. The
survey has reported on the changes in all habitats,
including forests, as well as species composition at
ten year intervals since 1978. For example, between
1990 and 2007 the area of broadleaved woodland
increased but the number of species declined by
9.3%. The full results, including the flows between
different habitats over time, are given in www.
countrysidesurvey.org.uk. The integration between
records of forest cover and other habitats enables
planners to assess the changes that are taking place
in the entire countryside and develop appropriate
policies for conservation. Comparable studies have
been carried out in Portugal, Sweden and Northern
Ireland.
The methodology has also been used to develop an
integrated program for monitoring biodiversity in
Europe using a classification with 72 environmental
strata based on climatic and topographic data.
The methodology will be described in the paper
and is described in EBONE website (http://www.
wageningenur.nl/en/Expertise-Services/ResearchThe
Institutes/alterra/Projects/EBONE-2.htm).
framework has also been used to assess the
implications of climate change, for example showing
that Fagus sylvatica may expand northwards in
Sweden under the most likely scenarios.
Recently, a comparable classification has been
constructed for Estonia producing eight classes.
These have been validated by applying orthogonal
regression to the axes of Principal Component
Analysis of the satellite land cover map of Europe and
DECORANA on species distributions in Estonia. The
correlation coefficients were over 0.7 and were very
significant. The strata are now being used to survey
vegetation and species in Estonian habitats, such as
clear fell forests, abandoned fields and patches of
invasive species, to provide data for the modelling
of potential changes under different climate change
scenarios.
249
Universidad Mayor
Forest Monitoring
Patagonian forests under attack: increasing large-scale insect
outbreaks detected from MODIS images
R. O. Chávez1,*, S. A. Estay2, A. G. Gutiérrez3, R. Rocco1
1
Pontificia Universidad Católica de Chile, Instituto de Geografía. Valparaíso, Chile
Universidad Austral de Chile, Instituto de Ciencias Ambientales y Evolutivas. Valdivia, Chile
3
Universidad de Chile, Departamento de Ciencias Ambientales y Recursos Naturales. Santiago, Chile
2
Keywords: Patagonian forests, EVI, remote sensing, time series, anomaly
Abstract
Insect outbreaks are considered one of the major disturbances in temperate forests. These natural
events have dramatic consequences not only on industry, but also on ecosystem functioning and
biodiversity conservation. Worldwide, natural outbreak dynamics have changed in the last decades by
global climate change with consequences not fully understood. In this study, we used 16-days MODIS
EVI composites to 1) reconstruct the leaf phenology of Chilean Patagonian forests and 2) detect EVI
anomalies. Negative EVI anomalies (values below the normal range) are related to green canopy loss
caused, among other perturbations, by insect defoliators. We developed a mathematical algorithm
(R Script) to calculate the EVI annual phenological cycle and EVI anomalies at the pixel level. By
displaying pixel level anomalies, we quantified the intensity, spatial distribution, and temporal spread
of insect outbreaks for the whole Aysén Region, Southern Chile (which is about the size of England).
The analysis showed that massive outbreaks (> 20,000 hectares) occurred during the growing seasons
2008-2009, 2011-2012, and 2014-2015, showing an increasing trend on the affected area along time.
This increasing trend in outbreaks events may lead to extensive carbon loss from one of the largest
wilderness forested area in the Southern Hemisphere.
250
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Reconstructing forest changes in a fragmented landscape
of southwest France from multiple datasources: ecological
implications
P.- A. Herrault1, D. Sheeren2, M. Fauvel2, M. Paegelow3
1
Centre d’études spatiales de labiosphère (CESBIO) UPS/CNRS/IRD/CNES Toulouse, France
Université de Toulouse, INP-ENSAT, UMR 1201 DYNAFOR Av. de l’Agrobiopôle, BP 32607, Auzeville Tolosane, 31326
Castanet Tolosan cedex
3
Université de Toulouse, UTM, UMR 5602 GEODE 5, al A. Machado, 31058 Toulouse,France
2
Knowledge about changes in landscape can
contribute to better explain the current biodiversity
because of a possible time-lag in biological
responses. Past disturbances of forests were already
shown as a key driver to explain apart of the current
species richness of forest insects or plants. Historical
maps combined with image time-series data (old
aerial photographs and satellite images) are often
used to assess the temporal continuity of forests and
their spatio-temporal dynamics. However, adequate
methods are needed to recostruct automatically past
and current states of forests with their trajectory.
In this study, we present a global processing chain
to reconstruct forest cover with its changes in
a fragmented landscape of south west France.
Five spatial data sources from 1850 to 2010 were
selected to map the forest evolution: historical
maps dating from 1850 and 1900, black and white
aerial photographs acquired in 1954 and 1979,
and true color orthophotograph of 2010. First, a
new registration method based on kernel ridge
regression was proposed for georeferencing
old maps and managing their geometrical local
distortions. Then, animage-processing method was
defined to extract and vectorize the forests from the
old color map of 1850. The supervised classification
approach based on a Gaussian detector relies on a
color transformation of the map in the CIELab color
space, after a morphological filtering step applied
to remove overlapping elements (text, elevations
contour lines). An additional GEOBIA method
was performed on the orthophotgraph of 2010 to
capture the current state of forests. For the others
sources, the forests were digitized manually. In a
next step, datasets of woodlands produced for the
fifth time period were manually matched in order to
produce the corresponding relations for estimating
changes. Finally, metrics at the object and landscape
levels were computed to characterize the spatiotemporal evolutions of the forest cover. This paved
the way for measuring effects of history on current
species richness of forest hoverflies. Results show
high performances for the developed methods
and tools. Thiss tudy alsoprovides an interesting
methodological framework to drive futur estudies in
historical ecology.
251
Universidad Mayor
Forest Monitoring
Remote sensing of photosynthetic light use efficiency of
tropical ecosystems
Celio Helder Resende de Sousa 1 , Thomas Hilker 2 , Yhasmin Mendes de Moura 3
Forest Ecosystems and Society Peavy Hall 050E Oregon State University, Corvallis, OR 97330
celio.sousa@oregonstate.edu
2
Forest Ecosystems and Society Peavy Hall 231 Oregon State University, Corvallis, OR 97330
thomas.hilker@oregonstate.edu
3
Instituto Nacional de Pesquisas Espaciais - INPE Caixa Postal 515 - 12227-010 - São José dos Campos - SP, Brasil
yhas.mendes@gmail.com
1
Keywords: remote sensing, photosynthesis, photochemical reflectance index, MODIS, MAIAC, Amazon
Abstract
Tropical ecosystems play major roles in the global carbon, water and energy cycles and, as a result, in
global climate. The broad goal of this research is to monitor changes in plant physiological parameters, including status of pigments, and water use in connection with drought events in such ecosystems. Specifically, we focus on stress related changes in photosynthetic activity and monitoring of vegetation decline following major stress events by inferring light use efficiency (ε) from measurements
of reflectance from the MODIS data using the Photochemical Reflectance Index. To address our objective, we inferred light use efficiency (ε) from measurements of reflectance from the MODIS data from
2000 to 2012. Lower values of PRI were found during the driest months of the year (July, August and
September). We also conducted an exploratory analysis to assess the potential climate variables that
might drive the changes in photosynthetic activity in the Amazon. We were also able to demonstrate
close links between changes in the Photochemical Reflectance Index temperature and precipitation.
Our finding show clear seasonality of light use efficiency over tropical forests that are related to dry
and wet season cycles and correspond well to flux tower related measurements of photosynthesis. Finally, Multi-angle MODIS observations, while not optimal for measuring short term changes in ε, may
provide realistic estimates of photosynthesis over tropical regions.
252
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Satellite-based monitoring of invasive species in Central-Chile
Julian Cabezas1, Fabian Fassnacht1, Tobias Schmidt2, Birgit Kleinschmit2, Michael Foerster2
2
1
Institute for Geography and Geoecology, Karlsruhe Institute of Technology (KIT)
Institute for Landscape Architecture and Environmental Planning, Technical University of Berlin
Keywords: Disturbance detection, Nothofagus forest, Landsat, BFAST, Pinus radiate, Ulex Europaeus,
Acacia dealbata.
Chile has a large number of endemic species due
to its isolated location and is therefore one of the
biodiversity hots-spots of the Planet. At the same
time a number of invasive species occurred over the
last years and have shown to have a notable negative
effect on Chilean forest ecosystems. Between 2016
and 2018, the project SaMovar (Satellite-based
monitoring of invasive species in central-Chile) will
investigate the past and recent spread of selected
invasive species.
Most invasive species are ruderal species/strategists
and are strongly adapted to certain kinds of
disturbances. Therefore, one focus within the project
will be the reconstruction of the disturbance history
of natural vegetation areas of three administrational
regions in central Chile (Maule, Biobio, Isla Chiloe
in Los Lagos). The targeted time-period for the
analysis will be 1985-2016. Based on time-series
analysis using Landsat, MODIS and Copernicus data
as well as recent and historic reference information,
disturbance occurrences and the type of disturbance
will be recorded. The mapping of disturbances will
base on the creation of a vegetation mask followed
by disturbance mapping based on LandTrendr and
BFAST. After the mapping of the disturbances, each
disturbed area will be assigned to one of the most
common disturbance agents in the study areas.
These include forest fires, clear-cutting and biotic
agents. Methodically, the assignment to disturbance
agents will base on the analysis of the reflectance
signal and the shape of the disturbed area. The
development of the signal over the first few years
after the disturbance will also be considered as
additional information.
The disturbance history will subsequently be used
as one input layer to species distribution models
which will be developed to model the future spread
of Pinus radiata, Ulex europaeus and Acacia dealbata
assuming different climate scenarios. These three
non-native woody species have been observed to
have a prominent impact on the natural vegetation
of central Chile by invading non-managed areas that
often suffered from disturbances briefly before the
invasion. Specifically in the Maule Region. Natural
Nothofagus forest is being severely invaded by Pinus
Radiata, especially when the forest fragments are
surrounded by pine plantations or when the forests
are disturbed, creating opening in the vegetation.
The project thereby methodically targets on 1) the
adaptation of the methodical state-of-the-art in
remote sensing based mapping of invasive species
to the new data delivered by Copernicus and 2) the
detection and classification of disturbances as an
input to improve the identification and modelling
of invasion dynamics. The scientific interest mainly
lies in an increased understanding of the invasion
dynamics of the three target species in a highly
heterogeneous landscape that has undergone
drastic land-use changes over the last few decades.
Preliminary results of the project are presented,
especially regarding the detections of disturbances
(fires, harvesting) in the forest ecosystems of the
Maule Region using the BFAST algorithm applied to
the available Landsat archive.
253
Forest Monitoring
Universidad Mayor
Spectral manifestation and signal to noise ratio of forest
disturbance and recovery
Zhiqiang Yang, Department of Forest Ecosystem and Society, Oregon State University, Corvallis, OR
Warren Cohen, USDA Forest Service, PNW Research Station, Corvallis, OR
Sean Healey, USDA Forest Service, Rocky Mountain Research Station, Ogden, UT
Noel Gorelick, Google Switzerland GmbH, Zurich CH 8002
Robert Kennedy, College of Earth, Ocean, and Atmospheric Sciences, Oregon State University, Corvallis, OR
Landsat imageries have been widely used in
change detection from simple two-date image
differencing analysis to complex time series analysis.
Understanding how forest disturbance and recovery
are manifested spectrally can help making efficient
change detection with remote sensing data. In this
study, 7200 randomly selected pixel samples in
Contiguous United States were interpreted with
time series of Landsat images using piecewise linear
segmentation and all available aerial photos. For each
sample pixels, change processes including harvest,
fire, decline, recovery, and stable were recorded.
With the identified piecewise linear segments, the
noise level for the whole time series for each sample
pixels were quantified for all the 6 raw Landsat bands
and a series of vegetation indices (Tasseled Cap
254
Brightness, Greenness, Wetness, NBR, NDVI, NDMI)
The spectral signals for all identified disturbance
were calculated as the segment change in the bands
and indices. Both signal and noise varies greatly
by bands and spectral indices. Visible bands have
lowest level of noise (< 0.01) and also have lowest
level signal (<0.03). Shortwave infrared band have
relative large noise and signal, resulting in higher
signal to noise ratio. In this presentation, we present
how signal to noise ratio varies by disturbance types
in different spectral bands and indices (2.23 – 16.68).
Based on correlations among signal to noise ratio in
spectral indices, we present an application of using
signal to noise ratio information for forest change
detection for Contiguous United States.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Terra-i: A Pantropical Near Real Time Monitoring System for
Vegetal Cover Change
1. Louis Reymondin, Terra-i team leader and the system’s chief architect, 2. Paula Paz, research
assistant, Presenter, 3. Oscar Bautista, research assistant, 4. Jhon Tello, systems analyst
Keywords: Deforestation, near real time monitoring, sensing remote, pan-tropically, neural network.
Abstract
Terra-i is a system for near real time vegetal cover change monitoring using remote sensing and
data mining. It aims to detect vegetation loss resulting from human activities at pantropical level
and provide up-to-date information about vegetation status with a spatial resolution and frequency,
250 meters and every 16 days, relevant for decision markers. Terra-i is developed by the International
Centre for tropical Agriculture CIAT, University of Applied Sciences Western Switzerland HEIG-VD and
King´s College London and is currently funded by Global Forest Watch (GFW) and CGIAR consortium.
Terra-i data are available free of charge on www.terra-i.org. Data are also available in other platforms
as Terra-i is a core partner of Global Forest Watch.
The methodology is based on the premise that natural vegetation follows a predictable pattern of
changes in greenness from one date to the next brought about by site-specific characteristics and
climatic conditions in the preceding days. We use a Bayesian-probability based neural network to
learn how the greenness of a given pixel (MODIS-MOD13Q1) responds to a unit of rainfall (TRMM/
GPM), then apply the model to identify anomalies in the time series which can be attributed to human
activities.
The tool has been applied as an official early warning system for land cover and land-use change in
Peru through the collaborative framework agreement signed between the International Center of
Tropical Agriculture (CIAT) and the Peruvian Ministry of Environment (MINAM). Finally, the most
outstanding publication based on Terra-i data is a paper published in Science in January 2014: Drug
Policy as Conservation Policy: Narco-Deforestation. In this paper, a team led by Kendra McSweeney
used Terra-i data to show the relationship between drug trafficking and deforestation in eastern
Honduras.
The focus of this talk will be on the Terra-i methodology and its potential uses to inform decision
making.
About the Author
Paula Paz is a Topographical Engineer from
Universidad del Valle, Colombia with understanding
and general management of GIS and remote
sensing software tools. Paula is currently working
as research assistant at CIAT working on the Terra-i
project since three years ago. Her main tasks include
the downloading, processing and post-processing
of data and information for Terra-i to support
the monitoring of deforestation in the tropics.
Additionally Paula has been involved in projects
with the Peruvian government, and has participated
of fieldwork to validate of Terra-i data and know
the main drivers of deforestation in the Peruvian
Amazon.
255
Universidad Mayor
Forest Monitoring
Time series of Landsat images to determine burned area in
the context of the Latin American Network of Forest Fires
(RedLatIF).
Jesús A. Anaya1, Armando M. Rodriguez2, Walter Sione3
Universidad de Medellín. Colombia. Email: janaya@udem.edu.co
Fundación Amigos de la Naturaleza - Universidad Autónoma Gabriel R. Moreno. Bolivia.
3
Universidad Autónoma de Entre Ríos – Universidad Nacional de Luján. Argentina.
1
2
Keywords: Forest fires, time series, Landsat, GEE, burned area.
Abstract
Time series of the Normalized Burn Ratio (NBR) calculated from Landsat data were analyzed at
different ecoregions of Colombia and Argentina. The process included removing clouds and cloud
shadows from Landsat imagery and masking areas were vegetation was removed without evidence of
charcoal. Changes from vegetation to bare soils are usually classified as burned areas and therefore
validated as commission errors. Temporal change from prefire to postfire (dNBR) was evaluated using
three methods: i) The difference between the prefire and postfire normalized burned radio (dNBR)
(Key & Benson 2005); ii) The relative form of dNBR (Miller & Thode 2007), and iii) The dNBR where
a composite of the maximum reflectance was used as input for prefire. Each method generated a
surface of dNBR and different thresholds were used to classify each dNBR pixel into burned and not
burned. Metrics as minimum sum of errors and dice coefficient (Padilla et al. 2014) were considered
to define the optimum threshold. The analysis of burned area time series allow identifying the cause of
errors and generating prefire statistics. It was found that there is a pattern towards large commission
errors when trying to minimize omission and no difference was found from methods i) and ii). Time
series are important to replace prefire pixles contaminated by clouds, but other errors arise depending
on the metric selected from the time series.
Introduction
Large discrepancies have been found for burned
area (BA) estimation using remote sensing images
with data sets of medium resolution (500 m – 1
km) as GBA2000, GlobScar, WFA, (Boschetti et
al. 2004) or more recently MCD64A1, MCD45A1,
GEOLADN-2 (Moreno et al. 2014). Several aspects
of remote sensing technology such as spatial and
temporal resolution are identified as responsible
for inaccuracy (Anaya & Chuvieco Salinero 2012).
Indexes derived from spectral bands are commonly
used to enhance the detection of BA (Chuvieco et al.
2002; González et al. 2007; Roy et al. 2008). A simple
index is the single date of the Normalized Burn Ratio
based on NIR and SWIR (Key & Benson 2006) used
to determine fire severity. However, the detection
256
improves when the difference between two dates is
considered, one date to determine prefire NBR and
other date for post fire NBR.
The method to highlight areas affected by fire and
the selection of a threshold to classify BA (limit
between burned and unburned pixels) is challenging
because the large amount of factors affecting prefire
and post fire reflectance. Some of these factors are
the amount of vegetation, the type of vegetation
and the composition and color of soils (e.g. black
soils vs. white soils). By the other hand, there are
changes not associated to the effect of fire, such
as, illuminations and atmospheric differences in
the pre and post fire imagery, changes inherent to
vegetation phenology and climate conditions. The
sum of these factors generates unique per pixel
Universidad Mayor
ForestSAT 2016 Abstracts Summary
conditions. The present work aims to decrease the
uncertainty generated by changes in reflectance at
TOA before and after the fire using NBR and time
series of Landsat data.
Methodology
Two values of the NBR were calculated, one for the
charcoal signal (postfire) and one for vegetation
(prefire). The difference between these NBR values
was calculated for every pixel of a Landsat image, for
this reason, the result is a surface with dNBR values
(difference of the normalized burn ratio index).
Reflectance calculated on top of the atmosphere
(TOA) was used to calculate NBR; in this phase
clouds, shadows and water bodies were removed.
Two masks were used to remove the effect of clouds,
one for the large reflectance of the aerosol and the
other to remove the low reflectivity of their shadow.
Clouds were masked with the built-in algorithm
of GEE “simple cloud score” and shadows were
removed with a threshold for SWIR1. The last preprocessing step was to identify and remove water
bodies with the Normalized Difference Water Index
(NDWI).
Input data and BA index calculations
Landsat 5 and 8 were used for calculations based on
equation 1, 2, 3 and 4.
eq. 1
dNBR= NBRprefire – NBRpostfire (Miller & Thode
2007)
eq. 2
RdNBR= (NBRprefire – NBRpostfire)/NBRprefire
eq. 3
dNBRmax= NBRprefire(max) – NBRpostfire eq. 4
Where,
SWIR: 2.1-2.3 μm
RdNBR: Relative difference of NBR
NBRprefire(max): maximum NBR from prefire time
series
The RdNBR has been considered important when
NBR values are low at prefire conditions (eq. 3).
For instance, when the amount of vegetation is
low, dispersed (with high soil signal) or dry, which
is very likely in pastures during the dry season. The
maximum value of NBR for prefire conditions (eq. 4)
where selected to evaluate the effect of removing
clouds using time series. Pixels contaminated
with clouds are expected to result in low NBR
values. Three dNBR surfaces where generate using
equations 2-4. Then each surface was classified
into multiple BA maps using thresholds increments
of 0.1. When dNBR 0 it is assumed that the pixel
was not affected by fire. Each BA map generated
by this semi-automatic method was compared to
the reference information to generate accuracy
statistics in order to select the best method and
optimal threshold.
Results
The extent and size of BA reference information
varied notably. The average size of BA polygons for
the images of Colombia and Argentina was 4 and 132
ha respectively. With a highly fragmented landscape
in Colombia (801 polygons) when compared to
Argentina (276 polygons). NBR values behave very
similar to NDVI values, with high NBR values at
prefire, healthy vegetation, and low values after fire.
Figure 1, shows three land cover types of Colombia:
grassland, Shrubland (regrowth) and mature
amazon forests. The time series show low values of
NBR at date 23, when BA reference information was
gathered. The NBR values for bare soil and burned
areas are very similar in prefire and posfire, for this
reason bare soils are likely to be classified as burned
areas by the semi automatic method. Other aspect
revealed by the time series analysis is the variability
of NBR in the Forest class due to phenology and
remote sensing artifacts, with a minimum NBR of
0.58 and a maximum of 0.72. This variability is the
cause of commission at low threshold values.
The error of commission increases as the BA
identification increases. In this regard, it should be
noted that the metrics to select the best method
should include both, low errors and large BA
agreement. Selecting thresholds associated to low
sum of errors (commission and omission) heavily
influenced by commission will result in extremely
low values of BA detection.
No significant differences in accuracy were found
when using the relative form of dNBR (equation
3), the best identification of BA with the lowest
commission error was found when using equation
4. These preliminary results indicate that time series
257
Universidad Mayor
Forest Monitoring
Figure 1. Time series of prefire NBR values and other classes of land cover.
may improve BA classification in areas with heavy
cloud content, i.e. when prefire pixels contaminated
by clouds are replaced by a time series metric.
Conclusions
The main limitation was the availability of
consecutive images with low extent of cloud content.
The best burned area map was derived by including
soils mask and the maximum composite as input for
prefire NBR. This case study has been implemented
on the Google Earth Engine platform which allows
access to a vast amount of satellite imagery and
tools for algorithm development (Patel et al. 2015).
This platform was very valuable for the cooperation
and interaction among the Latin American Network
of Forest Fires (RedLatIF) members.
References
Anaya, J.A., Chuvieco Salinero, E., (2012). Accuracy
assessment of burned area products in the
258
Orinoco basin. Photogrammetric Engineering
and Remote Sensing 78, 53-60.
Boschetti, L., Eva, H.D., Brivio, P.A., Gregoire,
J.M., (2004). Lessons to be learned from the
comparison of three satellite-derived biomass
burning products. Geophysical research letters
31(L21501), doi:10.1029/2004GL021229.
Chuvieco, E., Martin, M.P., Palacios, A., (2002).
Assessment of different spectral indices in the
red-near-infrared spectral domain for burned
land discrimination. International Journal of
Remote Sensing 23, 5103-5110.
González, A.F., S., M.d.M., Cuevas, G.M., (2007).
Un nuevo algoritmo para la cartografía de áreas
quemadas a partir de información NIR, SWIR y
TIR. Revista de Teledetección, 97-105.
Key, C., Benson, N., (2005). Landscape assessment:
Ground measure of severity; the Composite
Burn Index, and remote sensing of severity,
the Normalized Burn Index, in: D. Lutes, R.K.,
J. Caratti, C. Key, N. Benson, S. Sutherland, L.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Gangi (Ed.), FIREMON: Fire effects monitoring
and inventory system. USDA Forest Service, pp.
1-51.
Key, C., Benson, N., (2006). Landscape Assessment
(LA) Sampling and Analysis Method, in: Rep.,
U.F.S.G.T. (Ed.), RMRS-GTR-164-CD, p. 51.
