bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
1
Beyond Traditional Methods: Innovative Integration of LISS IV and Sentinel 2A
2
Imagery for Unparalleled Insight into Himalayan Ibex Habitat Suitability
3
Ritam Dutta1,2,3, Bheem Dutt Joshi1, Vineet Kumar1, Amira Sharief1, Saurav Bhattcharjee1
4
Rajappa Babu3, Mukesh Thakur2, Lalit Kumar Sharma1,*
5
6
1
Zooological Survey of India, Prani Vigyan Bhawan, Kolkata, West Bengal 700053
7
2
University of Madras, Navalar Nagar, Chepauk, Triplicane, Chennai, Tamil Nadu 600005
8
3
Southern Regional Centre, Zoological Survey of India, Chennai, Tamil Nadu 600028
9
*Corresponding author: Lalit Kumar Sharma
10
lalitganga@gmail.com
11
12
Abstract:
13
Despite advancements in remote sensing, satellite imagery is underutilized in conservation
14
research. Multispectral data from various sensors have great potential for mapping
15
landscapes, but distinct spectral and spatial resolution capabilities are crucial for accurately
16
classifying wildlife habitats. Our study aimed to develop a technique for precisely discerning
17
habitat categories for the Himalayan Ibex (Capra sibirica) using different satellite imagery.
18
To address both spectral and spatial challenges, we utilized LISS IV and Sentinel 2A data and
19
integrated the LISS IV data with Sentinel 2A data along with their corresponding geometric
20
information. Employing multiple supervised classification algorithms, we found the Random
21
Forest (RF) algorithm to outperform others. The integrated (LISS IV-Sentinel 2A) classified
22
image achieved the highest accuracy, with an overall accuracy of 86.17% and a Kappa
23
coefficient of 0.84.
24
To map the suitable habitat of the Ibex, we conducted ensemble modeling using the
25
Land Cover Land Use (LCLU) of all three image types (LISS IV, Sentinel 2A, Integrated)
26
and other predictors such as topographical, soil type, vegetation, and water radiometric
27
indices. The integrated model provided the most accurate prediction of suitable habitat for the
28
Ibex, surpassing the other two LCLU classes derived from individual images. The Soil
29
Adjusted Vegetation Index (SAVI) and elevation were identified as crucial factors in
30
identifying suitable habitats.
31
These findings hold valuable implications for the development of effective conservation
32
strategies, as accurate classification schemes enable the identification of vital landscape
33
elements. By precisely classifying LULC satellite images and identifying crucial habitats for
34
the Ibex, this pilot study provides a new and valuable strategy for conservation planning. It
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
35
enhances our ability to preserve and protect the habitat of wildlife species in the mountain
36
ecosystem of the Himalayas.
37
Keywords: Image classification, Integration image, Ensemble species distribution model,
38
Himalayan Ibex.
39
1. Introduction:
40
Satellite imagery is full of detail and contributes significantly to the dissemination of
41
geographical information (Muhammad et al. 2012). The utilisation of satellite and remote
42
sensing images provides both quantitative and qualitative data, which streamlines fieldwork
43
and shortens research times (Chaichoke et al. 2011). Satellite remote sensing techniques
44
capture images at regular intervals. Satellites with various band ranges, geographical
45
resolutions, and spectral resolutions have made it possible to gather more types of remote
46
sensing data from the same area (Wang et al., 2001). The significance of proficient analysis
47
and processing of remote sensing images has been on the rise owing to the exponential
48
growth of remote sensing data (Yin et al. 2021). Because of its wide range of potential
49
applications in fields as diverse as geography, ecology, city planning, forest monitoring, and
50
the military, remote sensing image scene classification is receiving increasing amounts of
51
research and development funding (Cheng et al. 2017). There are several different types of
52
multiresolution and multispectral data now accessible. The data acquired through
53
multispectral remote sensing is characterised by narrow spectral bands that possess a
54
relatively larger bandwidth. Consequently, the gathered data can be employed to examine the
55
spatial characteristics of ground substances (Vohra and Tiwari 2020). The limitations of a
56
single source of satellite data in accurately extracting ground objects are attributed to spectral
57
resemblance among different objects or spatial proximity between the objects. Consequently,
58
for the enhancement of data evaluation precision, it is imperative to appropriately construe
59
object characteristics such as configuration and spatial interconnections, in conjunction with
60
the spectral response (Vohra and Tiwari 2020).
61
It has become a challenge in remote sensing technology to figure out how to combine
62
different sourced data and produce the most useful image possible. However, in recent years
63
have seen an increase in the importance of image fusion within image processing applications
64
as a result of the plethora of available acquisition methods. Integrating many images from
65
various sensors into a single, useful one for analysis, is the goal of image fusion, which is a
66
relatively a new field (Ma et al. 2019). The comprehension of digital image fusion techniques
67
can enhance the interpretation of multiresolution and multi-sensor data, resulting in improved
68
images that are more suitable for both human perception and seamless computer analysis
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
69
tasks such as extraction of features, segmentation, and object recognition (Luo et al. 2016;
70
Berger et al. 2015). The fusion of data from multiple sensors and resolutions has been found
71
to be beneficial in enhancing the quality of low-resolution data. Additionally, this approach
72
offers supplementary information collected from the same geographical location, which can
73
be better comprehended than relying solely on data from a single sensor (Ma et al. 2018).
