Australian Google Earth Engine Burnt Area Map

A Rapid, National Approach to Fire Severity Mapping

Department of Agriculture, Water and the Environment

P a g e 1 | 25
© Commonwealth of Australia 2020

Ownership of intellectual property rights

Unless otherwise noted, copyright (and any other intellectual property rights) in this publication is owned by the
Commonwealth of Australia (referred to as the Commonwealth).

Creative Commons licence

All material in this publication is licensed under a Creative Commons Attribution 4.0 International Licence except
content supplied by third parties, logos and the Commonwealth Coat of Arms.

Inquiries about the licence and any use of this document should be emailed to copyright@[Link].

Cataloguing data

This publication (and any material sourced from it) should be attributed as: Department of Agriculture, Water and
the Environment, Canberra, July 2020, CC BY 4.0.

Cover photo: Drone footage, G. Pickford, NSW ParkAir

Contents photo: Sentinel 2 false colour, A. Roff, NSW DPIE, ESA

Department of Agriculture, Water and the Environment

GPO Box 858 Canberra ACT 2601
Telephone 1800 900 090
Web [Link]

The Australian Government acting through the Department of Agriculture, Water and the Environment has exercised
due care and skill in preparing and compiling the information and data in this publication. Notwithstanding, the
Department of Agriculture, Water and the Environment, its employees and advisers disclaim all liability, including
liability for negligence and for any loss, damage, injury, expense or cost incurred by any person as a result of
accessing, using or relying on any of the information or data in this publication to the maximum extent permitted by
law.

Acknowledgements

AUS GEEBAM was developed by the Remote Sensing and Landscape Science Branch, Science Economics and Insights
Division, New South Wales Department of Planning, Industry and Environment.

P a g e 2 | 25
Contents
Figures................................................................................................................................................................................... 4
Tables .................................................................................................................................................................................... 4
Executive Summary ......................................................................................................................................................... 5
Background ......................................................................................................................................................................... 6
Aim.......................................................................................................................................................................................... 7
Methods ................................................................................................................................................................................ 7
Overview of workflow .......................................................................................................................................... 7
Data and Inputs ............................................................................................................................................................ 8
National Indicative Aggregated Fire Extent Dataset (NIAFED). .......................................................... 8
National Vegetation Information System (NVIS) ....................................................................................... 9
Interim Biogeographic Regionalisation for Australia (IBRA) ............................................................ 10
Sentinel-2 (S2) Surface Reflectance (L2) ................................................................................................... 11
Creating time series mosaics of median surface reflectance ............................................................. 11
Vegetation Indices ............................................................................................................................................... 12
Relativised Normalized Burnt Ratio (RNBR) ........................................................................................... 12
Reclassifying continuous RNBR values as ordinal quantiles ............................................................. 13
Calibrating RNBR quantiles using visual interpretation of remote sensing ................................ 13
Evaluating AUS GEEBAM .................................................................................................................................. 13

P a g e 3 | 25
Results ................................................................................................................................................................................ 16
Discussion ......................................................................................................................................................................... 21
Appendix ........................................................................................................................................................................... 22
New South Wales Fire Extent and Severity Map ..................................................................................... 22
Victorian Fire severity map ............................................................................................................................. 23
References ........................................................................................................................................................................ 24

Figures
Figure 1. A comparison of workflows of fire severity maps from NSW/VIC and GEEBAM. .............. 8
Figure 2. National Indicative Aggregated Fire Extent Dataset (NIAFED) for July 1st, 2019 to
February 24th, 2020. ...................................................................................................................................................... 9
Figure 3. National Vegetation Information System (NVIS) symbolised by broad vegetation types.
............................................................................................................................................................................................... 10
Figure 4. Interim Biogeographic Regionalisation for Australia (IBRA) converted to a WGS84 40m
raster................................................................................................................................................................................... 11
Figure 5. Sentinel-2 Surface Reflectance WGS84 40m mosaic. .................................................................. 16
Figure 6. Australian Google Earth Engine Burnt Area Map 4.0 .................................................................. 18
Figure 7. South Eastern Australia – False Colour Mosaic - GEEBAM........................................................ 19
Figure 8. South Eastern Australia – Victorian fire severity map – DELWP ........................................... 19
Figure 9. South Eastern Australia – AUS GEEBAM fire severity classes – AUS GEEBAM ................. 20
Figure 10. South Australia – Kangaroo Island - False Colour Mosaic - GEEBAM ................................ 20
Figure 11. South Australia – Kangaroo Island - GEEBAM fire severity classes – AUS GEEBAM ... 21

Tables
Table 1. Date ranges of time series mosaics calculated as the median pixel value. ........................... 12
Table 2. AUS GEEBAM classes with descriptions. ............................................................................................ 13
Table 3. Fire severity mapping from Victoria (VIC), New South Wales (NSW) and South Australia
(SA) ...................................................................................................................................................................................... 14
Table 4. Fire severity map evaluation class equivalent for four classes ................................................. 15
Table 5. Fire severity map evaluation class equivalent for two classes ................................................. 15
Table 6. Reclassifying continuous RNBR Values as ordinal quantiles. .................................................... 17
Table 7. NSW fire severity map – 4 class confusion matrix ......................................................................... 22
Table 8. NSW fire severity map – 2 class confusion matrix ......................................................................... 22
Table 9. VIC fire severity map – 4 class confusion matrix ............................................................................ 23
Table 10. VIC fire severity map – 2 class confusion matrix .......................................................................... 23
Table 11. SA fire severity map – 4 class confusion matrix ........................................................................... 24
Table 12. SA fire severity map – 2 class confusion matrix ........................................................................... 24