Miller, J.D., Thode, A.E., (2007). Quantifying burn
severity in a heterogeneous landscape with a
relative version of the delta Normalized Burn
Ratio (dNBR). Remote Sensing of Environment
109, 66-80.
Moreno, J.A., Garcia, J.R., del Águila, I., Hernández,
P., (2014). Burned Area Mapping in the North
American Boreal Forest Using Terra-MODIS LTDR
(2001–2011): A Comparison with the MCD45A1,
MCD64A1 and BA GEOLAND-2 Products.
Remote Sensing 6, 815-840.
Padilla, M., Stehman, S.V., Chuvieco, E., (2014).
Validation of the 2008 MODIS-MCD45 global
burned area product using stratified random
sampling. Remote Sensing of Environment 144,
187-196.
Roy, D.P., Boschetti, L., Justice, C.O., Ju, J., (2008).
The collection 5 MODIS burned area product
-- Global evaluation by comparison with the
MODIS active fire product. Remote Sensing of
Environment 112, 3690-3707.
259
Forest Monitoring
Universidad Mayor
Understanding the large area disturbance history of Australian
Sclerophyll forests
Simon Jones 13*, Mariela Soto-Berelov 13, Samuel Hislop 13, Trung Nguyen 13, Salahuddin Ahmad 12, Shirley Famelli 1,
Andrew Haywood 1234
1
Centre for Remote Sensing, School of Mathematical and Geospatial Sciences, RMIT University, Melbourne, Australia
2
Department of Environment, Land Water and Planning, East Melbourne, 3002, Victoria, Australia
3
Coopertive Research Centre for Spatial Information, Carlton, 3053, Victoria, Australia
4
European Forestry Institute, Kuala Lumpur, Malaysia
Keywords: remote sensing, woody attribution, disturbance history, feature extraction.
Abstract
This paper presents a methodology for attributing and characterising the disturbance history of Sclerophyll forested landscapes over large areas. First we define a set of woody vegetation data primitives
(e.g. canopy cover, leaf area index (LAI), bole density, canopy height), which are then scaled-up using
multiple remote sensing data sources to characterise woody vegetation features. The advantage of
this approach is that vegetation landscape features can be described from composites of these data
primitives. The proposed data primitives act as building blocks for the re-creation of past woody characterisation schemes as well as allowing for re-compilation to support present and future policy and
management and decision making needs.
A large public land manager –the Victorian Department of Environment, Land Water and Planning set
up 786 permanent field plots (VFMP), sampled on a stratified systematic random framework, to measure woody and non-woody forest data primitives. Results are presented of data primitives including:
1) LAI –validated using ground based hemispherical photography, LAI 2200 PCA, CI-110 and terrestrial
and airborne laser scanners; 2) canopy height and vertical canopy complexity derived from airborne
LiDAR validated using ground plot observations; and, 3) time-series characterisation of disturbance
history and associated land cover change using LandTrendr analysis of Landsat LEDAPS data. kNN is
then used to link the derived features to the VFMP field plot sites.
260
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Use of New Technologies in Monitoring Mountain
Forests’ Condition
Radomir Bałazy 1, Mariusz Ciesielski 1, Tomasz Hycza 1, Patryk Waraksa 1
1
Forest Research Institute, Braci Lesnej 3 Street, 05-090 Raszyn
Keywords: LiDAR, satellite images, ALS, GIS, mountains, monitoring, remote sensing, deforestation
Mountain ecosystems, though very abundant
and diverse, are extremely difficult to monitor.
An obstacle during research may not only be
large differences in altitudes, but also completely
different climatic conditions on various aspects and
mountain slopes. It is in such conditions that remote
sensing techniques are particularly applicable, which
enable fast, precise, though not necessarily cheap,
monitoring of the occurring processes.
Upon the request of General Directorate of The
State Forests, the greatest monitoring of mountain
ecosystems has been carried out in the south of
Poland since 2012, which is based, among others,
on data obtained from airborne laser scanning
(ALS), satellite imaging repeated three times
within every growing season (Black Bridge), aerial
photographs, and ground measurements including
terrain laser scanning (TLS). The application of
all those technologies at one place and at a time
ensures unparallel possibilities of observing the
occurring processes. The project named “Creation
of a Forest Information System Covering the Areas
of the Sudety and the Western Beskidy Mountains
within the Scope of Forest Condition Monitoring and
Assessment” is at the same time the greatest GIS
project at The State Forests, which has been awarded
an ESRI prize in 2014 for exceptional achievements
in the scope of GIS.
The purpose of the presented project, apart
from the development of a modern GIS system,
is mainly monitoring of mountain deforestation
process, and an attempt to understand complicated
dependencies which influence this process. Several
dozen of various analyses have been carried out so
far for whole forest stands, as well as for single trees
(based on LiDAR data), which have demonstrated
many interesting relationships. The presentation
will include the results of selected analyses, among
others:
● The impact of topography on deforestation
process based on the analysis of a detailed terrain
model (DTM).
● The dependence of spruce growth (based on
ALS measurements of hundreds of thousands of
individual trees) on climate conditions, performed
treatments and soil in the context of their health
condition in the past and today.
● Mechanics
of
mountains’
deforestation
processes, their directions, and processes’
accompanying dependencies based on ALS data
and satellite imaging.
● Relations between selected physical and chemical
features of soils and assimilation apparatus
based on over 600 circle areas in the context of
topography and observed deforestations.
261
Universidad Mayor
Forest Monitoring
POSTER
Using historical satellite time-series to test an hypothesis of
forest susceptibility to the bark beetle outbreak
Martin Hais1,6*, Jan Wild2,5, Luděk Berec3, Josef Brůna2, Robert Kennedy4, Justin Braaten4,
Kateřina Hellebrandová6, Zdeněk Brož1
Department of Ecosystem Biology, Faculty of Science, University of South Bohemia, Branišovská 31,
37005 České Budějovice, Czech Republic
2
Institute of Botany, The Czech Academy of Sciences, CZ-252 43, Průhonice, Czech Republic
3
Department of Ecology, Institute of Entomology, Biology Centre CAS, Branišovská 31, 37005
České Budějovice, Czech Republic
4
Department of Forest Ecosystems and Society, Oregon State University, 321 Richardson Hall,
Corvallis, OR 97331, United States
5
Faculty of Environmental Sciences, Czech University of Life Sciences Prague, Kamýcká 129, Praha 6 – Suchdol,
CZ-165 21, Czech Republic
6
Department of Forest Ecology, Forestry and Game Management Research Institute, Strnady 136, 252 02 Jíloviště
1
Forest disturbances significantly influence the
character of forest ecosystems, and many studies
have focused on disturbance or post-disturbance
processes. However, few studies have investigated
whether satellite data can provide insight into
precursor conditions for disturbance in forests.
Based on experience from the Bavarian Forest
and other temperate spruce forests we modelled
risk of bark beetle attack based on commonly
used environmental predictors together with predisturbance spectral trajectories from Landsat
Thematic Mapper (TM) imagery. Our study area is
located in the central part of the Šumava Mountains,
in the border region between the Czech Republic
and Germany, Central Europe. The areas of bark
beetle attacked forest were delineated from aerial
photographs taken in 1991 and every year in the
period 1994 -2000. The environmental predictors
represent forest stand attributes (e.g. tree density or
distance to the attacked forest from previous year)
and common abiotic factors like topography-climate
variables or geological and pedological background.
Pre-disturbance spectral trajectories were defined
by the Tasseled Cap linear transformation (Wetness,
Brightness) calculated from 16 Landsat TM scenes
for the period 1984 – 1999. As pre-disturbance
262
spectral trajectory we consider a slope of linear
regression of either Wetness or Brightness index
related to years from 1984 until one year before
the bark beetle attack. Using logistic regression
and multimodel inference, we calculated predictive
models separately for any single year in 1994 – 2000
to account for a possible shift in importance of
individual predictors during disturbance. Inclusion
of pre-disturbance spectral trajectories (Wetness
slope and Brightness slope) improved predictive
ability of bark beetle attack models. Wetness slope
appeared most prominent even in comparison to
environmental predictors and relatively stable over
the years, while Brightness slope improved the
model only in the middle of the disturbance (1996).
Moreover, these pre-disturbance predictors were
not correlated with other environmental predictors
and are therefore able to explain additional data
variability. Predictive power of our fitted models
expectedly decreased with further in the future, but
this trend was slower at the beginning of disturbance.
The pre-disturbance spectral trajectories are
valuable not only for assessing the risk of bark beetle
attack, but also for detection of long-term gradual
changes even in non-forest ecosystems.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Validating a Forest Canopy Disturbance Map in
North Central USA
Mark D. Nelson (corresponding author)
U.S. Department of Agriculture, Forest Service, Northern Research Station, 1992 Folwell Avenue, St. Paul, MN 55108,
USA, phone 651-649-5104, e-mail mdnelson@fs.fed.us
Brian G. Tavernia
The Nature Conservancy, Colorado Field Office, 2424 Spruce Street, Boulder, CO, 80302, USA, phone 720-974 7014, e-mail
brian.tavernia@tnc.org
James D. Garner
U.S. Department of Agriculture, Forest Service, Northern Research Station, 1992 Folwell Avenue, St. Paul, MN 55108,
USA, phone 651-649-5107, e-mail jamesdgarner@fs.fed.us
Stephen V. Stehman
State University of New York, College of Environmental Science and Forestry, 322 Bray Hall, 1 Forestry Drive, Syracuse,
NY, 13210, USA, phone 351-470-6692, e-mail svstehma@syr.edu
Charles H. Perry
U.S. Department of Agriculture, Forest Service, Northern Research Station, 1992 Folwell Avenue, St. Paul, MN 55108,
USA, phone 651-649-5191, e-mail charleshperry@fs.fed.us
Keywords: forest canopy disturbance, early successional forest, Landsat Time Series Stack,
accuracy assessment
Abstract
The aim of this study was to produce a geospatial dataset of early successional forest (ESF; age 1-20
years) and other land cover classes, and to conduct a comprehensive and robust suite of map validation
procedures for estimating accuracy and assessing utility of the dataset. We used summer Landsat time
series stacks (LTSS) and a vegetation change tracker algorithm (VCT; Huang et al. 2010) to classify
forest canopy disturbance during 1990-2009 in Michigan, Wisconsin, and Minnesota, USA. To reduce
commission of ESF we incorporated LTSS of snow covered winter imagery (VCTw; Stueve et al. 2011).
Stand age was inferred from year of canopy disturbance and binned into four 5-year age classes that
were assigned to forest pixels of the U.S. National Land Cover Database of 2011 (Homer et al. 2015). To
reduce omission of ESF, we reclassified a subset of NLCD2011 Shrub/Scrub and Grassland/Herbaceous
class to forest where corresponding VCTw pixels were classed as forest (Garner et al. 2016). Land use
and stand age data from U.S. Department of Agriculture Forest Service, Forest Inventory and Analysis
(FIA) plots (n=27,219) were used to produce estimates of map accuracy and post-stratified estimates of
area following “good practices” in Olofsson et al. (2014), which were augmented by additional site- and
non-site specific validation assessments. Reported estimates are accompanied by ±95% confidence
intervals. Overall accuracy of generalized land cover classes (forest/nonforest/water) increased
relative to NLCD2011 from 87.6% (±0.39%) to 89.5% (±0.36%) after reclassifying about 11,000 km2
of NLCD2011 nonforest classes to forest. We estimated 216,563 (±1,532) km2 of forest land area, of
which 26,635 (±1,134) km2 was disturbed within the past 20 years. About 12% of total forest area was
ESF; 8.5% in Michigan, 10.9% in Wisconsin, and 16.2% in Minnesota. Overall accuracy for seven classes
(1-5 year old, 6-10 year old, 11-15 year old, 16-20 year old, persisting forest over 20 years, nonforest,
and water), was 84.9% (± 0.42%). Two thirds of the error was attributed to allocation disagreement
and one third to quantity disagreement (Pontius and Millones 2011). Producer’s accuracies ranged
from 15.0% (± 1.89%) for 16-20 year old forest to 91.5% (± 0.45%) for persisting forest class; user’s
accuracies ranged from 38.8% (± 6.28%) for 1-5 year old forest to 90.4% (± 0.52%) for nonforest class.
263
Forest Monitoring
Universidad Mayor
Overall accuracy increased insignificantly to 85.6% (± 0.42%) after aggregating the four 5-year age
classes into a single 20-year age class. Accuracies for individual 5-year age classes ranged from 15-31%
producer’s accuracy to 39-45% user’s accuracy, changing to 31.0% (± 1.46%) producer’s accuracy and
62.4% (± 3.26%) user’s accuracy after aggregating the four 5-year age classes into a single 20-year age
class. Similarly, accuracies for 5-year age classes improved to 24-36% (producer’s accuracy) and 5468% (user’s accuracy) under a fuzzy classification assessment approach whereby neighboring forest
age classes in the error matrix were also counted as correctly classified (Gopal and Woodcock 1994).
Northwestern subregions had lowest overall accuracies but highest producer’s and user’s accuracies
for ESF age classes. Estimates within equal-area polygons created using a space-filling curve procedure
(Lister and Scott 2009) revealed spatial clustering of omission errors. Employing multiple validation
assessments across multiple strata improved our understanding of data quality, utility, causes of
misclassification, and limitations of validation data.
264
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Water yield dynamics in forested watersheds: Using Landsat
annual time series for the assessment of eco services in
national forests of the Intermountain West, USA
Alexander J. Hernandez1, Sean P. Healey2, R. Douglas Ramsey1
1
Utah State University
United States Forest Service
2
Keywords: Landsat Time Series; Water Yield, Disturbance; SWAT; Forested Watersheds
Changes in forest structure invariably impact the
dynamics of the water balance of a given region.
Interception,
infiltration,
evapotranspiration,
and runoff are promptly affected after sudden
disturbances (i.e. harvests, fires) that eliminate
or reduce canopy cover. Assessment of long-term
water effects of disturbance at the catchment
level is limited by the lack of continuous, wallto-wall information about disturbance patterns.
Here, we present results of combining highlydetailed disturbance datasets in conjunction with
a physically-based hydrologic model to report
about water dynamics in forested landscapes. The
US National Forest System (NFS) has developed
a comprehensive plan for carbon monitoring that
requires a detailed temporal mapping of forest
disturbances across 75 million hectares. A longterm annual time series that shows the timing,
extent, and type of disturbance beginning in 1990
and ending in 2011 has been prepared for the entire
NFS using a mix of automated and manual mapping
processes. Maps also showed the magnitude of
each change event, which was modeled in terms of
inter-annual canopy cover loss. This information was
used as input to the Soil and Water Assessment Tool
SWAT to model water dynamics, with an emphasis
in water yield and sediment concentration, and
thus gain empirical insight into how disturbance
may predispose changes in water at a given outlet
in forested watersheds. We developed a hydrologic
baseline (business as usual) that included longterm measured climatic variables, topographic
relief, soil hydrologic groups, and the National
Land Cover Dataset NLCD for the year 1992, 2001,
2006, and 2011. Then, on a yearly basis from 1990,
we included our disturbance datasets in the SWAT
model to augment the spatiotemporal variability
of the land cover input, and modeled streamflow
and sediment loads for every year thereafter. This
permitted the comparison between undisturbed
and disturbed scenarios across the study area.
SWAT was calibrated in watersheds where monthly
hydrometric records were available. This allowed
the estimation of the model efficiency, and thus
report on the model uncertainty. The thematic,
spatial, and temporal resolution of our disturbance
maps allows the estimation of the relative impact
of disturbance through time in these watersheds.
Our results provide detailed transparent accounts of
water yield dynamics, and provide opportunities to
assess trends in ungauged watersheds with similar
ecological conditions. Water resources are subject to
numerous stress agents such as global change, and
climatic variability. Our research here quantifies the
effects of natural and anthropogenic disturbance on
water dynamics that can be used for managers and
stakeholders to plan for the future.
265
Forest Monitoring
266
Universidad Mayor
Forested wetland monitoring
Universidad Mayor
ForestSAT 2016 Abstracts Summary
3D mapping of mangrove forests along the Pacific Coasts of
Central and South America
Marc Simard1, Victor H. Rivera-Monroy2, Edward Castañeda2, Michael Denbina1, Hector Tavera Escobar3, Mireya Cajas
Pozo4, Mariko Burgin1
California Institute of Technology-Jet Propulsion Laboratory, Pasadena, California, USA
2
Louisiana State University, Baton Rouge, Louisiana, USA
3
Fundación Mar Viva, Bogotà, Colombia
4
Escuela Superior del Litoral, Guayaquil, Ecuador
1
Keywords: Mangroves, forests, interferometry, radar, polinsar, height, biomass
Mangrove forests are some of the most productive
ecosystems on Earth. In addition to extraordinary
carbon sequestration potential, they provide a
wealth of socio-economic resources to coastal
communities. Thus, it is imperative to generate
baseline maps to assess their status and vulnerability
to climate change and human activity. We present
novel approaches to map mangrove canopy height
as well as above ground biomass along the Pacific
coasts of central and south America. Mangroves
of these regions span a wide variety of structural
characteristics and socio-economic contexts,
providing a robust assessment of the remote sensing
capabilities.
We used several airborne and space-borne
instruments. In particular, we use a technique called
polinSAR (Polarimetric Interfometric Synthetic
Aperture Radar) to map mangrove forests in 3D.
In 2013 and 2015, NASA’s UAVSAR system, an
airborne L-band fully polarimetric radar enabled for
repeat-pass interferometry, collected data along
the coasts of Costa Rica, Colombia and Ecuador.
Using these data, we have developed an adaptive
temporal decorrelation algorithm to compensate
for changes in environmental moisture and wind
conditions between UAVSAR flights. To validate
the UAVSAR estimates of canopy height, we also
collected coincident in situ data in the Osa Peninsula,
the department of Chocó and the Guayas estuary.
Furthermore, we used data from Shuttle Radar
Topography Mission (SRTM) and the German’s
TanDEM-X spaceborne instruments to estimate
canopy height and cross-validate UAVSAR estimates
of canopy height. We obtained height uncertainties
in the order of 2m over a 20m pixels. The in situ
data was also used to derive canopy height-biomass
allometric equations to convert canopy height into
above ground biomass maps.
Interestingly, comparison of the new canopy height
maps from the UAVSAR and TanDEM-X systems with
those derived from the Shuttle Radar Topography
(SRTM) acquired in 2000 show strikingly rapid
changes in mangrove forest structure. Some forests
grew nearly 15 meters in the time interval. We also
found the location of the tallest mangrove of the
Americas (to be announced during the presentation).
Upcoming spaceborne missions with interferometric
capabilities (e.g. NISAR, SAOCOM and BIOMASS)
will provide further opportunities to monitor
mangrove structure and health. Thus it is crucial to
further develop technique adapted to repeat-pass
polinSAR and mangrove forests.
269
Universidad Mayor
Development of Methods
Finding the DAM signal: Utilizing time series of all
available landsat TM/ETM+ observations to map and
monitor beaver-related flooding events
Valerie J. Pasquarella1,2 and Curtis E. Woodcock3
Department of Environmental Conservation, University of Massachusetts Amherst, 160 Holdsworth Way,
Amherst, MA 01003; email: valpasq@umass.edu
2
Northeast Climate Science Center, University of Massachusetts, 233 Morrill Science Center, 611 North Pleasant
Street, Amherst, MA 01003
3
Department of Earth and Environment, Boston University, 675 Commonwealth Ave., Boston, MA 02215
1
Keywords: Landsat time series, forested wetlands, flood events, disturbance monitoring
Abstract
Multi-temporal imagery from Landsat family of satellites has been widely used to study a variety
of change processes common to upland forests, including fire, wind, harvest, and insect damage;
however, there have been relatively few studies using Landsat data to specifically map and monitor
beaver activity in lowland areas. In this study, time series of all high quality Landsat TM/ETM+
observations from 1985-2014 were used to investigate the spectral-temporal signatures of beaverrelated disturbances in Massachusetts, USA. We found evidence of a distinct spectral-temporal flood
response consisting of a decrease in mean annual Tasseled Cap Brightness occurring concurrently with
a decrease in the seasonal variability of Tasseled Cap Greenness. A targeted change detection algorithm
was developed based on this observed response and the algorithm was tested on reference datasets.
Results suggest that the timing of detected flood events matches well with records of beaver activity,
and the algorithm is flexible enough to identify flood events of different durations and magnitudes.
As North American beaver populations continue to expand, dense time series of Landsat observations
provide a new source of information for monitoring beaver activity over large spatial extents.
Introduction
North American beaver (Castor canadensis)
populations, once decimated by overhunting, are
returning to the landscapes of the conterminous
United States and reclaiming their role as
continental-scale agents of change. Like humans,
beavers are classic ecosystem engineers, having
a disproportionately large impact on their
environments and habitats beyond basic needs
for food and shelter (Naiman et al. 1988; Rosell
et al. 2005). Beaver dams alter the physical and
chemical properties of streams, changing currents,
temperature, and rates of sedimentation (Butler
& Malanson 2005), and this rapid alteration of the
abiotic environment often has dramatic impact
on riparian and aquatic habitats. Therefore,
understanding the spatial and temporal dynamics of
beaver activity is important for future management
of the beaver population, as well as the management
of forest and wetland ecosystems modified by
270
beaver activity.
Multi-temporal imagery from Landsat family of
satellites is widely used to study a variety of other
land cover change processes common to North
American landscapes, including fire, wind, harvest,
and insect damage (Cohen & Goward 2004; Wulder
et al. 2012). Flooding is often recognized as a driver
of vegetation change (Nielsen et al. 2008; Cohen
et al. 2016), and previous Landsat-based studies of
land cover change have quantified conversions from
forest to open water (Drummond and Loveland 2010)
and among wetland types (Kayastha et al. 2012;
Fickas et al. 2015). However, most previous efforts
to quantify beaver activity at landscape scales have
relied on expert interpretation of a series of highresolution aerial photos to identify beaver-related
wetland changes, and there have been relatively few
studies using Landsat imagery to specifically map
Universidad Mayor
and monitor beaver activity (Finn and Howard 1981;
Townsend, Walsh & Butler 1995; Townsend & Butler
1996).
With free and open access to the Landsat archive
now provided by the USGS, there has been growing
interest in using Landsat time series to analyze
ecological processes (Kennedy et al. 2014; Gomez
et al. 2016). Given the widespread, potentially
continental-scale impacts of growing beaver
populations, there is a critical need to acknowledge
beavers as an agent of change similar to other
widely studied patch-altering disturbances and to
investigate how time series of remotely sensed
observations, as opposed to individual images, can be
used to automate detection of spatial and temporal
patterns of beaver activity. In this study, time series
of all high quality Landsat observations were used to
explore the spectral-temporal signature of beaver
activity and develop an algorithm that can be used
to map beaver activity over large areas.
Methodology
To investigate spectral-temporal responses to
flooding events caused by beaver activity in
Massachusetts, we utilized stacks of all available
Landsat TM and ETM+ imagery with < 80% cloud
cover. Analysis was restricted to images acquired
from January 1984 through December 2014 to
ensure that full years of data were used in the
calculation of annual statistical metrics, resulting
in a 30-year time series for each pixel. To facilitate
the interpretation of reflectance data, Landsat’s six
optical bands were transformed into three Tasseled
Cap (TC) components—Brightness (TCB), Greenness
(TCG) and Wetness (TCW) (Crist 1985).