74
Combining the spatial and spectral characteristics of remote sensing images, image fusion
75
technology has gradually broadened its application field, and now the fusion image concept
76
has been applied to land cover land use (hereafter LCLU) categorization. Recently, there has
77
been an explosion of interest in using multi-sensor data fusion for LCLU classification.
78
Accuracy in land cover feature categorization can be improved through the integration of data
79
from many remote sensing sensors with varying resolutions. The demand for greater
80
precision in image and data analysis has spurred research into multiresolution and multi-
81
sensor data, as well as improved methods for gaining access to higher-resolution remote
82
sensing data (Rodriguez-Galiano et al. 2012a). The biophysical state of the Earth's surface
83
and immediate subsurface is characterised by the composition of topography, soil, surface
84
water, forest, grassland, groundwater, marsh and human structures, collectively referred to as
85
"land cover." Furthermore, land use for recreational purposes, wildlife habitat, and
86
agricultural land can the example of land use (Turner et al. 1995; Weng 1999; Sherbinin
87
2002).
88
It is obvious that land cover classification, derived from remotely sensed data, is still a
89
crucial societal need for natural resource management, surveillance, and development
90
strategies (Colditz et al. 2011; Topaloğlu et al., 2016; Khatami et al., 2016). Numerous
91
studies have generated information from remote sensing by making LCLU maps from various
92
data sources like multispectral, hyperspectral and radar aperture (Craig Dobson et al. 1995;
93
Soria-Ruiz et al., 2010; Pal and Foody 2010; Szuster et al., 2011; Miettinen and Liew, 2011;
94
Srivastava et al., 2012; Hütt et al., 2016; Fonteh et., 2016; Wei et al., 2016; Büyüksalih,
95
2016; Mohajane et al., 2018; Sirro et al., 2018; Juliev et al., 2019, Camargo et al., 2019,
96
Zafari et al. 2019), subsequently, the LCLU types were classified through the utilisation of
97
machine learning algorithms (Patenaude et al. 2005; Rosenqvist et al. 2003; Wulder et al.
98
2018). This type of mapping is useful for assessing landuse dynamics, identifying ecosystem
99
services, understanding the effects of global climate change, and formulating land use policy
100
(Fry et al., 2011; Burkhard et al., 2012; Gebhardt et al., 2014; Guidici and Clark, 2017; Noi
101
and Kappas 2017; Hussain et al., 2020).
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
102
Furthermore, with the growing statistical models nowadays Species distribution models
103
(hereafter SDM), is a great conservation tool which enhance the capabilities of conservation
104
managers to delaminate conservation priority areas on the base of species presence and the
105
association with their environment. SDM can directs to finds appropriate conservation
106
policies, estimating the area under invasion, evaluate species richness of an area and estimate
107
the probable habitat of any species (Franklin, 2010). SDMs perform crucial role in
108
quantitively ecology by systematically inspecting a species’ relation, in terms of evaluating
109
the biotic and abiotic factors for distributing an organism in a given area (Franklin, 2010).
110
The SDMs approach was developed based on Hutchinson's ecological niche theory, initially
111
introduced in the 1950s and later revised by Booth et al. (1988). Within the ecological
112
context, the species exhibits significant interactions with various factors such as dietary
113
resources, vegetation types, elevation profiles, and climatic elements (Morris et al., 2012;
114
Besnard et al., 2013). The SDMs employ several methods to determine how a species'
115
presence in the environment affects that species' ability to choose an area on a spatial surface
116
that would be favourable for that species (Guisan and Thuiller, 2005; Franklin, 2010;
117
Peterson et al., 2011). It is quite challenging to determine whether a species is actually
118
absent, even though occurrence records can be obtained through museums, published
119
literatures, and field studies. Several SDM methods have been created that solely use positive
120
presence data in order to address this difficulty (Phillips et al., 2006). Instead of using only
121
one modelling technique, the ensemble modelling strategy uses multiple SDM models, which
122
increases the accuracy of predictions about a species' geographic range (Araújo and New,
123
2007; Thuiller et al., 2009; Marmion et al., 2009). Due of the ambiguity in selecting one
124
strategy from numerous, ensemble modelling is more effective to a single SDM technique
125
(Pearson et al., 2006; Elith and Graham, 2009; Buisson et al., 2010; Garcia et al., 2011).