P a g e 4 | 25
Executive Summary
The Commonwealth Department of Agriculture, Water and the Environment (DAWE)
collaborated with the NSW Department of Planning, Industry and Environment (DPIE) to
develop a rapid, national approach to mapping fire severity.
The results will be used to quantify the potential impacts of the 2019/20 bushfires on wildlife,
plants and ecological communities and to identify appropriate response and recovery actions.
The DPIE Remote Sensing and Landscape Science Branch has developed the Australian Google
Earth Engine Burnt Area Map (AUS GEEBAM). The dataset uses Sentinel-2 satellite imagery
from before and after the fires of the 2019/20 bushfire season. AUS GEEBAM presents four fire
severity classes for the areas represented in the National Indicative Aggregated Fire Extent
Dataset (NIAFED) from the 25th of February 2020.
By leveraging the processing power of cloud computing, AUS GEEBAM can be produced rapidly.
This allows for fortnightly updates of fire severity during a fire season if required. This
document details the methodology used and provides a brief comparison to existing fire
severity maps from states and territories.
A set of random points were used to examine the relationship between AUS GEEBAM and fire
severity mapping from Victoria, New South Wales and South Australia. The points were
stratified by each state’s maps. When comparing four classes (Unburnt, Low and Moderate,
High, and Very High severity) the overall accuracy of AUS GEEBAM was between 48% and 82%.
When comparing two simple classes (Moderate and High severity) the overall accuracy rose to
between 72% and 92%. It is important to note that this is not a traditional validation as there is
no point of truth. Each map used in the comparisons will contain errors.
AUS GEEBAM was calibrated using visual interpretation of remote sensing data and no field
data was used. Currently, field data and detailed air photo analysis is collected independently in
each state and territory. It would be a powerful resource if it were collated at a Commonwealth
level and made available to researchers.
The intention of AUS GEEBAM is not to replace fire severity maps from each state and territory
but to provide a nationally consistent fire severity map. It is likely that individual state and
territory fire severity maps will more accurately represent variation in fire severity as they
benefit from local expertise and local calibration data. However, the state and territory datasets
do not fulfil the need for a national fire severity dataset built with a common methodology.

P a g e 5 | 25
Background
In late 2019 and early 2020, Australia experienced concurrent mega-fires throughout New
South Wales, Queensland, Victoria and South Australia and its most devastating fire season on
record (Nolan et al. 2020).
Fire can impact natural ecosystems and ecological communities by releasing sequestered
carbon (Abbott et al. 2016), altering nutrient cycles and water quality (Ice et al. 2004, Bladon et
al. 2014), altering plant distribution (Barlow and Peres 2008), and causing wildlife injury and
mortality (Esque et al. 2003).
When assessing the impact of a fire within an ecosystem it is of great importance to understand
how and where the fire has burnt, as often fires burn with varied intensities, leaving a landscape
burnt at a range of severity levels.
Here, the term fire intensity is used to describe fire behaviour and the energy that is released
from the fire, while fire severity refers to the effects the fire had on the ecosystem, vegetation or
loss in biodiversity.
The Australian Government’s Threatened Species Recovery Hub set a goal to provide an
immediate ecological response to Australia’s 2019/2020 wildfire events. This calls for
conservation practitioners to provide a “rapid assessment of biodiversity loss” and to “locate
and protect key refuge areas” (Dickman et al. 2020).
To mitigate against bushfire events and assist with the species recovery effort it is important to
understand the spatial extent of fires across species distributions and landscapes as well as to
rapidly identify refugia habitat for targeted post-bushfire conservation actions.
Potential bushfire refuges often persist in fire affected landscapes. Keppel et al. (2012) suggest
terrestrial biodiversity can potentially persist in this refugia habitat, which can remain intact in
an otherwise largely fire affected landscape (Collins et al. 2019), often due to factors of
topography, fuel loads, and fire intensity (Bradstock et al. 2010, Collins et al. 2012).
In order to meet these goals, a rapid spatial assessment tool is required to understand the scale
and severity of bushfire.
Techniques for fire severity mapping have traditionally relied on the difference between
vegetation indices from before and after fires (Keely, 2009). More recent advances, such as
relativised vegetation indices, use imagery from before the fire to help calibrate between fires
(Parks et al. 2014).
Victoria has developed a fire severity mapping method using machine learning classification
(random forests) of vegetation indices from before and after fire using Sentinel-2 satellite
imagery (Collins et al. 2018). The results highlight the benefit of using multiple vegetation
indices in a classification. It is based on a large collection of point data from previous fires and
visual interpretation of high spatial resolution remote sensing data. However, it has not been
implemented more widely.
Gibson et al. (2020) demonstrated a similar approach to mapping fire extent and severity
through a machine learning framework based on Sentinel-2 satellite imagery. It shared the
advantage of being calibrated with fires from previous seasons and included vegetation cover as
a component. However, while this is being implemented operationally with state agencies in
NSW, to date it remains incomplete for all fires in NSW.