From a change detection perspective, flooding
events associated with beaver activity typically
consist oftwo concurrent events—(1) a flood event
following initial damming and (2) changes in
vegetation cover and condition due to stress and dieoff of flood-intolerant plant species. While harvest,
fire, or insect infestation may also cause declines
in vegetation conditions, these agents of change
typically produce a post-disturbance increase in
reflectance (Schroeder et al. 2011; Senf et al. 2015),
with removal or die-off of vegetation revealing a
bright soil background. Flooding, however, would be
expected to decrease post-disturbance reflectance,
as vegetation succumbs to inundated conditions
ForestSAT 2016 Abstracts Summary
resulting in an open water or “ponded” state. Thus,
flooding events should have a spectral-temporal
trajectory that is distinct from other types of forest
change and wetland conversions. In comparing
the TCB, TCG and TCW trajectories of locations
impacted by beavers, a decrease in both TCB and
TCG was observed in all examples.
Using the observed relationship between TCB and
TCG, we developed an algorithm for identifying
beaverrelated flooding events using Landsat time
series. Mean TCB, maximum TCG, and minimum
TCG were computed for each year in the time series
using all clear observations for that year. Simple
conditional evaluations were then used to determine
whether or not a flood event had occurred (or remains
in progress) in a given year (Table 1). The first set of
conditions identifies abrupt change, as indicated by
a large, negative year-to-year change in the mean
of TCB, combined with a decrease in the seasonal
range of TCG. The second set of conditions identifies
long-term (cumulative) changes, as indicated by a
large negative decrease in the cumulative annual
difference in mean TCB combined with a decrease
in the seasonal range of TCG. Ultimately, a flood
event is flagged when the annual OR cumulative
annual difference in mean TCB from the previous
year is greater than a threshold (TCBdiff or TCBcusum)
AND the annual range of TCG in the current year
is less than some percentage of the starting range
of TCG (TCGstress). In the most general sense, this
algorithm is used to test the hypothesis that there
is a detectable, directional relationship between TCB
and TCG associated with beaver-related flooding.
Results
The proposed algorithm was tested on several sites
with a history of beaver activity, including Mass
Audubon’s Broadmoor Wildlife Sanctuary, located in
Natick, MA. Mapping flood events in the Broadmoor
wetlands by year (Figure 1) shows the timing and
extent of flooding associated with two beaver dams.
In 1990, a large area of change is detected in the
wetland upstream of Dam 1. The flooding event
caused by Dam 1 persists from 1990 through 1994,
with other potential events in 1998, 2005 and 2013.
The effects of Dam 2 first appear in 2003, with a flood
event persisting through the end of the time series.
The timing and location of flooding events is
generally in agreement with sanctuary records of
271
Universidad Mayor
Development of Methods
Table 1: General descriptions of key algorithm parameters and specific values used in this study
Threshold
Description
Value
TCBdiff
TCBcusum
TCGstress
Year-to-year change in mean TCB, captures abrupt flooding
Cumulative sum of the year-to-year change in mean TCB, captures gradual flooding
Range of TCG calculated as percent of range in year 1, captures vegetation stress
-0.05
-0.05
TCG t=1 *
(1 - 0.3)
Figure 4: Mapped results for Broadmoor, 1989-2013. Blue areas correspond to flood events detected in a
given year.
Known beaver dams are indicated by yellow circles. Landsat base image acquired 11-October 2008.
beaver
activity. The algorithm is able to detect year-toyear variability in the extent of flooded area, and is
sensitive to both short-term and longer-term events.
Furthermore, flood events are generally concentrated
in wetland areas with limited commission in upland
areas. These results suggest that flood events do
have a distinct spectral-temporal signature and can
be mapped and monitored an annual time step using
Landsat time series.
Conclusions
This study advances uses of time series of all high
quality Landsat observations to identify beaverrelated flooding events at an annual time step. We
have identified a spectral-temporal relationship that
272
distinguishes flooding from other types of change
events. To test this hypothesized spectral-temporal
response, we developed an algorithm that can be
used to detect timing, persistence and location of
flooding events. Unlike more complex segmentbased approaches (e.g. Kennedy et al. 2010; Zhu et
al. 2014; Czerwinski et al. 2014), we use year-to-year
variability in key spectral features to detect change
events at an annual time step. Testing indicates that
the timing of flood events detected by our algorithm
matches well with records of beaver activity, and the
algorithm is flexible enough to identify flood events
of different durations and magnitudes. Furthermore,
algorithm parameters are easily interpretable and
tunable, and should be generalizable to other study
areas. As the North American beaver population
continues to expand, dense time series of Landsat
Universidad Mayor
observations provide a new source of information
on the long-term large-scale impacts of these
industrious ecosystem engineers.
References
Butler, D. R., & Malanson, G. P. (2005).The geomorphic
influences of beaver dams and failures of beaver
dams. Geomorphology, 71(1-2), 48–60.
Cohen, W. B., & Goward, S. N. (2004). Landsat’s Role
in Ecological Applications of Remote Sensing.
BioScience, 54(6), 535.
Cohen, W. B., Yang, Z., Stehman, S. V., Schroeder, T.
A., Bell, D. M., Masek, J. G., et al. (2016). Forest
disturbance across the conterminous United
States from 1985–2012: The emerging dominance
of forest decline. Forest Ecology and Management,
360(C), 242–252.
Crist, E. P. (1985). A TM tasseled cap equivalent
transformation for reflectance factor data.
Remote Sensing of Environment, 17(3), 301–306.
Czerwinski, C. J., King, D. J., & Mitchell, S. W. (2010).
Mapping forest growth and decline in a temperate
mixed forest using temporal trend analysis of
Landsat imagery, 1987–2010. Remote Sensing of
Environment, 141, 188–200.
Drummond, M. A., & Loveland, T. R. (2010). Land-use
Pressure and a Transition to Forest-cover Loss in
the Eastern United States. BioScience, 60(4), 286–
298. http://doi.org/10.1525/bio.2010.60.4.7
Fickas, K. C., Cohen, W. B., & Yang, Z. (2015). Landsatbased monitoring of annual wetland change in the
Willamette Valley of Oregon, USA from 1972 to
2012. Wetlands Ecology and Management, 1–21.
Finn, J. T., & Howard, R. (1981). Modeling a
beaver population on the Prescott Peninsula,
Massachusetts: Feasibility of LANDSAT as an
input. NASA. Goddard Space Flight Center
Eastern Reg. Remote Sensing Appl. Conf.; p 155162
Gómez, Cristina, White, J. C., & Wulder, M. A. (2016).
Optical remotely sensed time series data for land
cover classification: A review. ISPRS Journal of
Photogrammetry and Remote Sensing, 116(C),
55–72.
Kayastha, N., Thomas, V., Galbraith, J., & Banskota, A.
(2012). Monitoring Wetland Change Using InterAnnual Landsat Time- Series Data. Wetlands,
32(6), 1149–1162.
Kennedy, R. E., Yang, Z., & Cohen, W. B. (2010).
Detecting trends in forest disturbance and
recovery using yearly Landsat time series: 1.
ForestSAT 2016 Abstracts Summary
LandTrendr--Temporal segmentation algorithms.
Remote Sensing of Environment, 114(12), 2897–
2910.
Kennedy, R. E., et al. (2014). Bringing an ecological
view of change to Landsat-based remote sensing.
Frontiers in Ecology and Environment, 12, 6, 339346.
Naiman, R. J., Johnston, C. A., & Kelley, J. C. (1988).
Alteration of North American streams by beaver.
BioScience, 753–762.
Nielsen, E. M., Prince, S. D., & Koeln, G. T. (2008).
Wetland change mapping for the U.S. midAtlantic region using an outlier detection
technique. Remote Sensing of Environment,
112(11), 4061–4074.
Rosell, F., Bozser, O., Collen, P., & Parker, H. (2005).
Ecological impact of beavers Castor fiber and
Castor canadensis and their ability to modify
ecosystems. Mammal Review, 35(3 4), 248-276.
Schroeder, T. A., Wulder, M. A., Healey, S. P., &
Moisen, G. G. (2011). Mapping wildfire and
clearcut harvest disturbances in boreal forests
with Landsat time series data . Remote Sensing of
Environment, 115(6), 1421–1433.
Senf, C., Pflugmacher, D., Wulder, M. A., & Hostert, P.
(2015). Characterizing spectral–temporal patterns
of defoliator and bark beetle disturbances
using Landsat time series. Remote Sensing of
Environment, 170(C), 166–177.
Townsend, P. A., Walsh, S. J., & Butler, D. R. (1995).
Beaver pond identification through a satellitebased ecological habitat classification. In Proc.,
American Society for Photogrammetry and Remote
Sensing and the American Congress on Survey and
Mapping Annual Meeting, Charlotte, NC (pp. 102111).
Townsend, P. A., & Butler, D. R. (1996). Patterns of
landscape use by beaver on the lower Roanoke
River floodplain, North Carolina. Physical
Geography, 17(3), 253-269.
Wulder, M. A., Masek, J. G., Cohen, W. B., Loveland,
T. R., & Woodcock, C. E. (2012). Opening the
archive: How free data has enabled the science
and monitoring promise of Landsat. Remote
Sensing of Environment, 122(C), 2–10.
Zhu, Z., & Woodcock, C. E. (2014). Continuous
change detection and classification of land cover
using all available Landsat data. Remote Sensing
of Environment, 144(C), 152–171.
273
Development of Methods
Universidad Mayor
Multi-Scale Remote Sensing of Mangrove Structure
and Biomass/Blue Carbon
Emanuelle Feliciano1, 2, Temilola Fatoyinbo1, David Lagomasino1, 3, Seung Kuk Lee1, 3
Biospheric Sciences Laboratory, NASA Goddard Space Flight Center, 8800 Greenbelt Road, Greenbelt,
MD 20771, United States.
2
NASA Postdoctoral Program, Universities Space Research Association, 7178 Columbia Gateway Dr., Columbia,
MD 21046, USA.
3
Goddard Earth Sciences, Technology and Research (GESTAR), Universities Space Research Association,
7178 Columbia Gateway Dr., Columbia, MD 21046, USA.
1
Abstract
Coastal blue carbon ecosystems such as mangroves have the highest total carbon densities of all
ecosystems. The high carbon sequestration coupled with the high risk of destruction make mangroves a
prime candidate for carbon mitigation initiatives such as the United Nations Collaborative Programme
on Reducing Emissions from Deforestation and Degradation in Developing Countries (UN-REDD
and REDD+). Our wetland group at NASA is investigating the use of multi-scale remote sensing
techniques coupled with ground measurements at various mangrove sites around the world including
Africa, Bangladesh and the United States. As part of these projects we have been working with local
governments, scientific institutions and international organizations such as the WWF, the UN-REDD
programme, USAID, and SilvaCarbon among others. The study areas of our case studies include
the mangrove forests of Mozambique in Africa, the Sundarbans in Bangladesh and the Everglades
National Park in the United States. To study the structure and health of the mangrove forests we
are using remote sensing data from airborne LiDAR (ALS) in conjunction with TanDEM-X (TDX) and
WorldView (WV) satellites to estimate canopy height. Our research strategy consists in using ALS to
upscale field estimates of biomass to a larger scale and enable validation of TDX and/or stereo WV
derived estimates of canopy height and biomass. This enable us to estimate and model above ground
biomass (AGB) as it has a proportional relationship with canopy height.
For Mozambique (Zambezi Delta), we are presenting results that include (1) the validation and
comparison of independent mangrove canopy height measurements from ALS, TDX and WV, (2) AGB
modeling and total AGB estimation using ALS data and field measurements. For the Sundarbans in
Bangladesh we show preliminary results of canopy height, AGB and below-ground biomass for the
largest continuous mangrove forest in the world. For the Everglades National Park in the United States
we present results of canopy height and AGB estimation using a combination of ALS, TDX and mangrove
allometry. Furthermore, we present results of AGB estimation using Terrestrial Laser Scanning (TLS) in
the Everglades. TLS is a tool that could help us reduce AGB uncertainty in the near future. In addition
to the findings of these case studies, we present preliminary results of mangrove growth rates using
a combination of Landsat data and canopy height data for test sites in Africa and Bangladesh. The
research of our wetland group is of upmost importance if a carbon monitoring system (CMS), such as
the proposed NASA CMS needs to be developed in blue carbon ecosystems for monitoring, reporting
and verification (MRV) activities.
274
Universidad Mayor
ForestSAT 2016 Abstracts Summary
P-Band DInSAR time series of river bank erosions:
Preliminary results and comparisons with field
measurements
Karlus A. C. de Macedo1*, Thiago L. M. Barreto1, Leandro Matos1, Dieter Lübeck1, Carlos T. C. Gamba2* and,
Daniel S.N.A. Albarelli 2, Pedro R. Crisma2, Adalberto A. Azevedo2 and João Bosco Jr.3*
1
Bradar Remote Sensing , Brazil
Instituto de Pesquisas Tecnológicas do Estado de São Paulo (IPT), Brazil
3
Santo Antônio Energia, Brazil
*Corresponding authors: karlus.macedo@bradar.com.br, carlosgamba@ipt.br, joaobosco@santoantonioenergia.com.br
2
Keywords: P-band, synthetic aperture radar, SAR, differential interferometry, DInSAR, time series, terrain
movement, forest, river banks, erosion processes.
Abstract
P-band differential interferometric (DInSAR) data were acquired from August 2015 until July 2016,
along 240km downstream from Santo Antônio Energia dam in Madeira River (Amazon Basin). The
objective is to measure and monitor erosions along the river banks, especially in forested areas,
with DInSAR. We present the processing methodology and time series analysis applied to this data
set. Preliminary field and radar measurements are compared in order to evaluate and validate the
proposed methodology.
Introduction
Methodology
Differential synthetic aperture radar interferometry
(DInSAR) and further time series analysis are established techniques and they have been applied successfully to spaceborne remote sensing. Recently,
it has been demonstrated that airborne DInSAR at
L- or P-bands are able to deliver coherent dense-grid
of high-resolution measurements over vegetated terrain along time, overcoming the limitations of spaceborne DInSAR [de Macedo et al., 2012 ; Jones et
al, 2012]. Nowadays, further airborne DInSAR investigations and developments are being carried on at L
or P-band. This is because, the centimeter accuracy
and long-term coherence, which are achieved when
using these bands, are very appropriate to measure
processes such as erosion, sedimentation occurring
in forested areas [de Macedo and Wimmer, 2015]
and landslides [Delbridge, 2015]. This work investigates how terrain movements associated with erosions along vegetated river banks can be measured,
identified and monitored by P-band DInSAR.
A sequence of 11 airborne SAR high-resolution images at P-band were acquired from August 2015 until
July 2016, along 240km downstream from Santo Antônio Energia dam in Madeira River (Amazon Basin).
The P-band data were acquired and were processed
with the OrbiSAR airborne system, of Bradar, Brazil.
Details on the airborne DInSAR requirements and
processing can be found in de Macedo et al. [2012].
We conducted the field campaigns to measure the
terrain movements synchronized with the week of
the very first flight in order to guarantee the same
reference readings (t0) for both field and SAR measurements. Further readings occurred synchronized with the week of the second, fourth, seventh,
ninth and eleventh acquisition flights. The timeseries analysis applied to our interferometric data set
is based on the work published in de Macedo et al.
[2015]. The preliminary time series of the accumulated terrain movements here presented are obtained
via the direct-integration.
275
Universidad Mayor
Development of Methods
Results
Fig. 1 shows the DInSAR mosaic of the whole surveyed region in order to understand the magnitude,
location and area extesion of the terrain movements
seen by the radar. A ROI (region o interest) of the
surveyed area was selected. Fig. 2 shows the field
and DInSAR time series of the ROI, overlaid with the
X-band SAR amplitude and contour lines from the
X-band elevation model, obtained from the same
airborne sensor.
Both SAR and field measurements show that the terrain is moving downwards. Fig. 2(j,l) show photos,
where a process of piping can be clearly identified.
Fig. 3 shows a preliminary statistical analysis over
95 field stakes. The scatterer plot shows that the correlation factor between the DInSAR and field measurements is 0.28. Although weak, we can conclude
that there are correlation between the radar and
field measurements.
Reasons for the weak correlation are: (1) The field
measurements are pointwise, while the radar measures the average movements within a resolution
cell of 17mx17m. (2) In field, we measure the vertical
movements and the planimetric movements relative
to the river banks, which is mostly oriented to the
East-West plane. While the radar, measures only in
the LOS (line-of-sight) direction, and are affected by
both horizontal and vertical movements within the
North-South Plane. Therefore, only the vertical movements are sensed by both the radar and field measurements. (3) The assumed reference stable stake
for the field measurements, some times, are located
within the area of movement, as revealed by the
Fig. 1: DInSAR Mosaic (left) and corresponding
Google Earth® Imagery (right)
276
DInSAR images in Fig. 2. In such cases, the field measurements are possibly underestimating the movements. (4) Its is important also to note the external
sources of interference in the radar measurements,
like soil moisture. Its is not fully understood up to
now, the degree of contribution of soil moisture in
the DInSAR measurements [Zwieback et al. 2015].
Conclusions
The preliminary conclusion is that the P-band DInSAR is sensible and precise enough to detect and
track the terrain movements along river banks, such
the ones caused by erosion processes. This unique
data set comparison between DInSAR and field data
allowed us to gain knowledge about the accuracy
and sources of interference for movements sensed
at P-band in dense forested areas. Further improvements in the processing chain, such as the generation of the redundant network of interferograms,
are being implemented.
References
de Macedo K.A.C., Wimmer C., Barreto, T. L., Lubeck D., Moreira, J. R., Rabaco L.L.M., Oliveira
W.J. (2012). Long-term airborne DInSAR measurements at X-and P-Bands: A case study on
the application of surveying geohazard threats
to pipelines. IEEE Journal of Selected Topics in
Applied Earth Observations and Remote Sensing, 5(3), 990-1005.
de Macedo K.A.C., Wimmer C. (2015, July). Time series of airborne DInSAR data over the Amazon
flooded vegetation: Water level changes. In 2015
IEEE International Geoscience and Remote Sensing Symposium (IGARSS) (pp. 5252-5255). IEEE.
Delbridge B., Bürgmann R., Fielding E., Hensley S.,
(2015, July). Kinematics of the slumgullion landslide from UAVSAR derived interferograms. In
2015 IEEE International Geoscience and Remote
Sensing Symposium (IGARSS) (pp. 3842-3845).
IEEE.
Jones. C.E, Bawden G., Deverel S., Dudas J., Hensley
S.,. Yun S. H. (2012, November). Study of movement and seepage along levees using DINSAR
and the airborne UAVSAR instrument. In SPIE
Remote Sensing (pp. 85360E-85360E). International Society for Optics and Photonics.
Zwieback, S., Hensley, S., Hajnsek, I. (2015). Assessment of soil moisture effects on L-band radar interferometry. Remote Sensing of Environment,
164, 77-89.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Fig. 2: DInSAR and field time series and photo of the geodynamic process.
Fig. 3: Preliminary radar vs.
field measurement comparisons
277
Universidad Mayor
Development of Methods
Poster Presentation, ForestSat 2016, Santiago, Chile.
Statistical correction of Lidar-Derived digital elevation models
with multispectral airborne imagery
Kevin J. Buffington, Oregon State University and US Geological Survey
Bruce D. Dugger, Oregon State University
Karen M. Thorne, US Geological Survey
John Y. Takekawa, Audubon California
Vertically accurate digital elevation models (DEMs)
are necessary for studying low-slope ecosystems
and hydrology. The biogeomorphology of coastal
wetlands are largely controlled by tidal inundation;
as sea levels rise, accurate wetland DEMs are
critical for modeling into the future. Airborne
light detection and ranging (lidar) is a valuable
tool for collecting vast amounts of elevation data
across large areas; however, the limited ability to
penetrate dense vegetation with lidar increases
the vertical uncertainty. Methods to correct lidarderived elevation data exist, but a reliable method
that requires limited field work and maintains
spatial resolution is lacking. We present a novel
method, the Lidar Elevation Adjustment with NDVI
(LEAN), to statistically correct lidar digital elevation
models (DEMs) with vegetation indices (NDVI) from
multispectral airborne imagery (NAIP) and surveygrade GPS measurements. Across 17 tidal marshes
278
along the Pacific coast of the U.S., we achieved an
average root mean squared error of 0.072 m, with a
40-75% improvement in accuracy from the lidar bare
earth DEM. Results from our method compared
favorably with results from three other methods
(minimum-bin gridding, mean error correction, and
vegetation correction factors), and a power analysis
applied to our extensive GPS dataset (>17,000
ground points) showed that on average 118 points
were needed to calibrate a site-specific correction
model for tidal marshes along the Pacific coast. By
using readily available imagery and field surveys,
we found that lidar-derived DEMs can be adjusted
for greater vertical accuracy while maintaining high
(1 m) horizontal resolution. While developed in tidal
marshes, the LEAN model approach could be applied
to any ecosystem where vertical accuracy of DEMs is
important and survey-grade GPS measurements can
be collected.
Forestry & Forest Management
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Analysing the relations between landscape structural changes
and hydrological response at subcatchment scale in temperate
forest basins: the case of Maule’s inner dryland
Benjamín Sotomayor*1,2, Mauricio Galleguillos1,2, Christian Little3, Antonio Lara2,4
Laboratoy of Ecology of Ecosystems, School of Agronomical Sciences, University of Chile
2
Center for Climate and Resilience Research (CR)2, Santiago, Chile.
3
Instituto Forestal, Valdivia, Chile.
4
Instituto de Conservación, Biodiversidad y Territorio, Universidad Austral de Chile, Valdivia, Chile
*Corresponding author: benjasotomayor@ug.uchile.cl -- Santa Rosa Avenue #11315, La Pintana, Santiago, Chile.
1
Keywords: Landscape composition, configuration, change, hydrologic response, land cover, correlation.
Abstract
We generated 8 land cover maps for the 2000-2014 period in Cauquenes and Purapel catchments,
located in Maule region’s inner dryland, Central Chile, to test for relations between landscape structural
and configurational changes and runoff coefficients at annual and seasonal scale at subcatchment
level. Overall accuracies ranged from 85% to 100%. A Spearman’s Correlation Matrix revealed several
positive and negative statistically significant correlations between landscape configuration and
structure metrics, suggesting that the landscape’s configurational and structural dynamic driven by
anthropic forestry management could have an impact in hydrological response at short term.
Introduction
South-central Chile inner dryland and coastal range
is a landscape composed of a matrix of extensive
industrial pine and eucalyptus plantations, degraded
shrubland and growingly fragmented native forests
patches, where industrial forestry has expanded
rapidly and intensively since mid-1970’s (Echeverria
et al., 2006). Today, this activity is the main driver
that triggers the landscape’s inner spatial dynamic,
driven by the short rotation period applied to
industrial plantations (10 to 20 years) (Armesto et
al., 2007). Several studies conducted in Chile and
elsewere have shown the effects that fast-growing
species plantations have on hydrological response
by analysing landscape composition, defined as the
total surface per land cover class, changes in the
long term (Leblanc et al. 2008, Little et al. 2009).
However, until now it’s unclear whether landscape
configuration dynamics, understood as the landscape
patches arrangement and position changes over
time, in short term could have any significant effect
over hydrological response at subcatchment scale.
The main objective of this investigation is to test
for relations between landscape structural changes
and the hydrological response at two forested
catchments in coastal Maule region, Chile, between
2000 and 2014.