126
Himalayan Ibex, also known as Himalayan Ibex is a member of Bovidae family is a
127
true goat species. This caprinae have a wide distribution range in the mountains of India,
128
Pakistan, China, Mongolia, Afghanistan, Tajikistan, Uzbekistan, Kazakhstan, Kyrgyzstan and
129
Russia (Shackleton 1997, Fox et al.1991, Fedosenko and Blank 2001). This species has a
130
wide residence which ranging from Karakoram and Hindu-Kush mountains of Pakistan
131
through the higher elevated areas of Lahaul-Spiti and some other patches of Himachal
132
Pradesh. The existence of this mountain goat species in this region since 1641 CE as per
133
Saini et al. 2019. This species has very wide range of habitat in its distributional range which
134
mainly configure by steep slopes, mountain ridges, rugged terrain, rocky outcrops, cold
135
deserts and foothills (Dzieciolowski et al. 1980; Clark et al. 2006, Khan et al. 2016). It is the
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
136
largest among the Capra genus, adult males can reach up to 5.6 feet and weigh up to 130 kg
137
(Fedosenko and Blank 2001). The Himalayan Ibex are affable, and its social groups are
138
composed of both single-sex and mixed-sex individuals, with the latter being observed only
139
during certain seasons. By affecting plant species composition, vegetation structure, and
140
nutrient cycling, ungulates play a significant role in preserving ecosystems (McNaughton
141
1979; Bagchi and Ritchie 2010). Therefore, maintaining and managing ungulate populations
142
and their habitat is a key goal of conservation management. Habitat is one of the prerequisites
143
components which supports any species to survive. Therefore, monitoring habitat is one of
144
the crucial elements which can help to make any conservation strategy. Himalayan Ibex is an
145
“Near Threatened” mammalian species as per IUCN Red list (Reading et al. 2020), and
146
categories as a Schedule I species under the Wildlife (Protection) Act, 1972 in India. This
147
caprinae species lost their habitat with the rapid urbanization, habitat destruction, hunting and
148
poaching.
149
The present study aims at the integration of LISS IV and Sentinel 2A images over the
150
Jispa valley of Lahaul Spiti district, Himachal Pradesh, which is under Trans-Himalayan
151
landscape. The purpose of this research is to combine high spatial and spectral data to better
152
distinguish different types of LCLU and find how the classification help to predict the habitat
153
of Himalayan Ibex in this landscape.
154
2. Study area:
155
The present study was carried out in the Jispa valley site. The landmass segment is located in
156
the eastern part of the Lahaul valley in the Lahaul Spiti district of Himachal Pradesh, India
157
(Fig 1). Total area of the study area encompasses 559 km2, and lies in UTM zone 43N. This
158
landmass falls under Trans Himalaya Ladakh Mountains (1A) Indian biogeographic zones,
159
which geomorphology is very distinct. High mountains, inclining slopes, and sparse
160
vegetation are the main features of this area. This area has only two clearly defined seasons.
161
High snowfall is frequent during the winter. Farming is one of the primary sources of income
162
in this area; the main commercial crops, which are only grown in the summer, are potatoes,
163
peas, cauliflower, and cabbage. This area is intersected by the River Bhaga. This landmass is
164
extremely important since it supports a variety of crops, biodiversity, pasture, linking roads,
165
and human settlements.
166
3. Methodology:
167
3.1.
168
3.1.1. Data Acquisition
Land Class and Land Use Classification using multiple satellite imagery
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
169
In implementing this study, we assessed two different source satellites viz. Linear Imaging
170
Self-Scanning Sensor (LISS) IV and Sentinel 2A. LISS IV is a multi-spectral sensor with
171
high resolution and data providing in three spectral bands (viz. B2 0.52 µm - 0.59 µm, B3
172
0.62 µm - 0.68 µm, B4 0.77 µm - 0.86 µm). A ground resolution of 5.8 m (at nadir) provides
173
by LISS-IV. Moreover, a rotating deck is attached to a Payload Steering Motor mounted with
174
LISS-IV, which can revolve by 26 degrees and allow for a 5-day revisit of any given ground
175
region. Since the system has 10-bit quantization and can cover 100% of the albedo with a
176
single gain, no gain commands are necessary. Furthermore, Sentinel 2A is an optical
177
multispectral imaging mission with a wide-swath and high resolution. The Global Navigation
178
Satellite System (GNSS), a dual-frequency receiver with an orbital accuracy-specific
179
propulsion system, assists in maintaining each satellite's position in orbit. The multispectral
180
optical instrument of the Sentinel 2A satellites contains 13 spectral bands (Visible, Near-
181
Infra-Red and Short Wave Infra-Red) with spatial resolutions of 10 m, 20 m, and 60 m for the
182
various spectral bands with a 290 km swath width. The satellite's sun-synchronous orbit is at
183
a mean altitude of 786 km, and it completes a 5-day cycle with the two satellites (Drusch et
184
al.,
185
(https://bhoonidhi.nrsc.gov.in/) (Table 1) and Sentinel 2A (Table 1) Level-1C multispectral
186
instrument
187
(https://earthexplorer.usgs.gov/).