P a g e 6 | 25
Severity classes derived from vegetation indices from before and after fires (Parks et al. 2014)
are widely used to map the severity of large bushfires (Collins et al. 2018). Using a single
vegetation index and setting thresholds has the advantage of being rapid and simple to
implement widely. However, there are shortcomings that include poor delineation of low fire
severity classes, and unsatisfactory performance when classification thresholds are applied to
new landscapes.

Aim
The aim of this project was to create a consistent, rapid, national approach to fire severity
mapping to inform bushfire response.
The requirements were to:
a) Create a national fire severity map with a consistent methodology and classification
system,
b) Create an approach that can be reproduced rapidly during and after a fire season,
c) Establish a relationship between a national fire severity map and the products in use at
state and territory agencies.
To meet these requirements, we introduce the Australian Google Earth Engine Burnt Area Map
(AUS GEEBAM).

Methods
Overview of workflow
AUS GEEBAM relies on Sentinel-2 satellite imagery and vegetation indices from before and after
fires. It differs to existing approaches in several ways. It harnesses the power of cloud
computing (Google Earth Engine) to create national time series mosaics without clouds,
shadows or artefacts.
The number of severity classes has been reduced by combining low and moderate severity fires.
AUS GEEBAM classes are calibrated systematically for each bioregion using visual interpretation
of Sentinel 2 false colour composites. Allocating fire severity classes is carried out by a single
interpreter. AUS GEEBAM is a scalable approach that can cover large areas rapidly, while still
providing reasonable accuracy at higher severity classes.
This method does introduce bias. However, any product that relies on the visual interpretation
of remote sensing for calibration is also subject to bias (Fisher et al. 2018; Collins et al. 2018;
Gibson et al. 2020). To mitigate observer bias, bioregions and vegetation mapping were used to
encourage a systematic approach to visual interpretation across the landscape (Figure 1). This
allows the observer to focus on, for example, rainforest and alpine heathlands and calibrate
them independently from tall forest.

P a g e 7 | 25
 NSW, VIC AUS GEEBAM

Image processing - Image processing -

SDC, Google Earth Engine Google Earth Engine

Calibration - Classify severity classes -

Visual interpretation of remote sensing Fixed relativised dNBR classes

Classify severity classes - Calibration -

Random forest Visual interpretation of remote sensing

Validation - Validation -
based on ground data when available based on ground data when available

Figure 1. A comparison of workflows of fire severity maps from NSW/VIC and GEEBAM.

In the aftermath of the 2019/2020 fires there is a wide range of field data and very high-
resolution aerial photography being acquired. These data will provide an opportunity to
conduct a more rigorous evaluation of AUS GEEBAM. When enough data is available it will be
possible to implement AUS GEEBAM classes based on multiple indices.

Data and Inputs

National Indicative Aggregated Fire Extent Dataset (NIAFED).
The extent of bushfires for AUS GEEBAM in 2019/2020 relies on the National Indicative
Aggregated Fire Extent Dataset (NIAFED).
NIAFED takes the national Emergency Management Spatial Information Network Australia
(EMSINA) data service, which is the official fire extent currently used by the Commonwealth and
adds supplementary data from other sources to form a cumulative national view of fire extent.
The EMSINA data service shows current active fire incidents while NIAFED shows the total fire
extent from 1st of July 2019 to the 24th of February 2020 (Figure 2). This dataset is released on
behalf of the Commonwealth Government and endorsed by the National Burnt Area Dataset
Working Group, convened by the National Bushfire Recovery Agency.

P a g e 8 | 25
Figure 2. National Indicative Aggregated Fire Extent Dataset (NIAFED) for July 1st, 2019 to February 24th,
2020.

National Vegetation Information System (NVIS)

The National Vegetation Information System (NVIS) was developed to assist in managing a
range of ecosystem services and practices such as biodiversity conservation, salinity control,
improving water quality and fuel-load management.
The NVIS framework was developed to enable the compilation of a nationally consistent
vegetation dataset from data collected by states and territories. NVIS describes the structural
and floristic patterns of groups of plants in the landscape. Collectively, the different levels in the
classification provide a description of vegetation that can be directly related to precise spatial
areas as a vegetation map.
For AUS GEEBAM the latest available vegetation maps from each state were compiled, assigned
NVIS vegetation types, and converted to a WGS84 40m raster (Figure 3).

P a g e 9 | 25
Figure 3. National Vegetation Information System (NVIS) symbolised by broad vegetation types.

Interim Biogeographic Regionalisation for Australia (IBRA)

The national and regional planning framework for the systematic development of a
comprehensive, adequate and representative ‘CAR’ National Reserve System is provided by the
Interim Biogeographic Regionalisation for Australia (IBRA).
The latest version, IBRA7, classifies Australia's landscapes into 89 large geographically distinct
bioregions based on common climate, geology, landform, native vegetation and species
information. For example, the Australian Alps, the Nullabor Plain and the Wet Tropics are
distinct bioregions. The 89 bioregions are further refined to form 419 subregions which are
more localised and homogenous geomorphological units in each bioregion. The bioregions and
subregions are defined in the IBRA7 bioregional map. IBRA is a more detailed subset of the
global ecoregions. For AUS GEEBAM the IBRA 7 bioregions were converted to a WGS84 40m
raster (Figure 4).