Methods
The study area is located in the inner dryland of
the Maule Region and comprises the Purapel (270
km2) and Cauquenes (619 km2) subcatchments,
both characterized by a pluvial hydrologic regime
and Mediterranean climate. Natural vegetation
consists of a mixture of deciduous Nothofagus forest
281
Universidad Mayor
Forestry & Forest Management
and thorny schlerophyllous shrubs (Luebert and
Pliscoff, 2006), but in present days, the landscapes
is dominated by a matrix of fast-growing Pinus
radiata and Eucalyptus spp. plantations and isolated
fragment of native forests in remote places and
heavily degraded anthropogenic savannas in the
lower reaches (Echeverria et al. 2006, Van de Wouw
et al. 2011). Landscape composition for 2000-01,
2001-02, 2004-05, 2005-06, 2006-07, 2008-09, 201011 and 2013-14 summer seasons was represented
using Landsat-derived land cover maps using the
Maximum Likelihood Classifier (MLC), including
a 1985-86 map as a reference situation. Six land
cover classes were established based on previous
field data: Agricultural-Meadows (A), Bare Soil
(BS), Industrial Plantations (P), Native Forest (NF),
Shrubland (SH) and Impervious land (I). 20 training
zones per class were defined per class through
Google Earth images interpretation and previous
field reference data. We used Landsat VNIR+SWIR
bands from the Landsat Climate Data Record
(CDR) plus 3 vegetation indices (Summer NDVI,
GNDVI and ΔNDVIwinter-summer) to enhance spectral
separability among classes, and topographic
correction was applied with the C-Correction
method as proposed by Hantson and Chuvieco
(2011). Accuracy assessment was conducted with a
Confusion Matrix using 10 validation zones per land
cover class for each date and study site. Landscape
configuration was represented through the land
cover’s general position in the landscape using
a Topographic Position Index-based landscape
landform classification (Weiss, 2001) derived from
an SRTM 1-arcecond DEM. Hydrological response
was represented with the Runoff Coefficient (RC),
calculated as the ratio between accumulated rainfall
and runoff at annual and seasonal (four-month) scale.
Finally, the existence of relationships between runoff
coefficients and landscape configuration features
was addressed using Spearman’s Correlation Matrix
with a significance level of 0.05.
Results and Discussion
Overall accuracies of the land cover maps ranged
from 85% to 100%, indicating a strong agreement
with the reference data. An intense land use/land
cover change process was observed in the 1985-2000
period in both catchments. In 1985-86, the landscape
composition was dominated by a matrix of degraded
Shrubland and Bare Soil in both catchments, but
by 2000-01, it had changed significantly due to the
explosive expansion of Industrial Plantations, which
increased by 461% in Cauquenes and 680% in Purapel
(figure 1). On the other hand, Native Forest shrunk by
43% in Cauquenes and 53% in Purapel at an annual
Fig.1: Land cover maps between 1985-2000-2015 in Cauquenes and Purapel catchments
282
Universidad Mayor
rate of -3.63% and -6.29%, respectively. In the 20002014 period, the intensity of this process decreased,
being the dominant matrix a mixture of Industrial
Plantations and Shrubland (figure 1). In fact, no
significant trend in total surface was observed in any
land cover in both catchment in this period (figure
2), indicating that the landscape has stabilized in a
functioning state defined as Dynamic Equifinallity
by Patterson y Hoalst-Pullen (2011). An exception
was found in the case of Industrial Plantations in
Purapel, where a significant decrease in surface was
observed in this period due to extensive harvest
operations. Nevertheless, this specific land use has a
strong spatial persistence (Miranda et al. 2015) so it
is expected to increase in surface as trees grow older
and can be precisely classified by remote sensing
methods.
ForestSAT 2016 Abstracts Summary
Hillsides and the annual and autumn runoff coefficient
in Cauquenes (figure 3). CONAF (2011) indicated
that the majority of newly stablished industrial
plantations in the study area between 1999 and 2008
were set in shrublands and abandoned cropland,
and in that sense, Farley et al. (2005) and Huber et
al. (2008), analysing the effects of replacing shrubs
and prairies with Pinus radiata over hydrological
cycle, reported a decrease runoff generation mainly
due to increase in interception, percolation and
evapotranspiration at local scale. The correlation
obeys to stochastic behaviour of plantations due to
the lack of integrated management at catchment
level, as can be seen by analysing the temporal
distribution of data in the correlation (figure 3).
This wasn’t observed in Purapel because of the
significant reduction in this land cover’s surface due
to extensive harvests operations (figure 2).
A negative and significant correlation was observed
between the surface of Industrial Plantations at
Fig. 2: Composition trend per land
cover class in 2000-2014 period in
Cauquenes and Purapel catchments.
Fig. 3: Statiscally significant correlations
between landscape structural features and
runoff coefficients in Cauquenes and Purapel
between 2000-2014.
283
Forestry & Forest Management
Conclusion
Statistically significant correlations were detected
at annual and seasonal scale in Cauquenes
subcatchment and suggest that the landscape’s inner
configurational dynamic triggered by the extensive
and intensive forestry use could affect hydrological
response in the short term in the context of highly
anthropized landscapes. The investigation only
shows preliminary exploratory results, hence more
effort should be done to improve our knowledge
about the effect of landscape configuration in
hydrology in order to provide a sustainable use
of the territories by integrating this effect in the
management at catchment level.
References
Armesto, J; Arroyo, M; Hinojosa, L. 2007. The
Mediterranean environment of central Chile. The
physical geography of South America 2007: 184199.
Aronson, J; del Pozo, A; Ovalle, C; Avendaño, J;
Lavin, A; Etienne, M. 1998. Land use changes
and conflicts in central Chile. Landscape
disturbance and biodiversity in Mediterraneantype ecosystems 1998: 155-168.
Echeverria, C; Coomes, D; Salas, J; Rey-Benayas, JM;
Lara, A; Newton, A. 2006. Rapid deforestation
and fragmentation of Chilean Temperate Forests.
Biological Conservation 130(4): 481-494.
Farley, KA; Jobbágy, EG; Jackson, RB. 2005. Effects
of afforestation on water yield: A global synthesis
with implications for policy. Global Change
Biology 11(10): 1565-1576.
284
Universidad Mayor
Hantson, S; Chuvieco, E. 2011. Evaluation of different
topographic correction methods for landsat
imagery. International Journal of Applied Earth
Observation and Geoinformation 13(5): 691-700.
Huber, A; Iroumé, A; Bathurst, J. 2008. Effect of
Pinus radiata plantations on water balance in
Chile. Hydrological Processes 22(May 2007): 142148.
Leblanc, MJ; Favreau, G; Massuel, S; Tweed, SO;
Loireau, M; Cappelaere, B. 2008. Land clearance
and hydrological change in the Sahel: SW Niger.
Global and Planetary Change 61(3-4): 135-150.
Little, C; Lara, A; McPhee, J; Urrutia, R. 2009.
Revealing the impact of forest exotic plantations
on water yield in large scale watersheds in SouthCentral Chile. Journal of Hydrology 374(1-2): 162170.
Miranda, A; Altamirano, A; Cayuela, L; Pincheira, F;
Lara, A. 2015. Different times, same story: Native
forest loss and landscape homogenization in
three physiographical areas of south-central of
Chile. Applied Geography 60: 20-28.
Patterson, M; Hoalst-Pullen, N. 2011. Dynamic
equifinality : The case of south-central Chile’s
evolving forest landscape. Applied Geography
31(2): 641-649.
Weiss, a. 2001. Topographic position and landforms
analysis. Poster presentation, ESRI User
Conference, San Diego, CA 64: 227-245.
Van de Wouw, P; Echeverría, C; Rey-Benayas, JM;
Holmgren, M. 2011. Persistent acacia savannas
replace Mediterranean sclerophyllous forests in
South America. Forest Ecology and Management
262(6): 1100-1108.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Large area tree species mapping in mixed temperate
forests from multi-temporal RapidEye satellite images
and LiDAR data.
1
Paul Magdon1, Collins Kukunda1, Hans Fuchs1 and Christoph Kleinn1
Chair of Forest Inventory and Remote Sensing, Georg-August University of Göttingen
Keywords: Forest management, multi-temporal, RapidEye, RandomForest
The importance of forests particularly in densely
populated areas of Central Europe has clearly
shifted from a pure source of timber, wood and
other tangible products towards a source of multiple
ecosystem services including water protection,
carbon sequestration, biodiversity preservation,
recreational functions, and also production of
wood. Close-to-nature forest management and
conservation practices have been developed
and adopted to manage forest lands sustainably
so that the long-term provision of this suite of
environmental, ecological and economic services
is secured, simultaneously on the same piece of
land. Current forest management emphasizes
specific silvicultural practices that include natural
regeneration, selective logging, dead wood
management, habitat tree conservation and
economic target tree identification. Consequently,
forest structure in many regions has changed away
from the traditional age-class forests towards
highly complex – closer to nature – forest stands
with trees of different species, ages/development
stages and dimensions. In order to implement these
management and conservation practices, as well as
to effectively monitor their impact on the ecosystem
services, high quality inventory data is essential.
Stand inventory approaches were so far based on
compartments as the basic spatial management
units, for which a homogenous forest structure was
assumed. However, with increasing complexity and
with more diverse silvicultural practices that focus
on individual trees instead of entire compartments,
new challenges arise for forest monitoring.
Remote sensing techniques can operationally
support the development of forest management
and conservation plans when the demanded
target variables can be observed or modelled with
sufficient accuracy. Tree species composition and
spatial distribution is among the fundamental
input information for forest management and
conservation planning and decisions in these closeto-nature forest types.
In this study we research into the potential of
combining multi-temporal RapidEye satellite images
and LiDAR canopy height models (CHM) to create
maps of the main European tree species. We analyze
the seasonal changes in the spectral signatures of the
species in order to i) identify seasonal characteristics
that can be utilized to differentiate between the
species, ii) to identify the seasons /dates which
are most important for the discrimination and iii)
to evaluate the minimum number of acquisition
required per vegetation period.
This work is part of the DFG funded “Biodiversity
Exploratories” Priority Program, a large collaborative
research project that covers different forest types
across three regions (each ~1000km²) in Germany.
We base our analysis on 78 RapidEye L1B satellite
images collected between 2014 and 2015, which were
orthorectified with a high quality local elevation model
and a systematic set of field survey GCPs resulting in a
subpixel accuracy of the image co-registration. Cloud
detection and atmospheric corrections were applied
using ATCOR software and in-situ spectral reflectance
measurements. The training dataset was compiled
from a full census of trees on 32 plots with a size of
one hectare each, where a total number of n=12,144
trees of 28 tree species were recorded in 2015. The
tree species information from the field observations
was merged with tree crown delineations derived
from the LiDAR CHM using seed growing and
watershed segmentation methods. From each of the
segments RapidEye pixels were extracted and fed
into a spectral signature database. Using graphical
285
Forestry & Forest Management
and statistical analysis methods we identify the
species specific characteristics and compare the
intra- and inter-species variation. Furthermore, we
research into the impact of the crown architecture
on spectral reflection by extracting 3D crown shapes
from the LiDAR CHM. Doing so we wish to contribute
to the understanding of the sources of variability of
canopy reflectance measurements by separating
the signal variation into species specific variation
and variation caused by the crown structure. Finally,
non-parametric RandomForest models are trained to
produce tree species distribution maps for all three
regions. The validation of the maps is done with an
independent sample based forest inventory which
covers all three regions based with a randomly placed
systematic grid.
Given the large and systematic set of images,
training and validation data and the extent of the
study area, we assume that the results from this
study will contribute to the development of remote
sensing assisted forest monitoring systems of
highly complex mixed temperate forests, which
are urgently demanded by forest managers,
conservation planners and researchers.
286
Universidad Mayor
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Remote Sensing Contributions to Indicators of Biological
Diversity in the U.S. National Report on Sustainable
Forests—2015
Mark D. Nelson
U.S. Department of Agriculture, Forest Service, Northern Research Station, 1992 Folwell Avenue, St. Paul, MN 55108,
phone (651) 649-5104, e-mail mdnelson@fs.fed.us,
Curtis H. Flather
U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station, 240 West Prospect Road, Fort Collins,
CO, 80526, phone (970) 498-2569, e-mail cflather@fs.fed.us,
Kurt H. Riitters
U.S. Department of Agriculture, Forest Service, Southern Research Station, 3041 East Cornwallis Road, Research Triangle
Park, NC, 27709, phone (919) 549-4015, e-mail kriitters@fs.fed.us,
Carolyn Sieg
U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station, 2500 South Pine Knoll Drive, Flagstaff,
AZ, 86001, phone (928) 556-2151, e-mail csieg@fs.fed.us, and
James D. Garner
U.S. Department of Agriculture, Forest Service, Northern Research Station, 1992 Folwell Avenue, St. Paul, MN 55108,
phone (651) 649-5107, e-mail jamesdgarner@fs.fed.us,
Guy Robertson
U.S. Department of Agriculture, Forest Service, National Program for Sustainability Assessment, 1400 Independence Ave.,
SW, Washington, DC 20250, phone (703) 605-1071 e-mail grobertson02@fs.fed.us,
Keywords: biological diversity, land use, land cover, remote sensing
Abstract:
Forest biological diversity contributes to human welfare through multiple ecosystem services. The
Montréal Process (MP) provides a standard international framework for assessing a set of criteria and
indicators (C&I) of the sustainability of temperate and boreal forest ecosystems and their ecological,
social, and economic components in twelve countries, including the United States. The National Report
on Sustainable Forests—2015 relies on the MP C&I to organize and present data relevant to U.S. forests.
The first of seven criteria addressed in the 2015 Report is Conservation of Biological Diversity, which is
organized into nine indicators that address three sub-criteria: ecosystem diversity, species diversity,
and genetic diversity. The aim of this presentation is to report results for U.S. 2015 indicators of forest
conservation of biological diversity, with emphasis on remote sensing and geospatial contributions
and challenges.
Ecosystem diversity indices of U.S. forest land use, composition, and structure are based on national
forest inventory (NFI) field observations, post-stratified by satellite image-based land cover datasets
to increase estimate precision. Protected vegetation land cover classes are assessed by combining
30-m Landfire existing vegetation types with a protected areas geospatial database. Landscape
metrics of forest fragmentation are obtained from 30-m National Land Cover Databases of 2001 and
2011. Total area of forest land use increased by 14 million acres since the previous report. Timberland
area increased by 14 million acres in large diameter size classes and decreased by 7 million acres in
287
Forestry & Forest Management
Universidad Mayor
medium and small diameter size classes. Woodlands (41%) and forest (31%) have a higher percentage
of area protected than other natural vegetation cover types (16%). Between 2001 and 2011, net loss of
interior forest cover varied from 7 to 20 percent depending on the landscape scale of measurement.
Indices of species and genetic diversity are based on national biological field inventories. Forestassociated taxa showed a general increase in the proportion of possibly extinct or at-risk species since
2003. Greatest declines in number of forest bird species occurred in oak ecoregions of the southern
Appalachians, pine and northern hardwood ecoregions of the upper Midwest and Great Lakes, and
montane and arid high plains ecoregions in the intermountain West.
Declining populations and shrinking geographic ranges provide indirect indications of genetic diversity
loss. Between 1966 and 2011, about 19% of forest-associated bird species increased in populations
and 20% decreased; decliners include species associated with early successional or wetland habitats.
Ten percent of forest-associated species no longer fully occupy their former range; substantially higher
rates of range shrinkage occur for infraspecies than for species.
Harmonization across indices is challenging due to different definitions in remote sensing and field
inventory datasets, e.g., land cover versus land use. Understanding the interaction among indicators
and the potential causes of indicator dynamics will necessarily involve linking indicators spatially
and exploring patterns of covariation. Indicators derived from traditional on-the-ground surveys
(e.g., species richness, species imperilment) could be modeled using indicators of habitat that are
traditionally derived from remotely sensed imagery (e.g., area, configuration, and composition of land
uses and land cover).
288
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Potential of using data assimilation to support forest planning
Rami Saad, Peichen Gong, Tomas Lämås and Göran Ståhl
Keywords: uncertainty; suboptimal loss; remote sensing; combining data; Bayesian statistics.
Abstract
Uncertainty in forest information typically results in economic losses in addition to other losses as a
consequence of suboptimal management decisions. Several techniques have been proposed to handle
such uncertainties; however, these techniques are often complex and costly. Data assimilation (DA) has
recently been advocated as a tool which may reduce the uncertainty and thereby reduce complexity
in implementing techniques to handle uncertainty in forest planning. It offers an opportunity to
make use of all new sources of information in a systematic way, and thus provide accurate and up-todate information to forest planning. New remote sensing techniques can deliver information about
forest stands with short intervals at low cost; DA offers a unified framework to make use of these
new sources of forest information. In this study we review the literature on handling uncertainties in
forest planning as well as related literature from other scientific fields in order to assess the potential
benefits of using DA in forest planning. We identify five major potential benefits: (i) The accuracy of
the information will be improved, (ii) The information will be kept up-to-date, (iii) The DA procedure
will provide information with known accuracy, (iv) Bayesian decision making can be applied, whereby
the accuracy of the information can be utilized in the decision making process and (v) DA data and
Bayesian decision making allow for the analysis of optimal data acquisition decisions.
289
Universidad Mayor
Forestry & Forest Management
Use of remotely sensed data to spatially predict optimal final
stand density, value and the economic feasibility of pruning
for even age plantation forests
Michael S Watt1, Mark O Kimberley2, Jonathan P Dash2, Duncan Harrison2
Scion, PO Box 29237, Fendalton, Christchurch, New Zealand
2
Scion, PO Box 3020, Rotorua, New Zealand
1
Keywords: 300 Index; LiDAR; productivity surfaces; pruning; radiata pine; Site Index; stocking
Two of the key end-uses for which forest plantations
are grown are clearwood timber, in which the lower
branches of the trees are removed (pruned) and
structural grade timber in which trees are not pruned.
Forest managers in many plantation growing regions
have to determine the relative profitability of growing
for clearwood or structural grade to ascertain
whether pruning will be cost effective. Regardless of
the regime chosen managers also need to determine
the final crop stand density (Sopt) that will maximise
value. Using a comprehensive series of simulations
extracted from a New Zealand growth model for
Pinus radiata a model of optimal Sopt that maximises
the volume and value of clearwood and structural
grade regimes was developed. At its core this model
uses either GIS surfaces, satellite imagery or LiDAR
data to predict two site productivity indices (Site
Index, 300 Index) which in turn are used to spatially
predict Sopt. Using log prices over the last 21 years
the model was used to spatially predict variation
in Sopt for New Zealand’s largest plantation forest
(Kaingaroa Forest) and throughout New Zealand.
The model was also used to predict the relative value
of clearwood vs. structural regimes, through time.
Of the three remotely sensed data sources LiDAR
data most accurately predicted both Site Index and
300 Index and the final models had respective R2
values of 0.88 and 0.79. The wide predicted variation
in Sopt for Kaingaroa forest and at the national level
was clearly a function of the two productivity indices
300 Index and Site Index. For all three regimes the
highest values of Sopt (ca. 500 – 700 stems ha-1) were
290
found on sites with high 300 Index (high volume) and
low to moderate Site Index ( moderate height). The
lowest stand densities (ca. 200 – 300 stems ha-1)
were found on sites with low to moderate 300 Index
and moderate to high Site Index. Mean predicted
Sopt for the existing plantation resource within New
Zealand was 603 stems ha-1 for the structural regime
and 569 and 449 stems ha-1, for the two clearwood
regimes, that, respectively, targeted production of
small and large diameter pruned log.
Temporal analyses show pruning profitability has
declined markedly over the last two decades in New
Zealand primarily in response to reductions in the
pruned log premium. Spatial analyses show pruning
to be profitable within at least 95% of plantations
using values from 1995-2000 and unprofitable within
at least 95% of plantations using values from 20112015. The most profitable areas for pruning were
found to be located in regions where 300 Index
ranges from moderate to high and Site Index is
relatively low.
This research clearly shows an application of
precision forestry that can use LiDAR or satellite
imagery to guide management operations at
both broad spatial scale and fine resolution. The
application of the presented model suggests there
is considerable scope for increasing plantation
value through optimising Sopt by site to values that
are higher than those generically prescribed and
highlighting sites where pruning is likely to be most
profitable.
Latin American Forests
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Comparing Generalized Linear Models and random forest to
model vascular plant species richness using LiDAR data in a
natural forest in central Chile
Javier Lopatin1,2, Klara Dolos1, H.J. Hernández2, M. Galleguillos3, Fabian Ewald Fassnacht2
Institute of Geography and Geoecology, Karlsruhe Institute of Technology (KIT), Kaiserstraße 12,
76131 Karlsruhe, Germany
2
Laboratory of Geomatics and Landscape Ecology, Faculty of Forest and Nature Conservation,
University of Chile, 11315 Santiago, Chile.
3
Department of Environmental Sciences, School of Agronomic Sciences, University of Chile, 11315 Santiago, Chile.
1
Keywords: Species richness, LiDAR, GLM, random forest, alpha-diversity, bootstrap validation.
Biodiversity is an essential element of the Earth
system, related to many important ecosystem
services. Cost-efficient and precise monitoring
systems are needed to support the conservation of
biodiversity in a quickly changing world. Here, we
tested the suitability of airborne discrete-return
LiDAR data for estimating vascular plant species
richness of second growth native forest ecosystems
in central Chile. We modelled the vascular plant
richness of four layers (total, tree, shrub and herb
richness) using twelve LiDAR-derived variables.
As species richness values are typically nonnormally distributed count data, the corresponding
asymmetry and heteroscedasticity in the error
distribution has to be considered. We therefore
compared the suitability of random forest (RF) and
a Generalized Linear Model (GLM) with a negative
binomial error distribution. In both cases, a feature
selection to identify the most relevant LiDAR
predictors was applied as a first step. In both model
types, the three most important predictors for all
four layers were altitude above sea level, standard
deviation of slope and mean canopy height.
This agreed with our preconception of LiDAR’s
suitability for estimating species richness which we
hypothesized to be its capacity to capture three
types of information: micro-topographical, macrotopographical and canopy structural. Generalized
Linear Models showed higher performances (r2:
0.66, 0.50, 0.52, 0.50; nRMSE: 16.29%, 19.08%,
17.89%, 21.31% for total, tree, shrub and herb
richness respectively) than RF (r2: 0.55, 0.33, 0.45,
0.46; nRMSE: 18.30%, 21.90%, 18.95%, 21.00% for
total, tree, shrub and herb richness, respectively).
In addition, the best GLM models were more
parsimonious (three predictors) and less biased than
the best RF models (twelve predictors). We explain
this with the mentioned non-symmetric error
distribution of the species richness values, which RF
could not capture properly.
From an ecological perspective, the predicted
diversity maps agreed well with the known
vegetation composition of the area. High species
numbers were found at low elevations and along
riversides. In these areas, overlapping distributions of
thermophile sclerophyllos species, water demanding
Valdivian evergreen species and species growing in
Nothofagus obliqua forests occur. Our three main
conclusions were: 1) appropriate model selection
is important when working with biodiversity count
data; 2) RF has troubles when applied to data with
non-symmetric error distributions; and 3) structural
and topographic information derived from LiDAR
data is useful for predicting local plant species
richness.
293
Universidad Mayor
Latin American Forests
Detection of low density natural forest in the Andes region
using LANDSAT 8 imagery
Author: Vega Isuhuaylas, Luis Alberto.
Affiliation: Forestry and Forest Products Research Institute (FFPRI)
Co-author: Hirata, Yasumasa.
Affiliation: Forestry and Forest Products Research Institute (FFPRI)
Keywords: Andean forest, remote sensing, forest detection, spectral test, Landsat 8
Abstract
Andean ecosystems in South-America are among the most diverse and threatened ecosystems in the
world. Due to a long history of exploitation that still continues, the remaining natural forests in the
Andes region are relict tree populations in inaccessible locations and open forests with low density of
individuals. From the perspective of forest conservation, the detection via remote sensing techniques
of such low density forests in large areas with freely available medium resolution satellite data
constitutes a challenge.