188
The exact geometric correction and registration of two images is the most fundamental
189
requirement for accurate image classification (Balcik, F.B. and Sertel, E., 2002). Sentinel 2A
190
data was pre-processed in the Sentinel Application Platform (SNAP Desktop, Version 6.0.0)
191
for resolution enhancement of the bands by the highest resolution of the other bands and all
192
the bands were converted into the same resolution (Brodu,2017). The satellite image band
193
was stacked in ArcGIS 10.6 environment with layer stacked function. The amalgamation of a
194
higher spatial resolution image and a higher spectral resolution image is advantageous for
195
remote sensing research (Chitade and Katiyar, 2012). Image fusion usually involves two main
196
processes, (a) Image to image geometric registration of two datasets geometry and (b)
197
combining spectral and spatial data to produce a new, enriched dataset that differs from the
198
originals (Lu and Weng, 2007).
2012).
LISS
scenes
IV
were
imagery
downloaded
was
from
collected
United
from
States
Bhoonidhi
Geological
portal
Survey
199
Image to image geometric registration is the process of superimposing images (two or
200
more) of the same scene that were captured at various times, from various angles, and or by
201
various sensors (Brown, 1992). The image-to-image geometric registration method was
202
assigned to the satellite images in order to run an accurate fusion procedure. We have used
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
203
the Coregistration QGIS plugin (Scheffler et al. 2017) for the procedure, where we use the
204
LISS IV image as a reference image and the Sentinel 2A image as a target image.
205
3.1.2. Integration of the two different satellite images:
206
As we mentioned, we used SNAP for the Sentinel 2A bands to convert the 20 meters bands at
207
the lowest resolution i.e., 10 meters. After that, we applied the nearest neighbour resampling
208
method (Wenbo et al. 2008) to minimize the loss of spectral information of Sentinel 2A
209
images and to gain a spatial resolution of 5 meters, like the LISS IV images. Furthermore, in
210
the integrating image the three bands of LISS IV assembled with the Sentinel 2A bands by
211
the replacement of the 3 bands of Sentinel 2A (i.e., B3, B4, B8; Green, Red and NIR bands
212
respectively). Since, the LISS IV has B2, B3 and B4 (Green, Red and NIR, respectively)
213
bands and the B2, B3 and B8 of Sentinel 2A bands have a similar spectral resolution. After
214
the integration of the LISS IV and Sentinel 2A bands total bands were 10, now it has good
215
spatial and spectral resolution as well as it gains bands number.
216
3.1.3. Image Classification:
217
The extraction and interpretation of valuable information from enormous satellite images
218
necessitate an effective and efficient statistical methodology. Image classification is a method
219
that involves the categorization of each pixel in an image or raw information obtained from
220
remote sensing satellites, with the aim of generating an appropriate set of labels for specific
221
land cover themes (Lilles and, Keifer 1994; Abburu and Golla 2015; Karlsson 2003).
222
In supervised satellite image classification methods, the training sample is the most crucial
223
element. Analyst input is required for supervised classification methods. The fundamental
224
process of supervised classification involves the examination of input data, generation of
225
training sets and signature files, and the assessment of the training sets and signature files'
226
efficacy (Abburu and Golla 2015).
227
In this study we classify the images (LISS IV, Sentinel 2A, and the integrated image)
228
by five supervised machine learning classification techniques viz. Maximum Likelihood, K-
229
nearest neighbourhood, Support Vector machine, Guassian mixed model and Random Forest
230
algorithm. In this present study the images classified into nine classes of LCLU types:
231
agriculture land, sparse vegetation, barren land, scrub areas, juniper patch, settlements,
232
permafrost, water bodies and road ways. The training data utilised in this classification
233
process was gathered for all eight LCLU classes, with the exception of permafrost, through
234
field surveys conducted within the study landscape. We have used the same training polygons
235
to classify the LISS IV, Sentinel 2A as well as the integrated image.
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
236
The idea is to evaluate the performance of various satellite images using their bands
237
as variables and to calculate the precision of mapping quantification by applying remote
238
sensing data to actual ground-truth circumstances. The accuracy evaluation of each defined
239
class is determined by an error matrix that compares map information with reference data and
240
the sampled area or points. Such errors are attributed to the accuracy of the producer and user
241
(Congalton and Green 2019; Foody 2002). The accuracy is derived from a final classification
242
error matrix made up of many multivariate statistical studies, where the overall accuracy,
243
kappa, and F-statistics represent the accuracy of various classified classes (Congalton et al.,
244
1983; Congalton 1991; Foody 2002; Nitze et al. 2012; Keshtkar et al. 2017, Gumma et al.
245
2019).
246
As a parametric statistical and supervised classification method, the Maximum
247
Likelihood (hereafter ML) classifier is the most popular technique used in remote sensing
248
applications (Jia et al. 2011). This approach uses an average with values, variance, and
249
covariance classification technique that is based on statistics and takes into account the values
250
of the variables (Günlü 2021). However, Gaussian mixture model (hereafter GMM) is
251
another machine learning algorithm. They are employed to divide data into various groups in
252
accordance with the probability distribution. GMMs uses to classify satellite images based on
253
probabilistic concept (Gu et al. 2007; Chellappa et al. 2009; Okwuashi et al., 2011; Lakshmi
254
et al. 2015). One of the simplest machine learning and supervised learning methods is the K-
255
nearest neighbour algorithm (hereafter KNN).