P a g e 10 | 25
Figure 4. Interim Biogeographic Regionalisation for Australia (IBRA) converted to a WGS84 40m raster.

Sentinel-2 (S2) Surface Reflectance (L2)

Sentinel-2 is a wide-swath, high resolution, multispectral imaging mission with a global 5-day
revisit frequency from the European Space Agency (ESA). The S2 Multispectral Instrument
(MSI) samples 13 spectral bands: visible and Near Infra-Red (NIR) at 10 meters, red edge and
Short Wave Infra-Red (SWIR) at 20 meters, and atmospheric bands at 60 meters spatial
resolution. It provides data suitable for assessing state and change of vegetation, soil, and water
cover.
For use in Google Earth Engine, the Sentinel-2 surface reflectance (L2) data are downloaded
from the Copernicus Open Access Hub. Removal of clouds and atmospheric and topographic
corrections are performed using sen2cor. See the Sentinel-2 User Handbook for details.

Creating time series mosaics of median surface reflectance

AUS GEEBAM relies on a mosaic of a series of Sentinel-2 images. The SWIR bands at 20 m
mitigate smoke and haze in the atmosphere, and thus provides useful imagery for bushfire
events.
AUS GEEBAM uses five national mosaics created as the basis for vegetation indices. The dates
were chosen to represent the late fires in southern Australia up to May 15th, 2020. However, in
this mosaic regrowth in northern NSW and QLD was clearly apparent as a “greening up” of the
image. An early season image was used in tandem to capture fires earlier in the season up to the

P a g e 11 | 25
15th of February 2020. To ensure the pre-fire mosaic was independent of these time series, and
free of cloud and shadows, a time series from April 2018 to April 2019 was used.
The tropics have an entirely different fire season so two mosaics were created to capture fires in
northern Australia up until the 15th of January 2020 (Table 1).
Table 1. Date ranges of time series mosaics calculated as the median pixel value.

Season Version Name Pre-fire Image Dates Post-fire Image Dates NIAFED Date
South
AUS GEEBAM 20180415 - 20190415 20191115 - 20200215 20200224
Early
South
AUS GEEBAM 20180415 - 20190415 20200115 - 20200515 20200224
Late
Tropics AUS GEEBAM 20181015 - 20190115 20191015 - 20200115 20200224

Vegetation Indices
AUS GEEBAM uses the difference between the normalized burnt ratio (NBR) before and after
fire. NBR is an index designed to highlight burnt areas in large fire zones. The formula is similar
to the normalized difference vegetation index (NDVI), but it uses near-infra red (NIR) and short
wave infra-red (SWIR2) wavelengths.
NIR − SWIR2
NBR =
NIR + SWIR2
The difference between the pre-fire and post-fire NBR is denoted as dNBR. A higher value of
dNBR indicates increased likelihood that the area has burnt, while areas with negative dNBR
values may indicate regrowth following a fire. The dNBR formula is demonstrated below:
dNBR = Pre − fire NBR − Post − fire NBR

Relativised Normalized Burnt Ratio (RNBR)

Parks et al (2014) proposed a Landsat-based burn severity metric, the relativized burn ratio
(RNBR). Simply put, RNBR is the dNBR divided by an adjustment to the pre-fire NBR. Adding
1.001 to the denominator ensures that the denominator will never be zero, thereby preventing
the equation from reaching infinity and failing.
A relativized version of burn severity is advantageous where pre-fire vegetative cover is low.
Low NBR will generally have low dNBR values regardless of the degree of fire-induced mortality
of the vegetation. As an absolute measure of change, dNBR simply does not allow for the
quantification of high severity in these cases, even if all pre-fire vegetation is consumed. As a
result, dNBR values are often correlated to pre-fire NBR. The denominator in the equation for
RNBR removes this correlation and allows the metric to be more sensitive than dNBR to
changes where pre-fire vegetation cover is low. As such, a relative index like RNBR is
theoretically more suited to detect changes to vegetation on a consistent scale:
dNBR
RNBR =
(NBRpre − fire + 1.001)
To incorporate RNBR values for both the early and late fire season Sentinel 2 mosaics, the
maximum RNBR value was selected for each pixel to create MaxRNBR used in creating severity
classes.

P a g e 12 | 25
Reclassifying continuous RNBR values as ordinal quantiles
AUS GEEBAM relies on Sentinel-2 satellite imagery and vegetation indices from before and after
fires. RNBR classes are created by dividing the frequency distribution into equal groups, each
containing the same fraction of the total population (~0.235 RNBR increments).