This study aims to evaluate the detection of Andean natural forests by remote sensing techniques
according to forest density using LANDSAT 8 satellite data, and to determine the most effective
detection variables.
The selected study area is the department of Cuzco (Peru). Field data on Andean forest condition was
gathered by plotless sampling survey in several locations and the forest stand density was calculated.
Low density forest was defined at different values for testing purposes. Next, spectral separability
tests between low density forest and other land cover classes were carried out using the mean values
of reflectance, vegetation indexes and other factors estimated via PCA and Tasseled cup analysis.
Finally, a land cover classification is carried out using the variables with better separability results and
the accuracy of classification is calculated.
294
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Effect of lidar pulse density on the prediction of aboveground
biomass change in Brazilian Amazon Rainforest
Carlos Alberto Silva1,2,3, Andrew Thomas Hudak2, Lee Vierling1, Carine Klauberg2, António Ferraz3, Mariano Garcia Alonso3,
Michael Keller4, Sassan Saatchi3
1
Department of Natural Resources and Society, College of Natural Resources, University of Idaho, (UI), 875 Perimeter
Drive, Moscow, Idaho, 83844, USA
2
USDA Forest Service, Rocky Mountain Research Station, RMRS, 1221 South Main Street, Moscow, ID, 83843, USA
3
Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA 91109, USA
4
USDA Forest Service, International Institute of Tropical Forestry, San Juan, PR, 00926, USA
Keywords: Airborne Lidar, Biomass, Tropical Forest, Forest Inventory, Pulse densit
Tropical forests are an important component of
global carbon stocks, but the response of tropical
forest biomass to climate is not sufficiently studied
or understood. Airborne lidar (Light Detection and
Ranging) is well suited for quantifying tropical forest
carbon stocks, however trade-offs exist between
lidar pulse density and acquisition cost. Our objective
was thus to evaluate the effect of lidar pulse density
on aboveground biomass (AGB) change prediction
using airborne lidar and field plot data in a tropical
rain forest located near Paragominas, Pará, Brazil.
Forest attributes such as tree density and diameter
at breast height were measured in 84 square field
plots of 50x50m in 2014. Using previously published
allometric equations, tree AGB was computed
and then summed to calculate total AGB at each
sample plot. The lidar data were acquired 2012
and 2014, and for each dataset the pulse density
was downscaled from its original density to lower
densities of 12, 10, 8, 6, 4, 2, 1, 0.8, 0.6, 0.4 and 0.2
Figure 1. Lidar point cloud by target pulse density. The 3d points are coloured by height.
295
Latin American Forests
pulses/m2. The effect of pulse density on the AGBC
prediction was evaluated under two scenarios i) lidar
data both from 2012 and 2014 were normalized to
height-aboveground using the DTMs created from
each density dataset; and ii) To simulate the effect
on subsequent acquisition, the DTM generated
at highest pulse density (12 pulse/m2) from 2012
was used to normalize height-aboveground for
downscaled lidar datasets both from 2012 and 2014.
For both scenario, we clipped the point cloud into
the sample plots to compute mean canopy height
(MCH). A power model was used to model AGB from
MCH in 2014, and to predict AGB for 2012 and 2014.
Leave-one out cross validation was applied and the
models were assessed according to the coefficient of
determination (R²), relative root mean square error
(RMSE,%) and Bias (%). The AGB change was then
computed as the difference between AGB predicted
from 2012 and 2014. To achieve consistent results;
we repeated all procedures 30 times. Our findings
showed that for both scenarios the AGB change
296
Universidad Mayor
prediction was only slightly affected by pulse density.
For models in scenario 1, the mean RMSE and Bias
slightly increased from 15.01 to 15.99%, 0.09 to
0.16% and the R² slightly decreased 0.75 to 0.72 as
pulse density decreased from 12 to 0.2 pulses/m².
For scenario 2 the mean RMSE (%) increase from
14.91 to 15.20%, the Bias (%) stayed constant at
0.09, and the R2 decreased from 0.76 to 0.75 as pulse
density decreased as well. Moreover, as pulse density
decreased from 12 to 0.2 pulses/m², the mean AGB
change prediction across sample plots slightly
decreased from 4.22 to 3.84 Mg/ha and from 6.12 to
5.82 Mg/ha, for scenario 1 and 2, respectively. This
study showed that a DTM generated from either low
or high pulse density lidar can be used to normalize
height aboveground for predicting AGB changes in
tropical forest. We therefore, we conclude that there
is good potential to monitor carbon pools in Brazilian
Tropical Rain Forest using airborne lidar data with
low pulse density.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Estimating of the leaf area index in a forest fragment of
mixed ombrophilous forest in Brazil, using remote sensing
techniques
Bruna Nascimento de Vasconcellos Schiavo¹
Ângela Maria Klein Hentz¹
Ana Paula Dalla Corte²
Carlos Roberto Sanquetta²
¹ PhD student, Federal University of Paraná, Department of Forest Engineering – Curitiba, Parana state, Brazil
² PhD, Federal University of Paraná, Department of Forest Engineering – Curitiba, Parana state, Brazil
The leaf area index (LAI) is considered as a
component of the most sensitive part in forest
structure, because the canopy leaves regulate
the fundamental processes in forest productivity.
This study was conducted to estimate the leaf
area index, in a fragment of Mixed Ombrophilous
Forest, located in São João do Triunfo, Paraná
state, Brazil, using remote sensing techniques. The
LAI was generated from orbital images of Pleiades
sensor, based on the relationship between some
vegetation indices tested in this work. First, the
orbital images were corrected for atmospheric and
radiometric distortion, then two vegetation indices
(VI) NDVI and SAVI were generated. The vegetation
indices were used to test the models proposed in
the literature by the Manual for the Energy balance
algorithms of the earth’s surface (SEBAL, 2002) that
uses the IV SAVI and Duchemin (2006) using the IV
NDVI. In addition, 312 samples were distributed in
the study area, where the LAI was estimated from
the terrestrial sensor IC-110 Plant Canopy Analyzer,
considered the real value as control. The data
obtained by the remote sensing techniques were
correlated with the data obtained from the field,
using the Pearson’s correlation coefficient. The VI
NDVI present a correlation with the control of r =
0.32, and the SAVI presented a correlation of r = 0.78.
It was observed that the model proposed by SEBAL
(SAVI) was highly correlated with the control data (r
= 0.66). However, the model proposed by Duchemin
(NDVI) did not present correlation (r = 0.23), so is
considered inefficient to estimate LAI in the Mixed
Ombrophilous Forest. In addition, both models
presented an underestimation compared to the real
value. The LAI estimated by terrestrial and orbital
sensors presented an ample variability among the
data, a factor possibly related to species diversity,
a characteristic of native forests. Among of these
variations, it was possible to identify the different
canopies in these models, so the pixels with the
highest values of the LAI correspond to the area with
a predominance of hardwoods. The remote sensing
techniques and the images of the Pleiades sensor
performed well to estimate the leaf area index in the
forest fragment of Mixed Ombrophilous Forest.
Keywords: Araucaria Forest; Pleiades; CI-110 Plant
Canopy Analyzer;
Introduction
Woodgate et al. (2015) considers that leaf area index
(LAI) is a primary descriptor of vegetation structure and
is an essential climatic variable. It is considered as part
of the most sensitive structure of the forest, because
canopy leaves regulate some ecophysiological
processes, such as photosynthesis and transpiration,
which are considered as fundamental elements of
forest productivity (LARCHER, 2004; BAMBI, 2007).
There are many feasible methods and equipment for
the non-destructive estimation of leaf area index and
photosynthetically active radiation, so that isolated
tree, planting, direct biometric measurements of
leaves or indirect measurements can be considered
using regression models based on branches, trunks
Or productivity in general (Guimarães et al., 2013).
The two main methods of estimation of leaf area
index through remote sensing are based on the
empirical relationship between vegetation indexes
and LAI, or between models of radiation transfer
using passive sensors or neural networks (MA, 2014).
297
Universidad Mayor
Latin American Forests
The objective of this study was to test sensing
techniques to estimate leaf area index in a fragment
of the Mixed Ombrophilous Forest in the state of
Paraná, Brazil.
Methodology
The study was carried out at the Rudi Arno Seitz
Experimental Station of the Federal University of
Paraná, located in the city of São João do Triunfo,
in the Center-South Region of the state of Paraná,
Brazil. 312 sample points were distributed in the
study area, where the LAITERRESTRIAL was
estimated from the CI-110 Plant Canopy Analyzer
sensor, considered as field truth (witnesses). The
CI-110 Plant Canopy Analyzer is a passive ground
sensor, used to measure the amount of incident
solar radiation in the visible spectrum. According
to Meyer Junior (2014) the CI-110 is composed of a
rod with 24 sensors, and at the tip of the rod, is the
fish eye lens that allows to calculate the LAI, the
coefficient of transmission for penetration Diffuse,
the canopy opening index, the average leaf angle
and the canopy extinction coefficient.
LAIORBITAL was estimated through vegetation
indices, which are mathematical formulas developed
to evaluate the vegetation cover qualitatively
and quantitatively, using spectral measurements.
Atmospheric and radiometric correction of the
image of the Pléiades sensor was performed, and
the NDVI and SAVI vegetation indices were tested,
which subsidized the LAIORBITAL estimation tests
by the models of SEBAL (2002) and Duchemin
(2006).
The Pearson correlation analysis was performed
between the leaf area index obtained from the CI110 Plant Canopy Analyzer (LAITERRESTRIAL),
considered as field truth and obtained from the
orbital image (LAIORBITAL).
Results
The data obtained through the techniques of Remote
Sensing were correlated with the data obtained in
the field through the Pearson correlation coefficient.
The value of the leaf area index for the Mixed
Ombrophilous Forest fragment obtained in this
study ranged from 6.01 to 8.01. The value of leaf
area index for forests, as presented in the literature,
varies from 0.40 for a low density of individuals and
298
16.9 for old stands. The highest values reported are
for conifers, with maximum values between 6 and 8,
for deciduous forests (JONCKHEERE et al., 2004).
Regarding the correlation of the vegetation indices
with the field truth, the NDVI presented r = 0.32, the
SAVI presented the correlation r = 0.78. Although
several studies have demonstrated the functionality
of NDVI in estimating vegetation properties, many
atmospheric influences and vegetation itself restrict
their use globally.Thus, improved indexes, such as EVI
and SAVI, incorporate an adjustment factor for soils
or atmospheric conditions, generating by-products
that have a greater reliability related to the field truth.
It was found that the model proposed by SEBAL
(SAVI) presented a high correlation with the field
truth (r = 0.66). However, the model proposed by
Duchemin (NDVI) showed no correlation (r = 0.23),
being considered inefficient for the estimation of
LAI in Mixed Ombrophilous Forest. In addition, both
models presented an underestimation of the truth
of the field. The LAI estimated by the terrestrial
and orbital sensors showed a great variability
between the data, a factor that may be related to
the diversity of species, characteristic of native
forests. Among these variations, it was possible
to identify the different canopies in these models,
so that the pixels with the highest LAI values
correspond to the predominantly hardwood area.
As in the LAI estimated through the CI-110
equipment, the LAI values generated by the orbital
method also presented a great variability between
the plots. This factor may be related to the different
forest dykes predominant in the Rudi Arno Seitz
Station.
Conclusion
Remote Sensing techniques and the images
of the Pleiades sensor presented a good
performance to estimate the leaf area index
in a Mixed Ombrophilous Forest Fragment.
The SEBAL and Duchemin models presented an
underestimation of the field truth, but the SEBAL
model was considered efficient due to the strong
correlation with the control.
Reference
BAMBI, P. Variação sazonal do índice da área
foliar e sua contribuição na composição da
serapilheira e ciclagem de nutrientes na
floresta de transição no norte do Mato Grosso.
Universidad Mayor
111f. Dissertação (Mestrado em Física e Meio
ambiente), Universidade Federal do Mato Grosso,
Cuiabá, 2007.
DUCHEMIN, B.; HADRIA, R.; ERRAKI, S.; BOULET,
G.; MAISONGRANDE, P.; CHEHBOUNI, A.;
ESCADAFAL, R.; EZZAHAR, J.; HOEDJES; J.C.B.;
KHARROU, M.H.; KHABBA, S.; MOUGENOT, B.;
OLIOSO, A.; RODRIGUEZ, J.C.; SIMONNEAUX,
V. Monitoring wheat phenology and irrigation
in Central Morocco: On the use of relationships
between evapotranspiration, crops coefficients,
leaf area index and remotely-sensed vegetation
indices. Agricultural Water Management, v. 79,
p.1-27, 2006.
GUIMARÃES, M.J.M., FILHO, M.A.C., PEIXOTO,
C.P., JUNIOR, F.A.G., OLIVEIRA, V. V. M.
Estimation of leaf area index of banana orchards
using the method LAI-LUX. Water Resources
and Irrigation, v.2, n.2, p.71-76, 2013.
JONCKHEERE, I., FLECK, S., NACKAERTS, K.,
MUY, B., COPPIN, P., WEISS, M. BARET, F.,
Review of methods for in situ leaf area index
determination Part I. Theories, sensors and
ForestSAT 2016 Abstracts Summary
hemispherical photography. Agricultural and
Forest Meteorology, v. 121, 19–35, p.2004.
LARCHER, W. Ecofisiologia Vegetal. Ed. Rima. São
Carlos. 2004. 531p.
MA, H.; SONG, J.; WANG, J.; XIAO, Z.; FU, Z.
Improvement of spatially continuous forest
LAI retrieval by integration of discrete airborne
LIDAR and remote sensing multi-angle optical
data. Agricultural and Forest Meteorology, v.
189-90, p. 60-70, 2014.
SURFACE ENERGY BALANCE ALGORITHMS FOR
LAND (SEBAL). Advanced Training and Users
Manual – Version 1.0. 2002. Disponível em <ftp://
ftp.funceme.br/Cospar_Funceme_2010/CLASS_
DAY_04.11.2010/LAB/quixere/quixere/Final%20
Sebal%20Manual.pdf>. Acesso em: 20 out. 2014.
WOODGATE, W., JONES, S.D., SUAREZ, L., HILL,
M., ARMSTON, J.D., WILKES, P., BERELOV ,M.S.,
HAYWOOD, A., MELLOR, A.; Understanding the
variability in ground-based methods for retrieving
canopy openness, gap fraction, and leaf area
index in diverse forest systems. Agricultural and
Forests Meteorology, p.83-95, 2015.
299
Universidad Mayor
Latin American Forests
Evaluating the ecological vulnerability of the remaining of
Araucaria Forest – Southern Brazil
Henrique Luis Godinho Cassol¹*; Bete Caria Moraes¹; Yosio Edemir Shimabukuro¹;
¹ National Institute for Space Research, Remote Sensing Division, Avenida dos Astronautas, 1.758 - Jd. Granja, São José
dos Campos, SP, CEP 12227-010; {bete;yosio@dsr.inpe.br}.
* Corresponding author: henrique@dsr.inpe.br +55(12)32086458
Key-words: Deforestation; Araucaria Forest; Landscape Metrics; Ecological Vulnerability;
Landscape Ecology;
Abstract
Nowadays Araucaria Forest is highly fragmented and disconnected, which gives it great susceptibility
and an extreme risk of biodiversity loss. However, just few studies have tried to establishing their
actual ecological vulnerability using landscape metrics and indices. Thus, we asked: What is the
current ecological vulnerability degree of the Araucaria Forest? And what is the distribution pattern
of forest patches? To answer these questions, we choose the southeast region of Parana State as
study case. The region showed around 15,000 ha of deforestation identified between of 2000 to
2010. The pattern distribution of forest patches was calculated through landscape metrics, whereas
the ecological vulnerability degree was assessed by interpreting results. The results suggest higher
ecological vulnerability on this region due to its highly fragmented landscape. The region contains
90% of forest patches less than 100 ha area, isolated in mean at more than 1.7 km and only 7% of the
remaining forest is under protection.
Introduction
Araucaria forest is one of the seven
phytophysiognomies of the Atlantic forest - the
luxurious forest of the Brazilian coast. As occurred
throughout the Atlantic Forest, Araucaria forest
was extensively explored not only due to the high
value of the Brazilian Pine, but also to give place
to agriculture lands, pastures and recently forestry
(HIGUCHI et al. 2012). Nowadays, the remaining
forest is susceptible and fragmented. For instance,
some researchers estimate their actual covering
lying between 10 to 20% of the original forest cover
(GALINDO-LEAL & CÂMARA, 2003; RIBEIRO et al.
2009; IESB et al. 2007).
The loss of forest has serious impact on the
ecosystem, such as biodiversity decline, habitat
loss, fragmentation and decreasing environmental
condition as water quality and soil fertility
(GALINDO-LEAL & CÂMARA, 2003; RIBEIRO et
al. 2009). The fragmentation extent, including
variations in connectivity (distance between forest
patches), is an important factor to assessing the
300
state of ecological vulnerability of an ecosystem. So,
the impact of deforestation and its fragmentation
may be indirectly measured by landscape metrics,
as number and area of forest patches, connectivity
between them, and other useful landscape metrics
(TURNER, 1989).
In this context, the ecological vulnerability degree
of an ecosystem is given by its exposure, sensitivity
and resilience (FORMAN & GODRON, 1986). The
exposure is related to the structure of the forest
as size of fragments and edge effects. Sensitivity
is related to the ecosystem function such as
interactions among plants and animals or trophic
chains and finally, the resilience can be associated to
the environment changes, i.e., the ability of forest to
return to the previous state (UUEMAA et al. 2013).
Thus, a small change in one of them may increase
the vulnerability of the whole forest ecosystem.
In this way we ask: what is the current ecological
vulnerability degree of the remaining Araucaria
Forest? And what is the pattern distribution of forest
patches?
Universidad Mayor
Methodology
The study area comprises a Southeast region of
Paraná State, Brazil, as a study case (Figure 1). The
original forest cover of Araucaria Forest was 78% of
SE region (~17,000km²), and remainder of Deciduous
forest (20%) and Grasslands (2%) (IPARDES, 2004).
The main species of Araucaria forest is the emergent
Brazilian pine (Araucaria angustifolia Bertz. O. Ktze),
co-occurring with other broadleaf species as Ocotea
porosa (Mez.) L. Barroso, Ilex paraguariensis St.
Hil, Cedrela fissilis and members of Lauraceae and
Myrtaceae (STEFENON, 2009).
Dataset
Forest patches of 2000 and 2010 were downloaded
from SOS Mata Atlântica/INPE, (2010) in shapefile
format. The polygons were converted to reference
system WGS84 and reprojected to coordinate
system UTM Zone 22 South. Scale format was
1:50000 meaning that the smallest forest patches
< 5 ha were not computed. Conservation Units
(CU) areas were obtained in shapefile format from
Ministry of Environment in scale 1:250000, datum
WGS84, (www.icmbio.gov.br; ICMBIO, 2013). Rivers
channels were downloaded from Institute of Land,
Cartography and Geosciences of Paraná (www.itcg.
pr.gov.br; ITCG/PR, 2010).
Ecological Vulnerability of Araucaria
Forest
In order to evaluate the current ecological
vulnerability of the remaining of Araucaria Forest,
ForestSAT 2016 Abstracts Summary
we applied some landscape metrics to the structure,
shape and connectivity of the forest patches
(RIBEIRO et al. 2009). Forest cover data and metrics
were computed in ArcGIS 9.3, Patch Analyst 5.0
(REMPEL et al., 2012).
Major landscape metrics chosen were: patch size,
structural connectivity, shape index, mean isolation,
distance to Conservation Units and distance to river
channel (Table 1). We utilize the multiscale approach
to assess the landscape metrics. In this approach,
each metric will be separated in class rules.
The distribution of the fragments into size classes
allows us to measure a forest cover in each of the
classes and their distribution in the total area. The
shape index refers to the complexity of the shape
of the fragments. Shape index assumes value one
for geometric forms and values greater than one for
more irregular and complex shapes. This parameter
is very useful, because more irregular forms tend to
have a greater edge effect, which may be limiting for
a survival of some species (UUEMAA et al., 2013). In
structural connectivity we interested of “probable”
distance to find a given fragment size class.
In mean isolation, the smallest forest patches are
successively removed from analysis to estimate the
minimum distance of a random point to the nearest
forest patch (RIBEIRO et al. 2009). This index is used
to infer the importance of small fragments in the
landscape and its connectivity to larger forested
areas (stepping stones).
Figure 1. Location of Southeast region of Paraná State, Brazil. In detail, relative coverage of forest for each
municipality area, and forest losses and growth from 2000 to 2010.
301
Universidad Mayor
Latin American Forests
Table 1. Major landscape metrics applied to assess the ecological vulnerability degree of Araucaria Forest.
Index
Description
Class rules
Patch size
Number of forest patches
and percentage of forest cover for different size classes
Patch size classes: 0-50ha; 50-100ha; 100-250ha; 250500ha; 500-1000ha; 1000-2500ha; 2500-5000ha; >5000ha
Structural Connectivity
Mean area of interconnected
forest patches considering
minimum distance classes
Linkage distances: 0-200m; 200-500m; 500-900m; 9001400m; 1400-2000m; 2000-2700m
Shape index
Shape index per forest patches size. Closer to one more
geometric is its shape
Patch size classes: 0-50ha; 50-100ha; 100-250ha; 250500ha; 500-1000ha; 1000-2500ha; 2500-5000ha; >5000ha
Mean isolation
Mean isolation of random
points to the nearest forest
patch. To evaluate isolation
the smallest forest patches
were successively removed
Size of forest patches removed: 0-50ha; 50-100ha; 100250ha; 250-500ha; 500-1000ha; 1000-2500ha;
Distance to CU
Minimum distance of any
forest patch to the nearest
Conservation Unit
Distance class: 0m (inside CU); 0-500m; 5000-1000m;
1.000-2.000m; 2.000-5.000m; 5.000-10.000m; 10.00020.000m; 20.000-50.000m; >50.000m
Distance to rivers
Minimum distance of any
forest patch to the nearest
river channel
Distance class: 0m (touch the river channel); 0-50m; 50100m; 100-250m; 250-500m; 500-1000m; 1000-2500m;
>2500m
Results and Discussions
100%
40%
80%
30%
60%
40%
20%
0%
%N
a)
We also noticed that 36% of forest patches are
distant less than 50 m of any channel river, but its
represents only 6 % of forest area (Fig. 2b). This
situation is worrying because the areas close to the
rivers are protected by law and are a refuge of the
wild allowing the flow of plants and animals.
% Class
% Class
Remaining forest cover of SE region is 17.5 % and
9% of Araucaria Forest estimated by Ribeiro et al.
(2009). The Figure 2 shows the relative distribution
of forest patches by number and area against size
class, respectively, and the minimum distance of any
forest patch to the nearest river channel by distance
classes. We can see that 76% of forest patches
have less than de 50 ha area, but it represents only
26% of the forested area (Fig. 2a). The largest
forest fragments are located in the CU of Serra da
Esperança (A > 5000 ha), between the municipalities
of Cruz Machado, Mallet and União da Vitória,
corresponding to 8% of remaining forest area (Fig.
2a).
1
2
3
4
5
6
7
10%
0%
8
%N
77% 14% 7% 2% 1% 0% 0% 0%
% Area 26% 17% 17% 10% 8% 6% 8% 8%
20%
b)
1
2
3
4
5
6
7
8
37% 5% 5% 13% 14% 16% 9% 0%
% Area 6% 1% 2% 5% 9% 18% 37% 21%
Figure 2. a) Distribution of forest patches by frequency and area by size classes: (1) 0-50ha; (2) 50-100ha; (3)
100-250ha; (4) 250-500ha; (5) 500-1000ha; (6) 1000-2500ha; (7) 2500-5000ha; (8) >5000ha. b) Distance (m) of
any forest patch to the nearest river channel by relative frequency and area by distance classes: (1) 0m (touch
the river channel); (2) 0-50m; (3) 50-100m; (4) 100-250m; (5) 250-500m; (6) 500-1000m; (7) 1000-2500m;(8)
>2500m.