256
classifying objects by utilising the proximity of training samples in the feature space
257
(Bremner et al., 2005, Nurwauziyah et al., 2018). Support Vector Machine (hereafter SVM)
258
is essentially a supervised machine learning technique. The structural risk minimization
259
concept and statistical learning theory serve as the foundation for the SVM (Günlü 2021).
260
This algorithm is typically used to address classification-related challenges, while it can also
261
be used to address regression-related complications. SVM differentiates between the objects
262
based on the separation of the hyper-planes (Huang and Zhang,2012, Luo et al., 2015,
263
Melgani and Bruzzone 2004).
The KNN algorithm is a method for
264
It has been proved that Random Forest (hereafter RF) is capable of producing
265
accurate LCLU maps (Rodriguez-Galiano et al., 2012b; Sonobe et al., 2014). RF constructs
266
decision trees before randomly combining them (Breiman 2001; Liaw & Wiener, 2002;
267
Ishwaran & Kogalur, 2007; Ishwaran et al., 2008; Noi and Kappas 2017, Sonobe et al., 2017).
268
The supervised classification of these three images performed by dzetsaka classification tool,
269
SCP tool in QGIS and ArcGIS 10.6 (Esri 2018, Karasiak 2019, Congedo 2021)
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
270
3.2.
271
3.2.1. Occurrence locations of the Himalayan Ibex:
272
During the extensive field study during 2019-2022 we recorded the occurrence of Himalayan
273
Ibex. The collection of the occurrences of the species’ we followed camera trapping method,
274
trail sampling, vintage point and questionnaire survey. We have sampled all types of habitats
275
in the search of Himalayan Ibex. However, the study area composed of rugged terrain, steep
276
slopes, high mountains, uncertain weather conditions so representative sampling was carried
277
out. During the field work we gathered 167 presence locations, furthermore we used spatially
278
autocorrelated 82 locations for final analysis.
279
3.2.2. Variables preparation and selection:
280
The variable selection for SDM is the critical stage in this analysis because the variables
281
should be plausible for the Himalayan Ibex presence. The Trans-Himalayan landscape is
282
mostly rugged and found gradients in elevation with less amount of vegetation growth which
283
makes this area complex and this complexity also observe in the habitat selection by wild
284
animals like Himalayan Ibex. We used Digital elevation model data from Alos Palsar at 12.5-
285
meter spatial resolution and used as primary source for derived slope and aspect. The LCLU
286
classes were generated from the good accuracy produced by classified images. Furthermore,
287
radiometric indices of soil, vegetation and water computed from Sentinel 2A image. All the
288
data was rasterized and resampled at same spatial scale i.e. at 5 meter by the help of spatial
289
analyst extension tool of ArcGIS 10.6. Primarily we prepared 23 ecological pertinent
290
variables for this study, however, for the final model building only selected uncorrelated
291
variables through Pearson correlation coefficients (r) higher than 0.8.
292
3.2.3. Selection of performing models and evaluation of modeling algorithms:
293
Because of the complex and varied nature of the association between species and their
294
environmental predictors, studies indicate that no single modelling approach is best in every
295
circumstance (Marmion et al., 2009). Modelling algorithms broadly classified as
296
Classification models, Regression models and Complex models. Hence, we applied
297
Multivariate adaptive regression splines (MARS) and Generalized linear model (GLM) from
298
Regression models, Boosted regression trees (BRT) from Classification models and
299
Maximum Entropy Model (MaxEnt), Random Forest (RF) from Complex models, each form
300
with 10-fold cross-validation (Hayes et al. 2015, Dutta et. al 2022). We developed the
301
modelling workflow using the SAHM module by utilising the VisTrails pipeline; using this
302
method, models were able to choose the predictors that best describe the model performance
303
(Morisette et al., 2013; Talbert and Talbert, 2012, Dutta et al. 2022). Each model generated
Habitat Suitability modelling for Ibex
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
304
an estimate of the potential habitat suitability for every pixel, presented as continuous values
305
ranging from 0 to 1, which was subsequently interpreted to represent the probable habitat
306
suitability for a specific pixel in the current study area. The predicted habitat suitability was
307
determined through the binary maps, with consideration given to the minimal training
308
occurrence as a threshold (Hayes et al., 2015). We generated 3 ensembled maps as follows,
309
LISS IV derived LCLU and other topographic and radiometric variables at 5-meter spatial
310
resolution, Sentinel 2A derived LCLU and other topographic and radiometric variables at 10
311
meter spatial resolution, Integrated image derived LCLU and other topographic and
312
radiometric variables at 5 meter spatial resolution. The average of all of the binary
313
predictions (0 or 1) of the five models are used in the ensemble maps for each map pixel,
314
moreover the count surface depicts the model agreement, where 0 refers no model agreement
315
and 5 refers all model agreement for suitability estimation (Dutta et al. 2022).