Calibrating RNBR quantiles using visual interpretation of remote sensing

To determine a reference unburnt condition, the NIAFED extent was buffered by 2km. For each
NVIS broad vegetation type in each IBRA bioregion a reference unburnt RNBR class was
determined. That value was available to calculate a standardised offset or a reference unburnt
value.
Each IBRA bioregion was systematically assessed to correct for obvious errors. For example, the
Very High severity class could be adjusted down by one RNBR Value for a fire where its extent
extended into an area of lower severity. Conversely, there were areas of shrublands that had
clearly burnt at Very High severity where all of the biomass is likely to have been consumed but
low pre-fire biomass had given it a lower RNBR Value.
Rainforest was calibrated independently to detect subtle changes in RNBR. Despite the apparent
low severity of fires in these systems, most rainforests are sensitive to fire and recovery can be
slow (Hjerpe et al. 2001, Cochrane and Schulze 1999, Brando et al. 2014). Post-fire monitoring
is warranted in these situations to ascertain the response of rainforest to low severity fires.
Each pixel of AUS GEEBAM contains the raw RNBR Value, the RNBR Class and the GEEBAM
Value. This allows an end user to observe which values have been adjusted during the
calibration away from the default global RNBR Value and allows for some transparency in the
process.
The process, while complex to explain, is simple to implement. Calibration of RNBR for AUS
GEEBAM was completed by a single operator in two days. Using a single operator helps reduce
observer error, as does the systematic process and the reference unburnt values. Given that
single index thresholds are known to feature poor delineation of low fire severity classes
(Collins et al. 2018), the Low and Moderate severity classes were combined (Table 2).

Table 2. AUS GEEBAM classes with descriptions.

GEEBAM GEEBAM
Description
Value Class
No data indicates areas outside NIAFED or NVIS categories that
1 No data
do not represent native vegetation (e.g. cleared land, water)
Little or no change observed between pre-fire and post-fire
2 Unburnt
imagery
Low and Some change or moderate change detected when compared to
3
Moderate reference unburnt areas outside the NIAFED extent
4 High Vegetation is mostly scorched
5 Very high Vegetation is clearly consumed

Evaluating AUS GEEBAM

AUS GEEBAM was calibrated using visual interpretation of remote sensing data and no field
data was used. Field data is still being collected and will be required to provide confidence
intervals and a validation of map accuracy.

P a g e 13 | 25
As an interim solution a multiple lines of evidence approach was taken to evaluate AUS
GEEBAM. To explore the relationship between AUS GEEBAM and fire severity mapping from
Victoria (VIC), New South Wales (NSW) and South Australia (SA), a set of points were randomly
stratified as the basis of a comparison. This will help users understand the strengths and
limitations of AUS GEEBAM and how it can be used to supplement the fire severity maps from
local agencies (Table 3).
The points were used to calculate a measure of overall accuracy for AUS GEEBAM (“Classified”)
relative to local products (“Ground Truth”). The points were stratified by fire severity mapping
from VIC, NSW and SA and intersected with those fire severity maps (resampled to 40m
WGS84) and AUS GEEBAM (40m WGS84).
It is important to note that this is not a traditional validation as there is no point of truth. Both
maps will contain errors.

Table 3. Fire severity mapping from Victoria (VIC), New South Wales (NSW) and South Australia (SA)

State or Dataset
Description Custodian Link
Territory Name

Fire severity is a metric of the loss of biomass

caused by fire. In collaboration with the NSW
Fire Extent
Rural Fire Service, DPIE Remote Sensing &
and
Regulatory Mapping team has developed a
NSW Severity DPIE NSW
semi-automated approach to mapping fire
Map
extent and severity through a machine
(FESM)
learning framework based on Sentinel-2
satellite imagery.

Fire severity classification of bushfires

(wildfires) impacting ~1.5 million hectares of
Fire
predominantly forested public land in eastern
VIC severity DELWP VIC
and north-eastern Victoria (and ~300,000 ha
map
of southern NSW), between November 2019
and March 2020.

Fire severity mapping based on Landsat 8 /

Sentinel-2 imagery has been completed for:
Fire Cudlee Creek, Keilira & Bunbury Fires (SE),
SA severity Kangaroo Island, and Miltalie. DEW SA
map
Fire severity = Pre-fire image NBR – Post-fire
image NBR.

Equivalence was determined for each severity map for all four AUS GEEBAM classes (Table 4)
representing “Unburnt”, “Low and Moderate severity”, “High severity” and “Very High Severity”.

P a g e 14 | 25
Table 4. Fire severity map evaluation class equivalent for four classes

AUS
Test AUS NSW Fire NSW VIC Fire VIC SA Fire SA
GEEBAM
Class GEEBAM Severity Value Severity Value Severity Value
Value
Non-FESM
mapped
Not Not Non-woody,
1 burnt area, 1 1 No data
used Native No data
No data,
grasslands
Unburnt
C_2 Unburnt 2 Unburnt 0 Unburnt 2 1
(< 0.25)
Low , Medium
Low and Low, Low
C_3 3 2, 3 canopy 3, 4 2
Moderate Moderate (0.25-0.50)
scorch
High High canopy Medium
C_4 High 4 4 5 3
Severity scorch (0.50-0.75)
Extreme High
C_5 Very High 5 5 Canopy burnt 6 4
Severity (> 0.75)

AUS GEEBAM is a rapid triage tool for decision makers where the question may simply be, “How
much has burnt and where are the really severe fires?”. Therefore, map evaluation class
equivalents were also created for two classes representing “Moderate” and “High”.
Classes that were not used in the severity equivalence classes include No Data, Not Native
Vegetation and burnt areas not mapped in NSW. Points were only selected that fell within the
NIAFED extent.