302
ForestSAT 2016 Abstracts Summary
10.0
30000
8.0
6.0
20000
4.0
10000
0
1
2
3
4
5
6
7
8
0.0
1
2
3
4
5
6
7
8
9
%N
6% 2% 2% 4% 11% 19% 29% 24% 5%
Connectivity 933 1113 1235 1453 1832 1970 3078 1770
%A
7% 2% 2% 3% 7% 15% 19% 15% 29%
Shape Index
71
Isolation
a)
2.0
35%
30%
25%
20%
15%
10%
5%
0%
1768 2240 5422 7797146442669834113
1.4
1.8
2.2
2.7
2.9
3.2
4.3
8.7
b)
400
350
300
250
200
150
100
50
0
Mean Area (ha)
40000
% Calss
Distance (m)
Universidad Mayor
61
54
50
40
48
39
38
335
Figure 3. a) Mean isolation (m) of forest patches by size class (ha); mean distance (m) of structural connectivity
among forest patches with the same size class, and shape index (unitless) by size class (ha). Size classes: (1)
0-50 ha; (2) 50-100 ha; (3) 100-250 ha; (4) 250-500 ha; (5) 500-1000 ha; (6) 1000-2500 ha; (7) 2500-5000 ha e
(8) >5000 ha. b) Minimum distance to of any forest patch to the nearest Conservation Unit by frequency and
area by distance class: (1) 0m (inside CU); (2) 0-500m; (3) 5000-1000m; (4) 1.000-2.000m; (5) 2.000-5.000m;
(6) 5.000-10.000m; (7) 10.000-20.000m; (8) 20.000-50.000m; (9) >50.000m. Continuous line refers to mean
area (ha) of forest patch by distance class.
Figure 3a presents mean isolation, strutural
connectivity and shape index by size of forest
patches. The mean isolation between a random
point to any forest patch was more than 1700m. If
the smallest patches were lost the average isolation
would raise exponentially. Structural connectivity
between forest patches with less than 50 ha was
calculated in 933 meters. As expected, small forest
patches are squared while biggest ones showed
more complex shapes (Fig. 3a).
The size, shape, proximity and isolation affects
directly the structure, function and biodiversity
of the remaining forest (GALINDO LEAL, 2003).
Small forest patches reduce diversity of birds
(ANJOS, 2001), tree species (SCHESSL et al. 2008)
and specially great mammals (CHIARELLO, 1999).
Considering the isolation and level of fragmentation,
species less generalist and with greater need of area
are the first to disappear (PARDINI et al., 2009).
The isolation of species due to fragmentation has
also negative effect to genetic conservation of
Araucaria angustifolia by increasing the endogamic
reproduction (STEFENON et al. 2009). Small forest
patches are not able to provide the self-reproduction
of individuals of Brazilian Pine, suggesting that the
colonization of border areas or abandoned areas
may not be sufficient for the conservation of this
important forest species (SOUZA et al. 2008).
In relation to the minimum distance to Conservation
Units, 5.75 % of remaining forests are contained in
any CU (7% of forest area), it means they are under
protection. The mean area size under protection is
only 71 ha (Figure 3b). These circa 6% of protection
is below the recommended 10% that must be
necessary to hotspots conservation by Secretariat
of the Convention on Biological Diversity (2002).
On this scenario, the creation of new CUs should be
priority to reduce the losses of biodiversity.
Conclusions
The Araucaria forest is highly fragmented and
susceptible: 91% of their fragments have less than
100 ha area; they are isolated in mean to 1.7km
and structural connectivity between forest patches
are above 900 m. Nevertheless, the Southeast of
Paraná State has still 17% of original forest cover;
7% of which are inside of Conservation Units. The
major challenge is connect the forest patches with
ecological corridors and above all connect them with
river channels (only 6 % of patches are less than 50
m of any river stream) combining the protection
of water sources with the possibility of animal
movement in these areas.
References
ANJOS, L. Bird communities in five atlantic forest
fragments in southern Brazil. Ornitologia
Neotropical. v.12. p. 11–27. 2001.
CHIARELLO, A. G. Effects of fragmentation of
the Atlantic Forest mammal communities in
southeastern Brazil. Biological Conservation. v.
89: p.71–82. 1999.
303
Latin American Forests
GALINDO-LEAL, C., CÂMARA, I.G., 2003. Atlantic
Forest hotspot status: an overview. In: GALINDOLEAL, C., CÂMARA, I.G. (Eds.), The Atlantic
Forest of South America: Biodiversity Status,
Threats and Outlook. CABS and Island Press,
Washington, p 3–11. 2003.
HIGUCHI, P.; SILVA, A.C.; FERREIRA, T.S.; SOUZA,
S.T.; GOMES, J.P.; SILVA,K.M.; SANTOS, K.F.
Floristic composition and phytogeography of the
tree component of Araucaria Forest fragments in
southern Brazil. Brazilian Journal of Botany, v.
35. n.2 p. 145-157. 2012.
INSTITUTO DE TERRAS, CARTOGRAFIA E
GEOCIÊNCIAS
DO
PARANA
(ITGC/PR).
Secretaria do meio ambiente do estado do
Paraná. 2010. Available: www.itgc.pr.gov.br.
Accessed: 20/07/2013.
INSTITUTO DE ESTUDOS SOCIOAMBIENTAIS DO
SUL DA BAHIA (IESB). Instituto de Geociências
da Universidade Federal do Rio de Janeiro
(IGEO/UFRJ), Departamento de Geografia da
Universidade Federal Fluminence (UFF), 2007.
Levantamento da Cobertura Vegetal Nativa do
Bioma Mata Atlântica. Relatório final. PROBIO
03/2004, Brasília, 84p.
INSTITUTO PARANAENSE DE DESENVOLVIMENTO
ECONÔMICO E SOCIAL (IPARDES). Leituras
regionais: mesorregião geográfica do Sudeste do
Paraná. Curitiba, 2004. 134p.
PARDINI, R.; FARIA, R.; ACCACIO, G.M.; LAPS,
R.R.; MARIANO-NETO, E.; PACIENCIA, M.L.B.;
DIXO, M.; BAUMGARTEN, J. The challenge of
maintaining Atlantic forest biodiversity: A multitaxa conservation assessment of specialist and
generalist species in an agro-forestry mosaic in
southern Bahia. Biological Conservation, v. 142.
p. 1178–1190. 2009.
304
Universidad Mayor
REMPEL, R.S., D. KAUKINEN., AND A.P. CARR.
2012. Patch Analyst and Patch Grid. Ontario
Ministry of Natural Resources. Centre for
Northern Forest Ecosystem Research, Thunder
Bay, Ontario. [online: http://www.cnfer.on.ca/
SEP/patchanalyst/].
RIBEIRO, M.C.; METZGER, J.P.; MARTENSEN, A.C.;
PONZONI, F.J.; HIROTA, M.M. The Brazilian
Atlantic Forest: How much is left, and how is the
remaining forest distributed? Implications for
conservation. Biological Conservation, v. 142. p.
1141–1153. 2009.
SCHESSL, M.; SILVA, W.L.; GOTTSBERGER, G.
Effects of fragmentation on forest structure
and litter dynamics in Atlantic rainforest in
Pernambuco, Brazil. Flora. v. 203. p. 215–228.
2008.
SOS Mata Atlântica. Shapefiles dos remanescentes
florestais 2000 e 2010. Available: http://mapas.
sosma.org.br/dados/. Accessed: 23/07/2013.
SOUZA, A.F.; FORGIARINI, C.L.; SOLON, J.; BRENA,
D.A. Regeneration patterns of a long-lived
dominant conifer and the effects of logging in
southern South America. Acta Oecologica, v. 34,
n. 2, p. 221–232, Sep. 2008.
STEFENON, V.M.; STEINER, N.; GUERRA, M.P.;
NODARI, R.O. Integrating approaches towards
the conservation of forest genetic resources: a
case study of Araucaria angustifolia. Biodiversity
and Conservation, v. 18. p.2433–2448. 2009.
TURNER, M.G.; O’NEILL, R.V.; GORDON, R.H.;
MILNE, B.T. Effects of changing spatial scale on
the analysis of landscape pattern. Landscape
Ecology, v.3, p.153-162. 1989.
UUEMAA, E.; MANDER, U.; MARJA, R. Trends in the
use of landscape spatial metrics as landscape
indicators: A review. Ecological Indicators. v.28.
p. 100–106. 2013.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Improving Observations of Tropical Forests with Optimized
Terrestrial Lidar Scanners
Crystal Schaaf1, Ian Paynter1, Edward Saenz1, Daniel Genest1, Francesco Peri1, Zhan Li1, Alan Strahler2, David Clark3,
Deborah Clark3, William Miranda4, Leonel Campos4
1
School for the Environment, University of Massachusetts Boston, USA
2
G eography Department, Boston University, USA
3
Department of Biology, University of Missouri St. Louis
4
La Selva Biological Station, Sarapiqui, Costa Rica
Terrestrial lidar scanners (TLS), which utilize
light detection and ranging (lidar) to create 3D
representations of environments, are increasingly
being used to capture properties of interest, such
as tree volume for deriving biomass, in a variety of
ecosystems. Tropical ecosystems present particular
challenges to TLS observation. The high level of
geometric complexity and the vertical stratification
of tropical forests creates complex occlusion, and
the inaccessibility, difficult weather conditions and
other deployment constraints limit observation
opportunities. TLS optimized for scanning speed,
portability and resilience, such as the Compact
Biomass Lidar (CBL, University of Massachusetts
Boston) can be used to meet these challenges and
facilitate high quality TLS observation of tropical
forests over meaningful spatial extents. The rapid
(33 second) scans permit a large number of scan
positions to be acquired in a single acquisition effort,
in order to mitigate occlusion. The light weight
(3.6kg) of the instrument enables deployment on an
11 meter high tripod, which facilitates sampling at
multiple heights at each scanning location.
Herein we present results from TLS observations of
a long-term CARBONO site at La Selva Biological
Station, Costa Rica. These data have been explored
through novel analysis techniques to describe forest
structure, as well as describe the TLS observation
density and quality. Occlusion patterns are described,
and used to infer sampling techniques that will further
improve observation quality. The contribution of
scans of different heights reveal that vertical scan
positions offer not only unoccluded observations
above the sub-canopy but also provide unique
information about the sub-canopy due to the
complexity of view-angles possible in the tropical
environment.
305
Universidad Mayor
Latin American Forests
Mapping forest degradation in the Valdivian Temperate
Rainforest ecorregion
Díaz-Hormazábal I.1, Chávez R. O. 2, Gutiérrez A. G. 1
1
Universidad de Chile, Facultad de Ciencias Agronómicas, Departamento de Ciencias Ambientales y Recursos Naturales,
Av. Santa Rosa 11315, La Pintana, Santiago, Chile.
2
Pontificia Universidad Católica de Valparaíso, Instituto de Geografía, Av. Brasil 2241, Valparaíso, Chile.
Key words: Forest gaps, forest degradation, image segmentation, WorldView3.
Governments have committed to mitigate climate
change by reducing forest deforestation and
degradation. In developing regions such as southern
South America, where large tracks of primary
forests still exist, this task is increasingly needed,
as people subsistence demands the use of intact
forests. However, there is a knowledge gap on 1)
where primary forests are located and 2) where
forest degradation is occurring. Such information is
necessary to support decision making for planning
forest management. Here, we present an approach
for mapping degradation of primary forests in
the Valdivian Temperate Rainforest ecorregion
in south-central Chile (39-40°S). We based our
analysis on forest gap structure of primary forests
using 1) ground-based sampling, 2) automatized
identification of gaps using high-resolution satellite
images, and 3) upscaling gaps metrics to regional
scale using lower resolution Landsat images. We
used sub-metric satellite images (50 cm) from
four study sites representing primary forests.
306
We segmented the panchromatic band using the
Mean-Shift algorithm. We then calculated the
Enhanced Vegetation Index (EVI) for each segment
and classifyied segments with low EVI as potential
gaps. We calculated area of segments distinguished
as gaps for each forests. Gap segmentation had an
average of 78% concordance with gaps identified
in the field. We found a good relation between gap
areas estimated from high resolution images to
vegetation index obtained from Landsat-8 images
allowing us to map the location of forests with low
gap areas (i.e. low degradation). Intact forests often
had lower gap frequency (e.g. gap area <10% of total
forest area). Our analyses suggest that 57.3% of oldgrowth forests in the study area have some degree of
degradation. We propose that primary forests have
a canopy structure that make them distinghishable
from degraded old-growth forests. Our work is a
step forward towards monitoring forests at large
spatial scales and a contribution for lowering forest
degradation in developing countries.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Modeling aboveground biomass from individual tree
LiDAR-derived metrics in tropical forest
Carlos Alberto Silva1,2,3, Andrew Thomas Hudak2, Lee Vierling1, Carine Klauberg2, António Ferraz3, Mariano Garcia Alonso3,
Michael Keller4, Sassan Saatchi3
1
Department of Natural Resources and Society, College of Natural Resources, University of Idaho, (UI), 875 Perimeter
Drive, Moscow, Idaho, 83844, USA
2
USDA Forest Service, Rocky Mountain Research Station, RMRS, 1221 South Main Street, Moscow, ID, 83843, USA
3
Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA 91109, USA
4
USDA Forest Service, International Institute of Tropical Forestry, San Juan, PR, 00926, USA
Keywords: Tree detection, Airborne LiDAR, Biomass, Tropical Forest, Forest Inventory
Tropical forest play an important role in mitigating
global warming. Airborne LiDAR (Light Detection
and Ranging) has been used to predict aboveground
biomass in tropical forest. However, few studies
have modeling and predicting tropical forest AGB
at plot-level from individual tree LiDAR-derived
metrics. The aim of this study was to model and
predict AGB at plot-level from individual tree
LiDAR-derived metrics in a Brazilian tropical forest.
Forest attributes such as tree density and diameter
at breast height were measured in 2014 at 84
square plots of 50x50m located in Paragominas,
Pará, Brazil. Using previously published allometric
equations, tree AGB was computed and then
summed to calculate total AGB at each sample plot
(TAGB). The LiDAR data also acquired in 2014 were
normalized to height aboveground, and individual
trees were detected using the Adaptive Mean Shift
3D (AMS3D) algorithm. Individual tree metrics,
such as height (TH), crow length (CL), crown radius
(CRd), crown ratio (CRt), crown based height (CBH),
crown projected area (CPA) and crown volume (CV),
Figure 1. LiDAR point cloud with colour coded by height A) and tree id derived from individual tree detection B);
Virtual Forest representation with trees colours coded by a single colour C) and by height D).
307
Latin American Forests
were calculated to each single tree using the rLiDAR
package, and summarized at plot-level as maximum
tree height (MAXTH), mean tree height (MEANTH),
mean crown length (MCL), mean crown radius
(MCRd), mean crown ratio (MCRr), mean crown
based height (MCBH), total crown volume (TCV),
and total projected area (TPA). Regression models
predicting TAGB from the individual tree LiDARderived metrics were developed and evaluated for
predictive power and parsimony. The best model
from a family of six models was selected based on
corrected Akaike Information Criterion (AICc) and
assessed through a leave-one out cross validation by
the adjusted coefficient of determination (adj.R²),
308
Universidad Mayor
relative and absolute root mean square error (RMSE)
and BIAS. Our findings showed that MAXTH, MCBH,
MCRd, and TCV were the most important metrics
to predict TAGB, with overall model adj.R² of 0.73,
absolute and relative RMSE and BIAS of 26.9 and
0.29 Mg/ha and 16.2 and 0.17%, respectively. This
study showed that individual tree LiDAR-derived
metrics can be used to predict TAGB in tropical forest
with acceptable precision. Therefore, we conclude
that there is good potential to monitor carbon
sequestration in Brazilian Tropical Rain Forest using
airborne LiDAR data and individual tree LiDARderived metrics.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Phenological observations from a hyperspectral camera in the
Amazonian Tapajos National Forest
Yhasmin M. de Moura1, Lênio Soares Galvão1, Thomas Hilker2, Cibele H. do Amaral3, Scott R. Saleska4, Jin Wu5,
Luiz E. O. C. Aragão1
Instituto Nacional de Pesquisas Espaciais (INPE), São José dos Campos - SP, 12227-010, Brazil, 2University
of Southampton, Department of Geography and Environment, Southampton SO17 1BJ, United Kingdom,
3
Federal University of Viçosa, Viçosa - MG, 36570-900, Brazil, 4University of Arizona, Department of Ecology
and Evolutionary Biology, Tucson, AZ 85721, USA, 5Environmental & Climate Sciences Department,
Brookhaven National Lab, Upton, New York, NY, 11973
1
Keywords: phenology; hyperspectral camera; principal component analysis; tropical ecosystem;
mixture model
Understanding seasonal and inter-annual variability
of vegetation phenology plays a major role for
Earth system modeling. However, scaling field
observations of changes in vegetation traits to
satellite remote sensing is often challenging due
to a mismatch in scale. Near surface camera based
phenology can help bridging this gap by using
optical principles similar to those used by satellite
sensors, but still allowing an interpretation similar
to that made by field observations. In this study, we
utilized data from a tower-mounted, hyperspectral
phenocam imaging system (SOC-710, Surface
Optics Corporation, San Diego, CA, US) located at
the Tapajós National Forest near Santarem, Pará,
Brazil (2.85˚ S, 54.97˚ W) to observe the phenological
behavior of a tropical forest ecosystem during the
dry season. Seventeen images (4-nm resolution
between 385 to 1050 nm) were carefully selected and
screened for data quality and clear sky conditions.
Reflectance measurements were calibrated using
a teflon panel, which was permanently installed
within the field-of-view of the camera. Images were
acquired at an angle of 45° off-nadir around solar
noon to minimize differences of shading effects
within the canopy. Principal component analysis
(PCA) showed that variations in leaf flushing
between July 27 and September 29, 2012 were
species-dependent. Using a linear mixture model,
we confirmed that differences observed from PCA
were expressed in the green vegetation fraction of
the hyperspectral images. We conclude that the data
analysis of tower-mounted hyperspectral cameras
contributes to a better understand of the seasonal
phenological signal observed by satellites.
309
Universidad Mayor
Latin American Forests
Recent trends of Land Surface Temperature and Vegetation
Indexes over the Temperate Rain Forest in Chile
V. Olivares-Contreras, C. Mattar, J.C. Jiménez-Muñoz and A. Gutiérrez.
Laboratory for Analysis of the Biosphere (LAB), Dept. of Environmental Sciences and Natural Renewable Resources,
University of Chile, Av. Santa Rosa 11315, La Pintana, Santiago, Chile.
2
Global Change Unit/Image Processing Laboratory, University of Valencia, Valencia 46980, Spain.
3
Dept. of Environmental Sciences and Natural Renewable Resources, University of Chile, Av. Santa Rosa 11315,
La Pintana, Santiago, Chile.
1
Key words: Temperate rain forest, land surface temperature, vegetation indexes, Modis, Aysen, Chile.
The current global warming has affected several
ecosystems at global scale generating different
consequences such as the decrees in carbon stocks
and productivity, the increase in mortality and
the persistent increment of fire risk. In Chile, one
of the most pristine ecosystems is the temperate
rain forest located in the Aysén region (43° - 50°S).
This rain forest covers more than 4.82 millions
of hectare which are mainly located in island and
isolated peninsula. This forest has suffered a strong
anthropogenic impact during the 30’ based on the
government policy to introduce grassland through
conversion plans from forest to agriculture terrains.
More than 6 millions of hectare were burned or
affected decreasing the forest land cover and
generating a dramatic change in this ecosystem.
The impact of the global warming over the
temperate rain forest is unknown. Thus, the aim of
this work is to analyze the spatial and temporal land
surface temperature and vegetation indexes trends
by using MODIS data (MOD11 and MOD13 products)
for the period covering 2002 – 2015. The LST, NDVI
and EVI were used to estimate the non-parametric
trends in the Aysén region (43° - 50°S ; 71 – 77°W)
310
by using the Sen’s slope test and to determine
the statistical significance applying the MannKendall test. Trends and statistical significance
were retrieved pixel by pixel at monthly mean and
seasonal periods assuming Summer (December,
January and February), Autumn (March, April and
May), Winter (June, July and August) and Spring
(September, October and November).
Results shown a generalized warming between 43°
and 47°S estimated in 1°C/decade. Southern the
47°S the land surface temperature trends were also
positive although it was noticed some effects of the
spatial resolution of MOD13 generating an impact
in te statistical confidence of the trend. In relation
to the vegetation indexes, at overall, the trends
are positive in the whole region at evaluating NDVI
and EVI. Nevertheless, a negative and statistical
significance trends were evidenced between 46° –
48°S mainly in 73°W is also retrieved. These results
contribute to better understand the dynamic of the
Aysén Temperate forest and to show the impact of
global warming in the land surface temperature and
vegetation trends in this Austral zone of Chile.
RDD+FREL-FRL and MRV
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Assessing the carbon and water balance of Boreal forests
using a process-based model driven by satellite images
Minunno F., Mutanen T., Aurela M, Häme T, Liski Y, Vesala T, Mäkelä A.
Keywords: Process-based modelling, Satellite Images, Carbon and water balance,
Boreal forests, data assimilation.
Abstract
Earth Observation (EO) data can potentially be used to drive vegetation models and improve simulated
carbon and water fluxes over large areas.
The objective of this work is to test the reliability of a process-based model in predicting the carbon
and water balance of Boreal forests when driven by EO measurements. We used a carbon-balance
based growth model combined with a soil carbon model.
After being calibrated and tested with ground data from various data sources the combined model
was run using LANDSAT data collected in 2015 at 20m resolution. Model predictions were validated at
two eddy-covariance sites: Hyytiälä in southern Finland and Sodankylä in Finnish Lapland.
At Hyytiälä, model predictions of annual gross primary production (GPP, 1143.9 gC m-2 y-1), annual net
ecosystem exchange (NEE, -205.6 gC m-2 y-1) and evapotranspiration (ET, 314.6 mm y-1) were within
the ranges measured at the eddy-site.
At Sodankylä GPP (573.1 gC m-2 y-1) and ET (236.4 mm y-1) predictions were consistent with the
measurements, while NEE (-87.27 gC m-2 y-1) was overestimated.
On this basis, we discuss the opportunities and challenges of using EO data for large scale simulations
with the aim of monitoring carbon and water fluxes.
313
Universidad Mayor
RDD+FREL-FRL and MRV
Comparison of local EO-based dense humid and dry
forest cover and change area estimates in the southwest
forest massif of Central African Republic using the UMD
global dataset
Christophe Sannier1, Ron Mc Roberts2, Loïc Faucqueur1 & Hajar Benelcadi1
1
SIRS, Parc de la Cimaise I, 27 rue du Carrousel 59650 Villeneuve d’Ascq France
Northern Research Station, U.S. Forest Service, Saint Paul, Minnesota, USA
2
Keywords: Model assisted regression, REDD+, national forest monitoring systems
The southwest forest massif of Central African Republic is covered by dense humid forest but also by
substantial dry forest areas. Most of current forest
monitoring studies focuses on tropical rain forest
because of their ecological interest and high carbon
stocks. But about 42% of tropical forests around the
world are considered as seasonally dry. Therefore,
dry Forest monitoring should not be overlooked because it is found over very large areas and is one of
the most seriously threatened forest types. Indeed,
the removal of dry forests potentially contributes
to a significant proportion of the 7-20% of the total
greenhouse gas emissions known to originate from
deforestation and forest degradation. However,
monitoring of dry forest from remote sensing is a
challenge due to its phenological variability and the
lack of hard boundaries between forest and non-forest.