316
For the purpose of comparing the performance of SDMs, numerous performance
317
metrics are frequently utilised. We employ area under the receiver operating characteristic
318
curve (AUC), Cohen's Kappa, True Skill Statistic (TSS), Proportion Correctly Classified
319
(PCC), sensitivity, and specificity to more clearly grasp comparative SDM performance
320
across the 5 model forms (Cohen, 1968; Allouche et al., 2006; Phillips 2010; Illán et al. 2010;
321
Jiménez-Valverde et al. 2013). Generally, AUC considered as a threshold-independent
322
statistic to evaluate models (Guisan and Zimmermann 2000; Brotons et al. 2004; Elith et al.
323
2006; Phillips et al. 2006; Pearson 2007; Grenouillet et al. 2011). We used the minimal
324
training presence threshold for metrics (specificity and sensitivity), which are threshold-
325
dependent (Peterson et al. 2011). The criterion for building ensemble model was based on
326
AUC threshold value of >0.75 of CV dataset. Variable importance was evaluated by the mean
327
AUC, which is the AUC values ratio and calculated by number of models runs for every
328
model.
329
4. Result:
330
4.1.
331
The outcome of the three classified images (LISS IV, Sentinel 2A and the integrated image)
332
quality assessment by different accuracy assessments have been discussed in the following
333
sections. The improvements in the integrated image have been examined by comparing the
334
integrated products with the LISS IV and Sentinel 2A images. The integrated image has a 5
335
metre spatial resolution like LISS IV, but it also improved spectral signature. We derived
336
nine LCLU classes of the study area from the three types of images with five different
337
supervised classification algorithms (Fig 2). 810 sample points (90 points for each class)
Classification of Land Class and Land Use
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
338
across the study area were obtained by the stratified random sampling approach to evaluate
339
the classification accuracy of the images. The study region consists of several LCLU features
340
such as sparse vegetation, scrub covers, juniper patches, road lines, water, agriculture field,
341
human settlements, barren land, permafrost area. Many of the aforementioned properties are
342
challenging to extract from LISS IV sensor data because of its fine resolution, which fail to
343
distinguish many classes at once. With the aid of integrated data, the classification result
344
increases and the features identified more accurately. As a result, fused data might be a cost-
345
effective alternative to better resolution multispectral data. The classification accuracy of
346
integrated classified images was compared to that of LISS IV and Sentinel 2A classified
347
images. Classification accuracy estimated by overall accuracy (OA), Kappa coefficients (κ)
348
Weighted Kappa, Producer’s accuracy (PA), User’s accuracy (UA), commission, omission
349
and F-statistics (F) (Table 2, 3, S1 – S4). (Congalton 1991).
350
classification accuracy not good by LISS IV image, no classifying algorithm perform well,
351
the highest overall accuracy achieved by SVM (OA = 60.61) and κ statistic calculated was
352
0.56, and the lowest performance by GMM algorithm (OA= 48.76). and κ statistic calculated
353
was 0.42, so the two classifying algorithms achieve moderate accuracy (Table 2, Fig S1 -S5)
354
(Landis and Koch 1977). Moreover, the other classier viz. ML, RF and GMM classifier
355
overall accuracy score was 57.40, 56.41, 48.76 and κ statistic score was 0.52, 0.51, 0.42
356
respectively that directed also moderate accuracy on the LISS IV image (Table 2, Fig S1 -
357
S5).
Result depicts that the
358
The accuracy estimation of the classification algorithm on Sentinel 2A imagery
359
evaluate that the RF model classify the area with highest overall accuracy i.e., 80.24 and κ
360
statistic 0.78 and the lowest performance by KNN algorithm i.e., 70.24 and κ statistic 0.67, so
361
the results show substantial agreement (Table 2, Fig 2, Fig S1 -S5). Other algorithms viz.
362
SVM, ML and GMM performing OA score 76.54, 73.08 and 71.35 respectively and the κ
363
statistic score was 0.73, 0.70, 0.68, respectively which is also substantial accurate as per
364
Landis and Koch (1977) (Table 2, Fig 2, Fig S1 -S5). Interestingly, the integrated image
365
shows higher classification accuracy compared to LISS IV and Sentinel 2A image (Table 2,
366
Fig 2, Fig S1 -S5). The highest overall accuracy produces by RF i.e., 86.17 and κ statistic
367
0.84 which is a perfect accuracy estimation (Table 2, Fig 2, Fig S1 -S5). KNN classifier also
368
perform great with the integrated image classification than the LISS IV and Sentinel 2A with
369
an overall accuracy value of 81.85 and κ statistic 0.80 and again it qualifies as perfect
370
classification (Table 2, Fig 2, Fig S1 -S5). The other classification algorithm such as the
371
SVM, ML and GMM predicted overall accuracy are 76.79, 71.72, 67.77 respectively and κ
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
372
statistic 0.74, 0.68, 0.64 respectively, which qualify as substantial classify (Table 2, Fig 2, Fig
373
S1 -S5). From this overall accuracy and κ statistic it is clear that integrated image
374
classification performs well than the single sensor images. However, the weighted κ also
375
depicts the same (Table 2, Fig S1 -S5).