Table 5. Fire severity map evaluation class equivalent for two classes

AUS
Test AUS NSW Fire NSW VIC Fire VIC SA Fire SA
GEEBAM
Class GEEBAM Severity Value Severity Value Severity Value
Value
Non-FESM
mapped
Not Not Non-woody,
1 burnt area, 1 1, 2 No data
used Native, No data
No data,
grasslands
Unburnt, Unburnt
Unburnt, Unburnt,
Low, Medium (< 0.25),
C_2 Low and 2, 3 Low, 2, 3 3, 4 1, 2
canopy Low
Moderate Moderate
scorch (0.25-0.50)
High Medium
High canopy
High and Severity, (0.50-0.75),
C_5 4, 5 4, 5 scorch/High 5, 6 3, 4
Very High Extreme High
burnt
Severity (> 0.75)

A confusion matrix was calculated on errors of omission and commission. It calculated the
user's accuracy and producer's accuracy for each class as well as overall accuracy. These
accuracy rates range from 0 –1, with 1 representing 100 percent accuracy.
User's accuracy shows false positives, where pixels are incorrectly classified as a known class
when they should have been classified as something else. Producer's accuracy is a false negative,
where pixels of a known class are classified as something other than that class. The overall

P a g e 15 | 25
accuracy is calculated as the total number of correctly classified pixels (diagonal elements)
divided by the total number of test pixels.

Results
The most powerful tool used in AUS GEEBAM is the automated creation of times series Sentinel-
2 mosaics. AUS GEEBAM uses the median from a time series of surface reflectance images to
mitigate cloud, shadow, and haze. This allows for the creation of multiple national mosaics from
a series of dates in two days. That includes date selection and visual assessment, downloading,
mosaicking and the creation of vegetation indices (Figure 5).

Figure 5. Sentinel-2 Surface Reflectance WGS84 40m mosaic.

The three mosaics that made up the southern MaxRNBR were pre-fire NBR, early season post-
fire NBR and late season NBR (Table 1). The MaxRNBR was divided into quantiles that are listed
in Table 6.

P a g e 16 | 25
Table 6. Reclassifying continuous RNBR Values as ordinal quantiles.

RNBR RNBR RNBR RNBR

Count Hectares
Value From To Class
-3 -4.999 -0.043 2 4347 701
-2 -0.043 -0.020 2 8911 1437
-1 -0.020 0.004 2 31028 5004
0 0.004 0.004 2 22584685 3642546
2 0.004 0.027 2 17902288 2887351
3 0.027 0.051 2 17575749 2834685
4 0.051 0.074 2 14700823 2371006
5 0.074 0.098 2 13379492 2157897
6 0.098 0.121 2 12742853 2055217
7 0.121 0.145 2 11989747 1933753
8 0.145 0.168 3 11029982 1778959
9 0.168 0.192 3 10100610 1629066
10 0.192 0.215 3 9249255 1491756
11 0.215 0.239 3 8410640 1356501
12 0.239 0.262 3 7601680 1226029
13 0.262 0.286 3 6845718 1104104
14 0.286 0.309 3 6138827 990094
15 0.309 0.333 3 5459173 880477
16 0.333 0.356 4 4819203 777260
17 0.356 0.380 4 4204033 678043
18 0.380 0.403 4 3636479 586506
19 0.403 0.427 4 3124301 503899
20 0.427 0.450 4 2674641 431377
21 0.450 0.497 4 4156869 670436
22 0.497 0.544 5 2925668 471863
23 0.544 0.591 5 1999976 322564
24 0.591 0.990 5 1972124 318072

Table 6 includes the RNBR Value with the corresponding range of raw RNBR values, the global
default RNBR class, the number of pixels and the number hectares this count represents. RNBR
classes are available in AUS GEEBAM and can be displayed as greyscale as a representation of
raw RNBR values.
Each pixel of AUS GEEBAM (Figure 6) contains the raw RNBR Value, the RNBR Class and the
GEEBAM Value. This allows an end user to observe which values have been adjusted during the
calibration away from the default global RNBR Value and allows for some transparency in the
results.

P a g e 17 | 25
Figure 6. Australian Google Earth Engine Burnt Area Map 4.0

Randomly select points were stratified by the fire severity map products from NSW (n = 1975),
VIC (n = 1841) and Kangaroo Island in SA (n = 252). The “ground truth” points from each State’s
map were then compared with the AUS GEEBAM classification results at the same locations.
The points were stratified by each state’s maps. When using four classes (Unburnt, Low and
Moderate, High, and Very High severity) the overall accuracy of AUS GEEBAM was between 48%
and 82%. When using two simple classes (Moderate, High severity) the overall accuracy rose to
between 72% and 92% (Appendix).
The figures below illustrate how the Sentinel 2 imagery (Figure 7), and AUS GEEBAM (Figure 9)
compare to fire severity mapping from Victoria (Figure 8). There are also images of the fires
impact on Kangaroo Island (Figure 10) in South Australia with the corresponding AUS GEEBAM
Classes (Figure 11).