An EO-based, IPCC compliant Land Use classification of the area was developed that distinguished
between dense humid and dry forest classes was developed for an area of approximately 70,000km² in
southwest CAR in collaboration with the Ministry of
environment. In addition, a probability sample was
developed to calibrate the identification of the lower boundary of the forest category (i.e. 30% forest
cover) based on the visual estimation of percent tree
cover from Very High Spatial Resolution satellite
imagery. This dataset was also used to calibrate the
University of Maryland (UMD) Global Forest Change
314
(GFC) data and optimal thresholds were determined
based on the comparison of tree cover percentage
calibration data with map data to match the selected national forest definition. Then, a model assisted
regression was developed based on a separate probability sample over the study area.
Wall-to-wall mapping provides a comprehensive
assessment of forest resources and input to land
use plans for management purposes, but wall-towall approaches require specialized equipment and
staff that are often not available. The UMD GFC map
products provide an alternative for tropical countries wishing to develop their own wall-to-wall forest
monitoring map products but without the resources to do so. However, these maps were produced at
global scale and cannot be used directly for national/
regional mapping because they need to comply with
a selected forest definition. This study provides an
approach on how the UMD GFC map can be used locally and a quantification of the loss of precision is
made. Initial results suggest that the number of sample units would need to be increased by 44% for the
area estimates based on the UMD GFC to be equivalent to the locally based map. The cost of collecting
these additional samples is discussed with a view to
providing guidelines on whether the additional cost
is worthwhile in relation to (i) the additional income
generated through a performance based payment
scheme, and (ii) the cost of developing a local based
capacity to map forest cover.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Comparison of local EO-based dense humid and dry
forest cover estimates with the UMD global dataset in
the Central African Republic
Hajar Benelcadi1, Christophe Sannier1, Loïc Faucqueur1 & Ron Mc Roberts2
1
SIRS, Parc de la Cimaise I, 27 rue du Carrousel 59650 Villeneuve d’Ascq France
Northern Research Station, U.S. Forest Service, Saint Paul, Minnesota, USA
2
Keywords: REDD+, Model Assisted Regression, forest estimates, dry forest, UMD GFC, relative efficiency
Abstract
This study provides an approach on how the UMD GFC map can be used locally and a quantification of
the loss of precision is made. Initial results suggest that the number of sample units would need to be
increased by 48% for the area estimates based on the UMD GFC to be equivalent to the locally based
map. The cost of collecting these additional samples is discussed with a view to providing guidelines
on whether the additional cost is worthwhile in relation to (i) the additional income generated through
a performance based payment scheme, and (ii) the cost of developing a local based capacity to map
forest cover.
Introduction
Most of current forest monitoring studies focuses
on tropical humid forest because of their ecological interest and high carbon stocks. But about 42%
of tropical forests around the world are considered
as seasonally dry. Therefore, dry Forest monitoring
should not be overlooked because it is found over
very large areas and is one of the most seriously
threatened forest types. The removal of dry forests
potentially contributes to a significant proportion
of the total greenhouse gas emissions known to originate from deforestation and forest degradation.
However, monitoring of dry forest from remote sensing is a challenge due to its phenological variability
and the lack of hard boundaries between forest and
non-forest. The REDD+ mechanism is included in the
Paris agreement representing the main outcome of
the UNFCCC COP21 after a long round of negotiations, as a key instrument to enhance the forests
sinks and preserve the forest. Many countries having
the tropical forest biome engaged or willing to engage under the REDD+ mechanism are developing
a National Forest Monitoring System to assess and
monitor their forests. However REDD+ is a result
based mechanism. Meaning that the payments will
be made to countries providing convincing evidence
that reduction targets have been achieved. The evidence must include accurate and precise estimates
of forest area that will be submitted by countries as
reference emissions levels to be reviewed as part of
the Monitoring Reporting and Verification process of
the REDD+ mechanism. Wall-to-wall mapping provides a comprehensive assessment of forest resources
and input to land use plans for management purposes, but wall-to-wall approaches require specialized
equipment and staff that are often not available.
The UMD GFC map products provide an alternative
for tropical countries wishing to develop their own
wall-to-wall forest monitoring map products but without the resources to do so. However, these maps
were produced at global scale and cannot be used
directly for national/regional mapping because they
need to comply with a selected forest definition.
Therefore, this paper aims to assess if global datasets such as the University of Maryland (UMD) Global
Forest Change (GFC) data can be used to fast track
the development of National Forest Monitoring Sys-
315
Universidad Mayor
RDD+FREL-FRL and MRV
tem (NFMS) in countries with complex landscapes
such as the Central African Republic in which several
biomes including dense tropical rainforest and dry
forest are present.
Study site
The study area is located in the southwest forest massif of Central African Republic represented by 3 provinces: Sangha Mbaéré, Lobaye, Ombella-M’Poko.
The total area is ~ 70 000 km². The study site was
selected because covered by dense humid forest but
also by substantial open dry forest areas. This was
part of the European FP7 project that was conducted
in collaboration with GAF AG and with the local Ministry of Environment and Ecology (MEE). The adopted forest definition is as follow: 1ha minimum area,
>30% tree cover, tree height >5m at maturity. Tree
plantations are excluded from forest area.
Data
A set of VHR data was acquired circa 2010 as part of
the FP7 REDDAF project through the ESA Data Warehouse according the location of Primary Sample
Unites (PSUs) from the probability sampling design.
All VHR Data were pan-sharpened and resampled to
an output pixel size of 0.5 meters to provide a homogeneous dataset optimized for visual interpretation.
Methodology
An IPCC compliant Land Use classification was developed that distinguished between dense humid and
dry forest classes. In addition, a probability sample
was implemented to calibrate the identification of
the lower boundary of the forest category (i.e. 30%
forest cover) based on the visual estimation of percent tree cover from Very High Spatial Resolution
satellite imagery. This dataset was also used to calibrate the University of Maryland (UMD) Global Forest Change (GFC) data and optimal thresholds were
determined based on the comparison of tree cover
percentage calibration data with map data to match
the selected national forest definition as suggested
by Sannier et al (2016). Then, a model assisted regression (McRoberts et al. 2012) was developed based on a separate probability sample over the study
area.
Processing of UMD GFC data to match selected forest definition in CAR for 2010
To process the UMD GFC dataset the following chain
is applied:
1) Application of the 30% tree cover percentage
threshold to the 2000 UMD GFC tree cover percentage map data 2000
2) Integration of « Forest Gains» within the NF class
defined
3) Integration of « Forest Losses » within the F class
defined
4) The circa 2010 F/NF map results from the combination of the 2000 F/ NF map with the forest gain
and losses
5) A 1 ha MMU filter is applied to both the F and NF
class
VHR data sampling design
The probability sampling design was performed with
simple random selection of sample units. A total of
46 segment of 5*5km based on an unaligned systematic random sample over the study area (~1.5%).
From 46 segments: a) 16 segments to calibrate the
GFC dataset and b) 30 segments to estimate the forest cover proportion. A set of 10 segments of 1 hectare were randomly selected within each 25 originals
5*5km segments for a total of 250 plots of 1ha. A
Table1: VHR data acquired circa 2010 processed and photo interpreted as
an alternative of ground truth data
316
Sensor
Number of scenes
Resolution (m)
GEOEYE
Ikonos
Worldview2
QuickBird
22
4
18
2
MS: 1.65/Pan:0.41
MS 3.2/Pan:0.82
MS: 1.85/Pan:0.46
MS: 2.4/Pan:0.61
Universidad Mayor
ForestSAT 2016 Abstracts Summary
number of 49 systematic pixels by plots were selected to be interpreted into Tree/ No-Tree to extract an
estimation of the tree cover within each 1ha plots
Results
Adjustment of the GFC Forest threshold
The VHR tree cover data and the GFC tree cover data
set were correlated to determine the adjusted forest
threshold of GFC data corresponding to 30% tree cover over the VHR data (Sannier et al. 2016).
Once the threshold determined (figure 1), we apply
it to the GFC tree cover map of the year 2000. And
to match the 2010 national data, we integrate the
gain and the forest losses and we apply a 1 ha MMU
so that the adjusted GFC map can be conform to the
adopted national forest map definition. The GFC
threshold of 51% appears to be better corresponding
to a tree cover of 30%.
Area uncertainty estimates for National and GFC
maps and the relative efficiency
Maps suffer from bias and the method of model assisted regression (Sannier et al. 2014) takes it into
consideration and correct the estimate of forests.
Forest cover and forest cover change estimates can
be produced based on samples alone (Direct estimate). Observations from reference samples and the
map can be combined to improve the precision of
Figure2: The national map with 30% tree cover and
the equivalent GFC map adjusted to 51% tree cover
Figure1: Scatterplot of average GFC values over 1ha plots against estimated tree
cover values from VHR imagery
317
Universidad Mayor
RDD+FREL-FRL and MRV
estimates (Model Assisted Regression), see (1) and
(2):
Pˆ MAR
Vaˆr Pˆ MAR
Pˆ map Biˆas Pˆ map
1
(1)
¦ 'i '
m
m m 1 i 1
2
(2)
The ratio between the variances of the direct area
estimate and the map based estimate gives the relative efficiency (McRoberts et al. 2016) of the map
which is:
t1
Vâr ( P )
Vâr ( P map )
,
t
Vâr ( P GFCmap )
Vâr ( P Nationalmap )
With the national map we are underestimating the
forest area with nearly 4205 km², but this is less than
applying the 30% threshold uncalibrated GFC global
map which is overestimating the forest proportion
with 8491km². However we have a relative efficiency of more than 2, which mean that direct estimate variance of the national map is higher than the
model assisted estimate and the number of sample
units would need to be increased by a factor of 2 if
the same level of precision was to be achieved with
samples alone as compared to the model assisted
approach combining the sample units with the national map (table2). The calibrated GFC seems to reduce the map bias compared to the national map.
This means that the calibration procedure effectively works. However, the 95% Confidence intervals
are considerably wider compared with the National
map resulting in a lower relative efficiency (Table2).
In fact it is possible to compute the relative efficiency
of the adjusted GFC map versus the national map,
we would need 48% more samples to have the same
level of uncertainty to that of the national map.
The Cost of collecting Additional reference data
set
The use of UMD GFC will be economically efficient if
the cost of collecting the additional n1-n samples is
greater than the cost of the National map minus the
cost of GFC data processing, given by:
n(t 1) p ! M
Nationa l
Study a rea
(km ²)
Direc t Estim a te a t 95%
Confidenc e Interva l
Map Bia s
MAR Estim ate
(%)
6010
8
85,4
(km ²)
±857
(%)
±1,21
(%)
(km ²) -4205
5,97
8491 -3699
12.06 5.26
(km ²) 61 328
59 429 60 402
(%)
84,4
85.8
±802 ±645
87.1
(km ²) ±529
MAR Estim ate at 95%
Confidenc e Interva l
(%)
Rela tive effic ienc y without
map
Rela tive effic ienc y of nationa l
vs GFC51
GFC51
70 391
(km ²)
Direc t Estim a te
GFC30
±0.75
±1.14
±0.91
2.62
1.13
1.76
1.48
Table2: Results of the Model Assisted regression estimates and the relative efficiency to compare the National map, the GFC at 30% threshold and the adjusted GFC
t<1, then
318
P GFCmap
ha s sma ller va ria nc e tha n
P Nationalmap a nd P Nationalmap
is less e ffic ient tha n P GFCmap .
Universidad Mayor
n is the original sample size, p is the cost of acquiring
each additional sample observation and M is the
additional cost of producing the national map compared to GFC.
For this case study:
- n=30 segments, t=1.48,
- With an image costs= 25 $ /km2, (min 25km2 for
archived image + sample interpretation) or
35$/km2 and (min 100 km2 for new acquisition+
sample interpretation).
- P= 17500$-98000$ this covers quite a range depending on whether the VHR imagery is available
from an archive or new acquisition need to be
programmed.
The use of the UMD GFC data will be economical if
P is less than the cost of producing the national
map minus the cost of processing the UMD GFC
data
Conclusions
UMD GFC data set can be processed for producing
national forest cover estimates. In addition, the calibration procedure is effective and reduces substantially map bias. 95% Confidence Intervals greater
by about 20% compared with that of national data
for forest cover area. To reach the same level of accuracy and so additional reference data is required.
UMD GFC is economical if the cost of producing the
ForestSAT 2016 Abstracts Summary
national map minus the cost of processing the UMD
GFC is greater than acquiring and processing new reference samples. However, additional benefits may
still be gained from producing the map at national
level such as the development of a national capacity
in this field, sovereignty and increase level of details.
References:
McRoberts, R. E., Vibrans, A. C., Sannier, C, Næsset, E, Hansen, M. C, Walters, B. F., Lingner, D.V.
(2016). Comparison of methods toward multi-scale forest carbon mapping and spatial uncertainty analysis: combining national forest inventory plot data and landsat TM images. Revue
canadienne de recherche forestière, 2016, 46(7):
924-932, 10.1139/cjfr-2016-0064
McRoberts, R. E., & Walters, B. F. (2012). Statistical
inference for remote sensing-based estimates of
net deforestation. Remote Sensing of Environment, 124, 394–401
Sannier C, McRoberts RE, Fichet L-V. 2016. Suitability of Global Forest Change data to report forest
cover estimates at national level in Gabon. Remote Sens. Environ. 173: 326-338
Sannier C, McRoberts R A, Fichet LV and Massard K.
Makaga R. (2014) Using the regression estimator
with Landsat data to estimate proportion forest
cover and net proportion deforestation in Gabon.
319
Universidad Mayor
RDD+FREL-FRL and MRV
Disentangling recent patterns in litter fall of European forests
with remote sensing data across a continental scale
1
Mathias Neumann1*, Hubert Hasenauer1
Institute of Silviculture, Department of Forest and Soil Sciences, University of Natural Resources and Life Sciences,
Vienna, 1190, Austria
Key words: climate change, NPP, evapotranspiration, LAI, decomposition, nutrients, MODIS, satellite
Remote sensing data is increasingly used for
studying the Earth’s vegetation across large scales.
Forests are particular interesting due to their key
role in the global carbon cycle and ongoing climate
change. By conducting photosynthesis forests
allocate atmospheric carbon absorbed via the
stomata and water and nutrients over the roots into
their organs such as stem, branches or foliage. Litter
fall is important for relocating nutrients from the
canopy back into the soil and thus plays an important
role in the global carbon cycle by accounting for
approximately one third of Net Primary production
within forests. Although the role of litter and its
decomposition process is evident, it remains a
poorly understood process due its high stochasticity
and effort to obtain litter fall data. In this study we
describe the current spatial and temporal patterns
of litter fall data in Europe, test state-of-the-art
methods in assessing litter fall estimates and
provide a method to enhance the accuracy of litter
fall predictions with remote sensing data. For our
analysis we obtain litter fall data from the ICP Forests
Level 2 network across European forests covering
323 plots and the time period 2000 to 2012. This
data set has never been used for such an analysis.
After a quality control and a data harmonization,
we calculated a mean annual litter fall (total and
foliage) of 214 gram carbon m-2 year-1, equal 38 % of
Net Primary production (NPP) for European forests
(foliage litter fall 129 gC m-2 year-1 equal 22%). We
detected a significant positive trend between 2000
to 2012 (increase per year 3.4 gC m-2 year-1, p<0.001)
which is consistent with observations from the 20th
century. Our results suggest that there is an ongoing
increase in litter fall over the last 100 years which
could be due to climate change impacts. The results
support previous research that air temperature
320
may be a predictor for litter fall both spatially and
temporally. Combining our data with the latest
remote sensing products (Leaf Area Index, NPP and
Evapotranspiration using MODIS data) reveal, that
remote sensing data may be used to enhance spatial
predictions of litter fall, which would allow temporal
explicit assessments of litter fall in the global carbon
cycle.
References:
Holland, E. A., Post, M. W., Matthews, E., Sulzman,
J., Staufer, R., & Krankina, O. (2015). A Global
Database of Litterfall Mass and Litter Pool
Carbon and Nutrients. Data set. Available online [http://daac.ornl.gov] from Oak Ridge
National Laboratory Distributed Active Archive
Center, Oak Ridge, Tennessee, USA. http://doi.
org/10.3334/ORNLDAAC/1244
Liu, C., Westman, C. J., Berg, B., Kutsch, W., Wang,
G. Z., Man, R., & Ilvesniemi, H. (2004). Variation
in litterfall-climate relationships between
coniferous and broadleaf forests in Eurasia.
Global Ecology and Biogeography, 13(2), 105–114.
http://doi.org/10.1111/j.1466-882X.2004.00072.x
Mu, Q., Zhao, M., & Running, S. W. (2011).
Improvements to a MODIS global terrestrial
evapotranspiration algorithm. Remote Sensing
of Environment, 115(8), 1781–1800. http://doi.
org/10.1016/j.rse.2011.02.019
Neumann, M., Moreno, A., Thurnher, C., Mues, V.,
Härkönen, S., Mura, M., Bouriaud, O., Lang, M.,
Cardellini, G., Thivolle-Cazat, A., Bronisz, K.,
Merganic, J., Alberdi, I., Astrup, R., Mohren, F.,
Zhao, M., & Hasenauer, H. (2016). Creating a
Regional MODIS Satellite-Driven Net Primary
Production Dataset for European Forests.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Remote Sensing, 8(554), 1–18. http://doi.
org/10.3390/rs8070554
Pitman, R., Bastrup-Birk, A., Breda, N., & Ratio, P.
(2010). MANUAL on methods and criteria for
harmonized sampling, assessment, monitoring
and analysis of the effects of air pollution
on forests Part XIII Sampling and Analysis of
Litterfall (p. 16).
321
Universidad Mayor
RDD+FREL-FRL and MRV
Estimating the dynamics of carbon stocks in forests with
remote sensing data
Michele Dalponte, Lorenzo Frizzera, and Damiano Gianelle
Dept. of Sustainable Agro-ecosystems and Bioresources, Research and Innovation Centre, Fondazione E. Mach, Via E.
Mach 1, 38010 San Michele all’Adige (TN), Italy
Forest aboveground biomass is increasing
and consequently the carbon stored in these
ecosystems. This increase is due to many factors. In
the Alps, for example, a reduction in the agricultural
exploitation of mountain areas decreased the areas
used as grasslands or pastures, and new forest
areas appeared. Climate change issues have also
responsibilities: as an example the increase of
nitrogen depositions and/or CO2 increased forest
productivity. Remote sensing data can be really
useful in predicting the changes of aboveground
biomass (and thus carbon) over time. In this study
we will estimate forest growth, and carbon stocks
dynamics, of a forest area in the Alps.
The study site, an evergreen forest, is located in
Lavarone (Trento) in the Italian Alps. In this area 50
field plots were located and a survey of DBH, height,
and species was carried out in 2007, and it will be
repeated in summer 2016. Moreover in 20 trees
322
the annual growth was continuously monitored by
means of dendrometers. The main species are Picea
abies L. Karst., Abies alba Mill., and Fagus sylvatica L..
Remote sensing data were acquired over the study
site in 2007, 2011, and 2015. In particular airborne
laser scanning (ALS) data with more than 10 pts/m2,
and hyperspectral data were acquired in all the three
dates. Other remote sensing data (SAR) will be also
acquired from data archives.
The forest increment will be estimated at plot
level and at individual tree crowns (ITC) level. The
individual tree crowns delineation will be carried
out on the airborne laser scanning (ALS) data using
the R library itcSegment. The variation in time will
be estimated directly or by means of comparison
among estimations at different dates. Different
remote sensing data will be used both alone and
combined together.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Linking Landsat 8 and forest inventory data for local biomass
mapping in open canopy woodlands
Belachew Gizachew1a, Svein Solberg1, Erik Næsset2, Terje Gobakken2, Ole Martin Bollandsås2; Johannes Breidenbach1;
Eliakimu Zahabu3, Ernest William Mauya3
1
Norwegian Institute of Bioeconomy Research, Post Box 115, 1431, Ås, Norway
Norwegian University of Life Sciences, Department of Natural Resource Management, Post Box 5003, 1432 Ås, Norway
3
Sokoine University of Agriculture, Faculty of Forestry and Nature Conservation, P.O. Box 3009, Chuo Kikuu,
Morogoro, Tanzania.
1a
Corresponding author
2
Keywords: AGB, Mapping, Modeling, Miombo woodlands, REDD+, EVI
Abstract
REDD+ implementation requires estimates of forest biomass, an input to estimate carbon emissions.
Field inventory and or remote sensing are the expected sources of data for a functional Monitoring,
Reporting and Verification (MRV) in the context of REDD+. By linking Landsat 8 derived spectral
indices to forest inventory data, a linear mixed effects models for the above ground biomass (AGB)
was developed. The model is then applied to composite images of selected districts in Miombo
woodlands in Tanzania, to further develop a 30 m resolution AGB map. Inventory data consisted of
tree measurements from 500 plots, and the Landsat 8 data comprised two Climate Data Record (CDR)
images covering the inventory area. Among a list of spectral indices and band responses tested, the
Enhanced Vegetation Index (EVI) correlated most strongly (rho = 0.44) to the AGB. The linear model
between the AGB and Landsat 8 derived EVI had a RMSE of 25 t/ha (49%). We mapped the distribution
of AGB in 13 sub-districts and compared the AGB density to that of default values of the IPCC, and
three recent pan-tropical AGB maps. The low biomass in the miombo woodlands, and the absence
of a spectral data saturation problem suggested that Landsat 8 data are suitable sources of auxiliary
information for carbon monitoring for low-biomass, open-canopy miombo woodlands.
323
Universidad Mayor
RDD+FREL-FRL and MRV
Mapping historical canopy cover change and recovery using
Landsat time series imagery (1972-2015)
Jody C. Vogeler 1, Justin D. Braaten 2, Michael J. Falkowski 3, Robert A. Slesak 4
1
Department of Forest Resources, University of Minnesota, St. Paul, MN, USA
Department of Forest Ecosystems and Society, Oregon State University, Corvallis, OR, USA
3
Department of Ecosystem Science and Sustainability, Colorado State University, Fort Collins, CO, USA
4
Minnesota Forest Resources Council, St. Paul, MN USA
2
Keywords: Landsat time series, forest canopy cover, forest disturbance mapping, Minnesota
Monitoring historic trends in change and recovery
of forest canopy cover can provide insights into
multiple aspect of ecosystem health and functioning.
Information on disturbance and recovery events
through time may help predict the effects of
management and climatic change on forest systems
into the future as well as aiding in the modeling and
monitoring of current wildlife habitat relationships
and historic populations trends. The objectives
of this study were to: (1) create historic maps of
annual predicted forest canopy cover using Landsat
time series imagery from 1972-2015 for the state of
Minnesota; and (2) characterize historic trends in
forest change and recovery for the study area.
Our project utilizes the full value of the Landsat
archive by calibrating and incorporating imagery
dating back to 1972 to provide >40 years of
comparable annual spectral trends for the state of
Minnesota. We introduce an R package, LandsatLinkr,
which automates the pre-processing steps
necessary for annual cloud-free atmospherically
and geographically corrected tasseled cap indices
calibrated between Landsat platforms including
the early MSS sensors. We used the LandTrendr
algorithm to smooth and segment the tasseled
cap trends. These smoothed annual tasseled cap
324
composite images were then correlated with
Minnesota Forest Inventory and Analysis plots and
4-band NAIP imagery to model forest canopy cover.