376
However, the accuracy estimation for identifying each feature class from these images we
377
calculated producer’s accuracy, user’s accuracy F- statistics, omission and commission error
378
rates. Regarding to the comparison of each feature class, permafrost class differentiate by
379
each classification algorithms on every image perfectly. Omission error rates depicts that the
380
highest pixels misclassified by GMM algorithm on LISS IV image (Table 3, Table S1 -S4,
381
Fig 2, Fig S1 – S5). Road, settlement and sparse vegetation classes are poorly classified by
382
this algorithm on the LISS IV image (Table 3, Table S1 -S4, Fig 2, Fig S1 – S5). However,
383
the commission error rate directs that the road, highly misclassified by ML classification on
384
the LISS IV image (Table 3, Table S1 -S4, Fig 2, Fig S1 – S5). However, classification of
385
LISS IV image by ML, GMM, SVM poorly classify settlement, scrub, juniper patch, barren
386
and agriculture classes (Table 3, Table S1 -S4, Fig 2, Fig S1 – S5). The result of producer’s
387
accuracy shows that road, settlement and sparse vegetation, agriculture and water poor
388
accuracy in terms of classified by GMM algorithm on LISS IV image (Table 3, Table S1 -S4,
389
Fig 2, Fig S1 – S5). The user’s accuracy and F- statistics depicts that the road, settlement and
390
sparse vegetation, agriculture and water class gained highest accuracy with the integrated
391
image classified by RF algorithm (Table 3, Table S1 -S4, Fig 2, Fig S1 – S5). Furthermore,
392
the KNN classifier on integrated image and RF classifier on Sentinel 2A image having
393
overall accuracy score >= 80%, κ statistic and weighted κ score >=0.80, which tells perfect
394
object classification but the best classification by RF on the integrated image among all the
395
classification (Table 3, Table S1 -S4, Fig 2, Fig S1 – S5).
396
4.2.
397
After applying supervised classification on these satellite imageries and evaluate the accuracy
398
of image classification we used these LCLU to evaluate the potential habitat distribution of
399
Himalayan Ibex, in this study landscape. Habitat suitability of a species can address by its
400
environment where the species thrive. The LCLU one of the important aspects which tells
401
about the habitat of the species. Performing the species distribution model is crucial for
402
conservation of species in their natural habitat. Our model builds with 82 uncorrelated
403
occurrence records of Himalayan Ibex. Furthermore, we assigned the LCLU derived from the
404
best classified image of three different images (Sentinel 2A, LISS IV and Integrated image).
405
Only LCLU are not enough when the species resides in complex and rough terrain. So, we
Habit Suitability Ensemble model
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
406
choose topographical variables and soil, vegetation and water radiometric indices along with
407
the LCLU for knowing the suitable habitat of this species. Classified LCLU maps helps to
408
understand the habitat more precisely, fine resolution can provide more insights to the habitat
409
of the species. However, we selected 21 predictors at starting (Table S5), after the
410
multicollinearity test, we removed correlated variables and use only variables which follow
411
the Pearson’s rule i.e., <0.8 (Fig S6 – S8). The AUC values of five different modelling
412
algorithm ranged from 0.77-0.92 when habitat class use from LISS IV classified image, AUC
413
values ranged from 0.77-0.91 when habitat class use from Sentinel 2A classified image and
414
when integrated classified image derived habitat class use for the suitability prediction the
415
AUC value ranged from 0.77-0.92 for training data sets, which reflects great performance by
416
the models (Table 4, Fig 3). The model performance also evaluated by other metrices like
417
True skill statistic (TSS), Percent Correctly Classified (PCC), Cohen's Kappa, specificity and
418
sensitivity and depicts good performance by the models (Table 4, Fig S9 – S14). The
419
development of the final ensemble model and ensemble count maps was achieved as every
420
participation model satisfied the AUC requirement of 0.75 and above (Fig 3, 4). The LISS IV
421
derived LCLU with other variables show most suitable area about 63.80 km2, the Sentinel 2A
422
derived LCLU with other variables show most suitable area about 72.42 km2 and the
423
Integrated image derived LCLU with other variables show increase in most suitable area
424
about 78.42 km2 (Table S6, Fig 6, 7). In all instances when the rest of the model uses some of
425
the variables, RF and MaxEnt methods used all uncorrelated variables. Furthermore, for all of
426
the models with all of the combinations, the Soil Adjusted Vegetation Index (SAVI) with the
427
highest mean AUC, found as a major contributor and most significant variable among the
428
variables (Fig S15 – S32). Different LCLU shows significance differently in the three types
429
of images (Fig S15 – S32). In the LISS IV image the Barren is the leading variable which
430
influence in the model building (Fig S15, S18-22), moreover Juniper patch shows higher
431
significance among the LCLU while using the Sentinel 2A and Integrated image (Fig S16 -
432
17, S23-32). When the Himalayan Ibex occurrence is more prevalent, the response curve for
433
these variables has high peaks, and the likelihood values fall as the distance increases (Fig 18
434
- 32).