P a g e 18 | 25
Figure 7. South Eastern Australia – False Colour Mosaic - GEEBAM

Figure 8. South Eastern Australia – Victorian fire severity map – DELWP

P a g e 19 | 25
Figure 9. South Eastern Australia – AUS GEEBAM fire severity classes – AUS GEEBAM

Figure 10. South Australia – Kangaroo Island - False Colour Mosaic – GEEBAM

P a g e 20 | 25
Figure 11. South Australia – Kangaroo Island - GEEBAM fire severity classes – AUS GEEBAM

Discussion
State and Territory fire severity maps are likely to be more accurate than AUS GEEBAM,
particularly where they use ground truth data to train the interpretation of satellite imagery.
However, all data based on interpretation of remotely sensed imagery is subject to some degree
of error. This makes validation of AUS GEEBAM against state and territory datasets challenging.
Where the datasets disagree, there is potential for error in both the state and territory data and
AUS GEEBAM. Future work will focus on validating AUS GEEBAM based on field data and/or
interpretation using very high-resolution remote sensing data. A nationally coordinated and
openly accessible set of fire ground truth data would be a valuable asset for this work.
As the intention of AUS GEEBAM was solely to provide a nationally consistent fire severity map,
AUS GEEBAM is likely to be less nuanced at low severities than state and territory products, and
will not reflect what is found in the field as well as products created with local expertise and
calibration data.
AUS GEEBAM is a rapid triage tool for decision makers where the question may simply be, “How
much has burnt and where are the really severe fires?”. It is designed to be as simple as possible
and to be reproduced nation-wide as required.
With a set of training data points, AUS GEEBAM could implement a more sophisticated approach
such as the use of multiple indices and deep learning classification without any compromise in
scalability and speed. Tools like AUS GEEBAM will become increasingly important as impacts of

P a g e 21 | 25
climate change are realised. They build the capacity for rapid conservation responses, and allow
further exploratory investigations into the magnitude and behaviour of bushfire events.
The rapid spatial data provided by the GEEBAM system will be a useful addition to inform
public policy decision making for fire and landscape management. Consistent national coverage
and classification are important when prioritising responses to widespread bushfire events,
with AUS GEEBAM delivering on these requirements.
It is important that there is continuous development of rapid assessment tools to provide
spatial data and identification of potential fire refuges and areas of very high severity.

Appendix
New South Wales Fire Extent and Severity Map
(cvmsre_NSWInterim_20200419_ag7l0)
[Link]
Fire Extent and Severity Map (FESM) covers most of the 5.7 million hectares affected by
bushfires in NSW between September 2019 and April 2020.
FESM defines fire severity as a metric of the loss of biomass caused by fire. The state-wide
severity map has standardised classes to allow comparison of different fires across the
landscape. The FESM severity classes include unburnt, low severity (burnt understory, unburnt
canopy), moderate severity (partial canopy scorch), high severity (complete canopy scorch,
partial canopy consumption), extreme (full canopy consumption).

Table 7. NSW fire severity map – 4 class confusion matrix

ClassValue C_2 C_3 C_4 C_5 Total U_Accuracy Kappa

C_2 128 141 33 2 304 0.421053 0
C_3 167 366 178 6 717 0.51046 0
C_4 97 165 337 51 650 0.518462 0
C_5 60 11 107 126 304 0.414474 0
Total 452 683 655 185 1975 0 0
P_Accuracy 0.283186 0.535871 0.514504 0.681081 0 0.484557 0
Kappa 0 0 0 0 0 0 0.279764

Table 8. NSW fire severity map – 2 class confusion matrix

ClassValue C_2 C_5 Total U_Accuracy Kappa

C_2 802 219 1021 0.785504 0
C_5 333 621 954 0.650943 0
Total 1135 840 1975 0 0
P_Accuracy 0.706608 0.739286 0 0.720506 0
Kappa 0 0 0 0 0.438166

P a g e 22 | 25
Victorian Fire severity map
[Link]
and-north-east-victoria-in-2019-20-version-1-
The Victorian fire severity map covers ~1.5 million hectares of fire affected areas of
predominantly forested public land in eastern and north-eastern Victoria (and ~300,000 ha of
southern NSW), between November 2019 and March 2020.
Fire severity mapping was derived using machine learning classification (Random forests) of
eight Spectral Indices (SI) from pre and post-fire Sentinel-2 satellite imagery. The fire severity
classification has 6 classes, (6) High canopy consumption resolution - 20% canopy foliage
consumed), (5) High canopy scorch - >80% of canopy foliage is scorched, (4) Medium canopy
scorch - Canopy is a mosaic of both unburnt and scorched foliage, 20 - 80%, (3) Low canopy
scorch - Canopy foliage is largely unaffected (90%) unburnt.

Table 9. VIC fire severity map – 4 class confusion matrix

ClassValue C_2 C_3 C_4 C_5 Total U_Accuracy Kappa

C_2 103 263 3 0 369 0.279133 0
C_3 35 429 148 5 617 0.6953 0
C_4 1 72 401 73 547 0.73309 0
C_5 0 9 129 170 308 0.551948 0
Total 139 773 681 248 1841 0 0
P_Accuracy 0.741007 0.554981 0.58884 0.685484 0 0.599131 0
Kappa 0 0 0 0 0 0 0.436746

Table 10. VIC fire severity map – 2 class confusion matrix

ClassValue C_2 C_5 Total U_Accuracy Kappa

C_2 830 156 986 0.841785 0
C_5 82 773 855 0.904094 0
Total 912 929 1841 0 0
P_Accuracy 0.910088 0.832078 0 0.870722 0
Kappa 0 0 0 0 0.741615