The best model was then applied to the entire stack
of time series Landsat composites to map annual
predicted forest canopy cover for over 4 decades.
In addition to the creation and validation of these
annual canopy cover maps, the cover model was
also incorporated into the change labeling portion
of the LandTrendr algorithm to better characterize
trends in forest change and recovery rates.
The products from this project include statewide
annual predicted forest canopy cover maps for the
past 4 decades in addition to maps of classified forest
change (e.g. fast, slow, high intensity, low intensity)
and rates of recovery from the various magnitudes of
disturbance. These mapping products are currently
being incorporated into a range of forest ecosystem
research and management applications including:
modeling wildlife habitat and population trends;
effects of various harvest methods and patterns
on post-harvest forest characteristics and recovery
rates; and creating maps classifying the agent of
disturbance (e.g. harvest, fire, insect, windfall,
urbanization, etc.) for the state of Minnesota.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Resumen Ejecutivo de Nivel de Referencia de Emisiones
Forestales / Nivel de Referencia Forestal Subnacional de Chile
Javier Cano Martin , Angelo Sartori Ruilova2
Unidad de Cambio Climático y Servicios Ambientales (UCCSA), Gerencia de Desarrollo y Fomento Forestal (GEDEFF),
Corporación Nacional Forestal (CONAF).
Profesional Responsable Contabilidad de Carbono y Sistema MRV, UCCSA, GEDEFF. javier.cano@conaf.cl
2
Jefe UCCSA, GEDEFF, CONAF. angelo.sartori@conaf.cl
Justificación
Cambio Climático
Según la Organización Meteorológica Mundial
la concentración atmosférica media mundial de
CO2 ha aumentado en un 41% en los últimos 50
años, superando la barrera de las 400 ppm de
concentración de CO2 lo que representa la mayor
concentración en los últimos 800.000 años.
El Panel Internacional de Expertos de Cambio
Climático (IPCC) ha identificado variaciones
en indicadores meteorológicos derivados del
aumento de concentración de Gases de Efecto
Invernadero (GEI) durante el Siglo XX, destacando
el aumento general de temperatura, días calurosos,
precipitaciones fuertes y frecuencia e intensidad
de las sequías y una disminución de los días de
frío y heladas, estos indicadores en las variaciones
climáticas han sido estudiados ampliamente,
existiendo un consenso científico general.
En Chile estas alteraciones han sido percibidas
durante los últimos años, provocando un aumento de
concientización en la sociedad ante las potenciales
amenazadas producto de variaciones climáticas
derivadas de las actividades humanas que alteran
la composición de la atmósfera global y conllevan al
denominado cambio climático.
El Rol de los Bosques
La principal causa del aumento de CO2 (principal Gas
de Efecto Invernadero) en la atmósfera es el uso de
combustibles fósiles, mientras que la deforestación
y degradación de los bosques representa, según
IPCC, el 17.3% de las emisiones GEI a nivel mundial.
Para el caso de América Latina, la representatividad
de las emisiones de CO2 producto de la deforestación
y degradación de los bosques supera el 40% (CEPAL,
basado en datos del World Resources Institute, 2010)
Este porcentaje adquiere mayor relevancia si se une
a la capacidad de los bosques para capturar y fijar
carbono atmosférico su relevancia como medio de
suministro de servicios ecosistémicos, provocando
que muchos de los esfuerzos realizados a nivel global
se concentren en identificar y poner en práctica
estrategias que frenen la deforestación, degradación
de los bosques, fomento de la conservación,
manejo forestal sustentable e incremento de masas
forestales.
REDD+
En el marco internacional, durante los últimos 20
años la estructura de estas políticas ha establecido
organismos e institucionalidad globales que marcan
las políticas a seguir. De especial interés para Chile
son las directrices establecidas en el Marco de
REDD+ de Varsovia, que determinan que cualquier
país que quiera acceder al pago por resultados
basados en desempeño, en términos de reducción
de emisiones y/o incremento de absorciones de
carbono forestal, deberá desarrollar cuatro ejes
centrales que se corresponden con: 1) contar con una
estrategia país que defina los enfoques de políticas
para promover incentivos positivos, 2) elaborar un
nivel de referencia de carbono forestal, 3) diseñar e
implementar un sistema de monitoreo de emisiones
y capturas de carbono y 4) velar por el cumplimiento
de consideraciones sociales y ambientales, conocidas
como salvaguardas en el contexto global.
ENCCRV
En este contexto, con el objetivo de apoyar la
recuperación y protección del bosque nativo y
325
Universidad Mayor
RDD+FREL-FRL and MRV
formaciones xerofíticas, así como potenciar el
establecimiento de formaciones vegetacionales
en suelos factibles de ser plantados como medidas
de mitigación y adaptación a los efectos del
cambio climático y lucha contra la desertificación,
CONAF, está desarrollando e implementando la
Estrategia Nacional de Cambio Climático y Recursos
Vegetacionales (ENCCRV) que pretende alcanzar
dicho objetivo mediante el diseño e implementación
de un mecanismo estatal que facilite el acceso
de las comunidades y propietarios de bosques,
formaciones xerofíticas y suelos factibles de ser
plantados, a los beneficios asociados a los servicios
ambientales de estos ecosistemas. Paralelamente,
la ENCCRV, pretende satisfacer los compromisos
internacionales que ha asumido Chile en materia
de cambio climático asociado a los bosques y
lucha contra la desertificación, donde destaca el
compromiso del Gobierno ante la CMNUCC de
manejar sustentablemente y recuperar 100.000
hectáreas de bosque, así como forestar otras 100.000
hectáreas, principalmente con especies nativas, en
el período 2016/2030.
Nivel de referencia
Los sistemas forestales/vegetacionales chilenos
Chile, debido a sus características geográficas (suelo,
clima, latitud, distribución de la población, entre
otras) posee una fuerte diversidad de ecosistemas
forestales, con presencia de pequeños bosques de
altura en el altiplano y formaciones xerofíticas en
la zona norte; bosques esclerófilos en zona central;
bosques dominados por especies caducifolias del
género Nothofagus en la zona de clima húmedo
templado, donde se insertan formaciones dominadas
por coníferas milenarias como la Araucaria y el Alerce,
para transitar a bosques compuestos por especies
perennes, denominado siempreverde, en el área de
la Selva Valdiviana y grandes extensiones de bosques
de lenga en las regiones más australes del país.
Cada uno de estos bosques y formaciones xerofíticas
representan un valioso aporte para la mitigación, en
base a la reducción de emisiones e incremento de
capturas de CO2, y la adaptación a los efectos del
cambio climático como medio de provisión de agua y
refugio, conservación de la biodiversidad, retención
de suelo, limitante del avance de la desertificación,
etc.
326
Significado y relevancia
El Nivel de Referencia de Emisiones Forestales
/ Nivel de Referencia Forestal (NREF/NRF),
tiene como objetivo caracterizar las emisiones
históricas de Gases de Efecto Invernadero (GEI)
por Deforestación y degradación de los bosques,
así como las absorciones producto del aumento
de existencias de reservas de carbono forestal, la
conservación y el manejo sustentable de los bosques
y las proyecta hacia el futuro, con la intención
de medir el desempeño del enfoque de políticas
asociadas a REDD+.
Durante el mes de enero de 2016, Chile consignó
el NREF/NRF de carácter subnacional, ante la
Convención Marco de las Naciones Unidas sobre
Cambio Climático (CMNUCC), que caracteriza las
emisiones históricas de GEI por Deforestación
y Degradación de los bosques, así como las
absorciones producto del aumento de existencias
de reservas de carbono forestal, la conservación y
el manejo sustentable de los bosques y las proyecta
hacia el futuro, para las regiones del Maule a Los
Lagos entre los años 1997 y 2013, con la intención
de medir el desempeño del enfoque de políticas
asociadas a REDD+.
El NREF/NRF fue elaborado siguiendo los criterios
determinados por el Panel Intergubernamental de
Expertos en Cambio Climático (IPCC) y mantiene
congruencia con el Inventario Nacional de Gases de
Efecto Invernadero (INGEI) utilizando las mismas
fuentes de información, principalmente el Catastro
de Bosque Nativo y sus actualizaciones, el Inventario
Forestal Continuo y Estadísticas de Incendios
Forestales.
En la actualidad el documento presentado está siendo
sometido a revisión de expertos internacionales
independientes que propondrán mejoras y ajustes
para optimizar los métodos e identificar las brechas
de información existentes. Durante este período
de revisión, CONAF, en conjunto con el equipo de
apoyo técnico, tomará las acciones pertinentes para
satisfacer las demandas de la CMNUCC CMNUCC
entre las que ya se han identificado las de mayor
relevancia, como son: 1) la creación de un sistema de
registro espacial para la delimitación de áreas bajo
Manejo Forestal que permitan estimar el flujo de
Universidad Mayor
carbono y estimar el Nivel de Referencia de Manejo
Sustentable, y 2) levantar información espacial de
uso de la tierra, complementaria al Catastro, para
homogeneizar la temporalidad de los datos y el
periodo de referencia.
De forma paralela se trabajará para extender el
Nivel de Referencia de escala subnacional a escala
nacional, con el objetivo de presentar el NREF/NRF
que registre las emisiones y absorciones de las cinco
actividades REDD+ a nivel nacional.
Equipo Técnico
El NREF/NRF Subnacional de Chile ha sido elaborado
por CONAF, a través de un trabajo conjunto liderado
por la Unidad de Cambio Climático y Servicios
Ambientales (UCCSA) y el Departamento de
Monitoreo de Ecosistemas Forestales, con el apoyo
técnico del Banco Mundial, la Consultora Winrock
Internacional, el Instituto Forestal (INFOR), la
Universidad Austral de Chile (UACH) y las agencias de
Naciones Unidas, FAO, PNUMA y PNUD, adscritas al
Programa de REDD de Naciones Unidas (UN-REDD).
Área de contabilidad
CONAF, ha decidido focalizar los esfuerzos para
la generación del Nivel de Referencia subnacional
en la zona Centro-Sur de Chile, desde la Región del
Maule hasta la Región de Los Lagos, área que cuenta
con una alta concentración y la mayor diversidad
de bosques del país, una fuerte presión antrópica
y con gran potencial para la reducción/absorción
de emisiones de GEI relacionadas con los bosques
y capacidad de producir beneficios ambientales no
relacionados con el carbono.
Conceptos clave
Bajo el contexto de REDD+, que rige el NREF/
NRF, Chile definió como bosque aquellas tierras
establecidas como Bosque Nativo en la legislación
vigente y determinó los conceptos nacionales
para las cinco actividades REDD+: Deforestación,
Degradación, Aumento de existencias, Conservación
Forestal y Manejo Sustentable.
● Para el cálculo de emisiones por Deforestación fue
estimada la superficie de bosque nativo convertida
a otros usos.
● En el caso de la Degradación Forestal se
consideraron las emisiones por pérdida de biomasa
en los bosques que permanecen como tal, las
emisiones de CH4 y N2O producto de la combustión
en incendios forestales, así como las emisiones
por sustitución de bosque nativo por plantaciones
ForestSAT 2016 Abstracts Summary
monoespecíficas de carácter industriales.
● El Aumento de existencias registra aquellas
absorciones producto de la forestación y la
restitución de bosque nativo, así como las
absorciones producidas por los bosques que
permanecen bosques durante el periodo analizado.
● Para determinar el Nivel de Referencia de
Conservación Forestal, fue analizado el flujo de
carbono en los bosques que permanecen bosques
en áreas bajo conservación formal.
● Debido a la falta de información oficial del país
que permita localizar y delimitar espacialmente
las áreas sujetas a Manejo Sustentable, esta
actividad no fue integrada en el NREF/NRF, si bien
las emisiones y absorciones producto de manejo
sí son registradas dentro de las actividades de
Degradación y Aumento de Existencias.
Métodos
El NREF/NRF fue elaborado siguiendo los criterios
determinados por el Panel Intergubernamental de
Expertos en Cambio Climático (IPCC) y mantiene
congruencia con el Inventario Nacional de Gases de
Efecto Invernadero (INGEI) utilizando las mismas
fuentes de información, principalmente el Catastro
de Bosque Nativo y sus actualizaciones, el Inventario
Forestal Continuo y Estadísticas de Incendios
Forestales.
La disponibilidad de información, hace que no
se pueda contar con un período de referencia
estandarizado para las cinco regiones, pero se
estandarizaron los datos para determinar las
emisiones y absorciones de CO2e entre los años
1997 y 2013.
Los métodos aplicados siguen las directrices IPCC
de mayor precisión y exactitud, aplicando Enfoques
de grado 3 (espacialmente expliciticos), para el
cálculo de datos de actividad y factores de emisión
de Nivel 2 y 3, (creados en base a datos, inventarios
y modelos, desarrollados en el país a escala
nacional o subnacional). Se aplicaron metodologías
diferenciadas para estimar emisiones y absorciones
en áreas de bosques que permanecen como tal y
áreas que registran cambios de uso de la tierra.
En el caso de los bosques que permanecen como
tal las emisiones por degradación, así como las
absorciones producto del aumento de biomasa,
fueron calculadas mediante una metodología que,
utilizando datos del Inventario Forestal Continuo
y análisis de imágenes satelitales, calcula el área
basal y número de árboles en una retícula continua
327
Universidad Mayor
RDD+FREL-FRL and MRV
de celdas de una hectárea, permitiendo identificar la
situación de degradación o aumento de existencias y
el contenido exacto de CO2e emitido a la atmósfera
o absorbido por el bosque durante el periodo.
Para la estimación de emisiones y absorciones
que implican cambio de uso de la tierra, como la
sustitución, restitución o forestación se analizaron las
matrices de cambio del Catastro y se establecieron
valores derivados del Inventario Forestales Continuo,
de la misma forma que se aplica en el INGEI.
Para determinar las emisiones de Gases de Efecto
Invernadero producto de la combustión en incendios
forestales se aplicaron factores estándar.
Resultados generales
El NREF/NRF subnacional de Chile representa un
flujo de carbono correspondiente a absorciones de
carbono 1,5 millones de Toneladas de CO2e anuales.
Este se divide en cuatro principales actividades:
● Deforestación; con 3,5 millones de Ton de CO2e
anuales emitidas, para una superficie superior a las
6.000 ha anuales
● Degradación Forestal; con 9,2 millones de Ton.
CO2e anuales emitidas, para una superficie superior
a las 500.000 hectáreas entre 2001 y 2010.
● Aumento de existencias; con 10,0 millones de
Ton.CO2e anuales absorbidas, para una superficie
superior a las 700.000 hectáreas entre 2001 y 2010
● Conservación Forestal; con 2,4 millones de Ton.
CO2e anuales absorbidas, para una superficie
cercana a 1.450.000 hectáreas de bosques
catalogados como áreas de conservación.
Es importante mencionar que el NREF/NRF no
contabiliza las absorciones y emisiones que se
producen en bosques nativos que se mantienen en
un umbral de resiliencia o recuperación natural así
como los flujos que se producen en las plantaciones
forestales. Estos flujos son recopilados, sin embargo,
en el Inventario de Gases de Efecto Invernadero de
Chile.
Cronología de ampliación
Durante el año 2016 el Nivel de Referencia será
ampliado a las regiones de Valparaíso, Metropolitana
y O’Higgins, donde ya se cuenta con una serie de
trabajos que han permitido tener avances puntuales.
Se prevé que durante 2017 el Nivel de Referencia se
extienda a nivel Nacional.
Relevancia del NR
Contar con un Nivel de Referencia ha sida un
punto clave para el proceso de elaboración de la
ENCCRV, ya que si bien, los principales causales de la
deforestación y degradación forestal a nivel nacional
e incluso a escala regional, como los incendios y el uso
insustentable de la biomasa (principalmente la leña)
son conocidos por todos los actores involucrados,
hasta la fecha no ha existido una cuantificación
específica, así como una identificación detallada de
áreas con mayor nivel de afección.
En el marco de la ENCCRV han sido planteadas una
serie de medidas de acción que tienen por objetivo
combatir los principales causales de Deforestación
y Degradación forestal, así como promover la
Conservación, el Manejo Forestal y el Aumento
de carbono en los bosques, la identificación de las
áreas prioritarias para cada tipo de intervención, así
como el control y monitoreo de los resultados de la
implementación de estas medidas serán analizados
en base al Nivel de Referencia.
La adecuada ejecución de estas medidas de acción,
deberán ser monitoreadas, reportadas y verificadas
ante entes internacionales, lo que permitirá a los
ejecutores recibir pagos por resultados, incrementando
la plusvalía de los bosques y mejorando las condiciones
de vidas de aquellos que mantienen una relación
directa con el bosque nativo chileno.
NREF/NRF
Actividad REDD+
Maule
Biobío
La Araucanía
Los Ríos
Los Lagos
Total
Deforestación
Degradación
Conservación
Aumentos
84.982
396.645
1.059.067
644.696
1.267.494
3.452.884
608.976
1.209.890
1.907.344
1.373.080
4.050.103
9.149.392
-14.780
-72.359
-334.741
-710.081
-1.298.478
-2.430.439
-1.182.162
-1.282.143
-1.517.894
-2.022.041
-4.007.772
-10.012.012
Tabla 1. NREF/NRF subnacional de Chile
328
Universidad Mayor
ForestSAT 2016 Abstracts Summary
POSTER
Synergistic use of sar and optical datasets for forest
biomass retrieval and characterization of forests in
temperate zone – a national case study Poland
Hoscilo Agata1, Ziolkowski Dariusz1, Lewandowska Aneta1, Sterenczak Krzysztof2, Bochenek Zbigniew1, Bartold Maciej3
Remote Sensing Centre, Institute of Geodesy and Cartography, Warsaw, Poland
2
Forest Research Institute, Sekocin, Poland
3
University of Warsaw, Faculty of Geography and Regional Studies, Warsaw, Poland
1
Hoscilo, A.: E-mail: agata.hoscilo@igik.edu.pl, Tel: +48 223291976
Keywords: biomass, SAR, carbon, temperate, forest parameters
Abstract
Assessment of forest above-ground woody biomass is essential for national and regional forest
carbon stocks and carbon stock changes estimation and reporting. The use of remotely sensed data
supports countries in advancing approaches to forest monitoring and management. It helps to obtain
accurate data on the status of forest resources and forest carbon stocks and support implementation
of national and international policies. The aim of this research is to develop methodology for the forest
above-ground woody biomass retrieval at the temperate forest, based on remotely sensed data. The
biomass assessment was conducted in the framework of the European Space Agency (ESA) funded
GlobBiomass project, which aims to better characterize and to reduce uncertainties of above ground
biomass estimates by developing an innovative synergistic mapping approach in five regional sites
(Sweden, Poland, Borneo, Mexico, South Africa) for the epochs 2005, 2010 and 2015 and one global
map for the year 2010. The authors present the approach for the biomass retrieval at the national
level over Poland – temperate forest. A synergy of radar ALOS PALSAR (L-band) and optical Landsat
missions data have been used to derive forest above-ground woody biomass at the national level. The
backscattering at HH and HV polarization, texture and ratios have been tested. The National Inventory
of Forest Condition (inventory plots) has been used as the reference data for the biomass retrieval. The
growing stock volume was converted into woody biomass using the IPCC approach based on biomass
expansion factors (BEFs) and wood density following IPCC guidelines. The method used for biomass
estimation in Poland was based on a machine-learning Random Forest regression. The Random Forest
models were calibrated separately for coniferous and deciduous forest using a set of training plots
located over the entire forested area except steep slopes in the mountains. The overall accuracy of
the biomass retrieval is around 50 tons per hectares; the accuracy of coniferous is better than for
broadleaf forest. The information on the spatial distribution of forest biomass was then related to
the forest parameters (tree species, type of forest site, forest composition) described by vegetation
spectral indices derived from high-resolution Landsat and SPOT5 images. The aim of this analysis
was to examine the relationship between forest biomass and various features characterizing forest
canopies within the selected area of interest. Remotely sensed forest parameters were determined
within WICLAP project, conducted within Polish-Norwegian Research Programme, financed by the
National Centre for Research and Development.
329
Universidad Mayor
RDD+FREL-FRL and MRV
Using leaf-on and leaf-off airborne LiDAR to model vegetation
structure and above-ground carbon storage in the critical zone
Kristen Brubaker
Hobart and William Smith Colleges, Geneva, NY
Keywords: understory biomass, critical zone, forest structure, lidar
Understanding patterns of above-ground carbon
storage across forest types is increasingly important
as managers adapt to the threats of climate change.
Although airborne lidar has been used extensively
to model above-ground carbon storage, it has not
been extensively used to understand the difference
between tree and shrub carbon storage in closed
canopy forests. We compared the fine-scale above
ground carbon storage in two watersheds; one
watershed was underlain by sandstone bedrock and
the other by shale. We measured tree and shrub
biomass across three topographic positions for both
watersheds, and calculated the carbon stored in
each. We then used leaf-on and leaf-off airborne lidar
datasets to construct a model for carbon storage
for each component across the watershed, using
a combination of terrain and point cloud metrics
330
from lidar. Since overstory vegetation can obscure
the understory vegetation from LiDAR, a correction
factor was applied using percent overstory canopy
cover. Using RandomForest, we modeled shrub and
tree forest carbon across both watersheds, using
a combination of leaf-on and leaf-off LiDAR point
cloud metrics and topographic metrics. We found
that there is an inverse relationship between tree
carbon storage and shrub carbon storage across
sites, and that LiDAR can be used to model these
relationships across a broader, watershed scale. We
also found differences in the tree carbon and shrub
carbon ratios between bedrock types. Watersheds
underlain with sandstone had a higher proportion
of their carbon stored in shrubs, while watersheds
underlain with shale had a higher proportion of their
carbon stored in trees.
Universidad Mayor
ForestSAT 2016 Abstracts Summary
Using satellite data to estimate gas emissions into the
atmosphere by burning biomass in Mexico
Maria Isabel Cruz Lopez
National Commission for Knowledge and Use of Biodiversity (CONABIO),
Liga Periferico –Insurgentes Sur 4903, Col. Parques del Pedregal
C.P 14010 Tlalpan, Mexico D.F., icruz@conabio.gob.mx
Keywords: Emissions, Forest fire, satellite data
Mexico is a biodiverse country as result of
geographical conditions. One of the most recurrent
threats to biodiversity are forest fires, which among
its consequences are emissions of greenhouse
gases (GHGs), which affect the nature, health and
human activities, and they have influence in the
global warming. These effects are expressed in
different spatial levels: local, regional and global,
so it becomes a relevant issue to national and
international context.
using satellite imagery, but the uncertainty is still
high.
Internationally, the issue of fires emissions has been
analyzed on different ways and techniques; one of
the most widely models used to estimate emissions
require calculating four main parameters: burned
area, amount of existing biomass, burning efficiency
(biomass consumed by fire) and emission factor.
Different authors have applied the model using
statistical data and satellite data. One of the most
difficult parameters to estimate is the efficiency of
burning, some authors have proposed alternatives
In order to make predictions of this variable in
dynamic form, it uses statistical learning methods,
as decision trees, and satellite products to find
the spatial and temporal relationships between
environmental features (18 variables of biomass,
vegetation, topography) and the burning efficiency
(field data, provided by Dr. German Flores). The
preliminary results show some patters to select
the main variables and introduce others, as fuel
moisture, and increase the accuracy of prediction.
Therefore the main objective of the research is
to develop a method for calculating the burning
efficiency with remote sensing data in three
representative sites of forest eco-regions of Mexico,
as input for the estimation of gas emissions from
the burning of biomass and know their spatial
distribution.
331
RDD+FREL-FRL and MRV
332
Universidad Mayor