South Australian Fire severity map

The SA fire severity map was calculated using the following methodology:
1. Pre and post-fire imagery using either Landsat 8 / Sentinel-2 imagery level 2 (bottom of
atmosphere surface reflectance) downloaded from USGS / Copernicus Open Access Hub/NCI.
2. Band stacking to datasets of same resolution.
3. Normalised Burn Ratio = (NIR – SWIR) / (NIR + SWIR) applied to pre and post images.
Landsat 8: Band 4 (NIR), Band 7 (SWIR).
Sentinel-2: Band 8A (NIR) / Band 12 (SWIR).
4. Fire severity = Pre fire image NBR – Post fire image NBR.
Classes are: Unburnt (< 0.25), Low (0.25-0.50), Medium (0.50-0.75), and High (> 0.75).

P a g e 23 | 25
Table 11. SA fire severity map – 4 class confusion matrix

Class Value C_2 C_3 C_4 C_5 Total User Accuracy Kappa
C_2 41 29 0 0 70 0.585714 0
C_3 3 19 6 0 28 0.678571 0
C_4 0 6 39 4 49 0.795918 0
C_5 0 0 14 91 105 0.866667 0
Total 44 54 59 95 252 0 0
P Accuracy 0.93181 0.351852 0.661017 0.957895 0 0.753968 0
Kappa 0 0 0 0 0 0 0.660687

Table 12. SA fire severity map – 2 class confusion matrix

Class Value C_3 C_5 Total User Accuracy Kappa

C_3 92 6 98 0.938776 0
C_5 6 148 154 0.961039 0
Total 98 154 252 0 0
P Accuracy 0.938776 0.961039 0 0.952381 0
Kappa 0 0 0 0 0.899814

References
Abbott, B. W., J. B. Jones, E. A. Schuur, F. S. Chapin III, W. B. Bowden, M. S. Bret-Harte, H. E. Epstein,
M. D. Flannigan, T. K. Harms, and T. N. Hollingsworth. 2016. Biomass offsets little or none of
permafrost carbon release from soils, streams, and wildfire: an expert assessment.
Environmental Research Letters 11:034014.
Barlow, J., and C. A. Peres. 2008. Fire-mediated dieback and compositional cascade in an Amazonian
forest. Philosophical Transactions of the Royal Society B: Biological Sciences 363:1787-1794.
Bladon, K. D., M. B. Emelko, U. Silins, and M. Stone. 2014. Wildfire and the future of water supply.
Environmental Science & Technology 48:8936-8943.
Bradstock, R. A., K. A. Hammill, L. Collins, and O. Price. 2010. Effects of weather, fuel and terrain on
fire severity in topographically diverse landscapes of south-eastern Australia. Landscape
ecology 25:607-619.
Collins, L., A. F. Bennett, S. W. Leonard, and T. D. Penman. 2019. Wildfire refugia in forests: Severe
fire weather and drought mute the influence of topography and fuel age. Global Change
Biology 25:3829-3843.
Collins, L., R. A. Bradstock, E. M. Tasker, and R. J. Whelan. 2012. Impact of fire regimes, logging and
topography on hollows in fallen logs in eucalypt forest of south eastern Australia. Biological
Conservation 149:23-31.
Collins, L., P. Griffioen, G. Newell, and A. Mellor. 2018. The utility of Random Forests for wildfire
severity mapping. Remote Sensing of Environment 216:374-384.
Dickman, C. R., D. A. Driscoll, S. T. Garnett, D. A. Keith, S. Legge, D. Lindenmayer, M. Maron, A.
Reside, E. G. Ritchie, D. M. Watson, B. A. Wintle, and J. Woinarski. 2020. After the
catastrophe: a blueprint for a conservation response to large-scale ecological disater.
Threatened Species Recovery Hub, Lucia, Australia.

P a g e 24 | 25
Esque, T. C., C. R. Schwalbe, L. A. Defalco, R. B. Duncan, and T. J. Hughes. 2003. Effects of desert
wildfires on desert tortoise (Gopherus agassizii) and other small vertebrates. The
Southwestern Naturalist 48:103-111.
Ice, G. G., D. G. Neary, and P. W. Adams. 2004. Effects of wildfire on soils and watershed processes.
Journal of Forestry 102:16-20.
Keeley, J. E. 2009. Fire intensity, fire severity and burn severity: a brief review and suggested usage.
International Journal of Wildland Fire 18:116-126.
Keppel, G., K. P. Van Niel, G. W. Wardell‐Johnson, C. J. Yates, M. Byrne, L. Mucina, A. G. Schut, S. D.
Hopper, and S. E. Franklin. 2012. Refugia: identifying and understanding safe havens for
biodiversity under climate change. Global Ecology and Biogeography 21:393-404.
Nolan, R. H., M. M. Boer, L. Collins, V. Resco de Dios, H. Clarke, M. Jenkins, B. Kenny, and R. A.
Bradstock. 2020. Causes and consequences of eastern Australia’s 2019‐20 season of mega‐
fires. Global Change Biology Online early.
Parks, S. A., G. K. Dillon, and C. Miller. 2014. A New Metric for Quantifying Burn Severity: The
Relativized Burn Ratio. Remote Sensing 6:1827-1844.

P a g e 25 | 25