435
5. Discussion:
436
The theme of this research was to compare the object classification methods derived from
437
different sensor imagery and the preparation of an integrated image and analysis the
438
classification performance of it, for the improvement of LCLU analysis which play pivotal
439
role for species habitat configuration. Two different sensors namely LISS IV and Sentinel
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
440
2A, were considered for the preparation of the integration image. The images were classified
441
by 5 different type of supervised classification technique viz. ML, GMM, KNN, SVM and
442
RF. As a result, five LISS IV, five Sentinel 2A and five integrated images were produced
443
(Fig 2). In total fifteen images were compared with each other by qualitatively measures
444
(Table 2, 3, S1-S4, Fig S1-S5). Consequently, all of the images were subjected to visual
445
interpretation and comparison. Subsequently, the integrated images revealed discrete
446
outcomes that evince enhancements of varying magnitudes (Table 2, 3, S1-S4, Fig S1-S5).
447
The results of the qualitative evaluation indicate that RF produced the most advantageous
448
result on integrated images, while GMM produced the least favourable results on LISS IV
449
images (Table 2).
450
Classified images by integration techniques were investigated also by comparing the
451
classification results of original LISS IV and Sentinel 2A images for each selected class
452
(Table 2, 3, S1-S4, Fig S1-S5). The results demonstrate a clear improvement in classification
453
when utilising the RF method on the integrated image. However, KNN, SVM and ML of
454
Integrated showed great accuracy estimation for the image classification. Moreover, classifier
455
algorithms perform also well in the Sentinel 2A image after integrated image (Table 2, 3, S1-
456
S4, Fig S1-S5). In the case of LISS IV image, the classifier did not perform good, which
457
easily interpret by the accuracy estimation. Among the all models GMM classification on
458
LISS IV image was the worst (Table 2, 3, S1-S4, Fig S1-S5). In a nutshell it has been
459
determined that the utilisation of multi-sensor data significantly improves the precision of
460
LCLU classification, resulting in a more dependable and superior map generation. The
461
integration of two different satellite image result showed promising result to differentiate the
462
LCLU types. All the supervised classification methods used to classify the images discussed
463
in this paper are capable of properly classify multispectral images with good accuracy except
464
LISS IV. Furthermore, some model imperfectly classify settlement, road, water bodies which
465
account by visual interpretation (Fig 2).
466
The habitat suitability model of Himalaya Ibex, derived from the best classification
467
method of three images depicts there is a change in most suitable habitat (Table S6, Fig 4-5).
468
From this study we conclude that in this rugged terrain the species have complex association
469
with habitat which can not only defined by land classes, it also highly influenced by the
470
physical character of the terrain (Fig S15-S32). From the distribution model assessment, we
471
found that SAVI and elevation have major role which shape their habitat (Fig S15-S32). Area
472
with little vegetation cover, SAVI is used to adjust the NDVI for the impact of soil
473
brightness. So, the improperly classified image of LISS IV predict least suitable area
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
474
followed by Sentinel 2A classified image (Table S6, Fig 4-5). Furthermore, the integrated
475
image derived LCLU with the other topographic and radiometric variables predict an increase
476
in suitable area (Table S6, Fig 4-5). SDM with all image derived LCLU and other variables
477
predicts 34.76 km2 common (Table S6, Fig 4-5). Whereas, 41.50 km2 area common in LISS
478
IV and Sentinel 2A derived LCLU SDM model, 44.91 km2 area common in LISS IV and
479
integrated image derived LCLU SDM model and 54.14 km2 area common in Sentinel 2A and
480
integrated image derived LCLU SDM model (Table S6, Fig 4-5). Therefore, the study's
481
findings show that an integrated image can distinguish LCLU more effectively than a single
482
satellite image of the two images employed in this study. The present study evaluates the
483
performance of the SDM in conjunction with the three image-derived LCLU datasets,
484
incorporating both topographical and radiometric variables. The results indicate that the SDM
485
exhibits favourable performance across all LCLU datasets. Notably, the integrated image
486
dataset demonstrates a high level of precision in identifying LCLU, which may prove useful
487
in characterising the habitat configuration of the Himalayan Ibex.
488
5.1.
489
The current study deal with the comparison between the contribution of LISS IV, Sentinel
490
2A, and the integrated image of these aforementioned multisensor image with five
491
classification algorithms for the enhancement of LCLU analyses. After that, we select the
492
best classified image of these three different types of images for extraction of LCLU and
493
predict suitable habitat of Himalayan Ibex and compare which is the best among them.
494
For the image classification we employed five supervised classification algorithms, namely
495
ML, GMM, KNN, SVM and RF. Subsequently, statistical analysis and visual interpretation
496
were performed on each image. Due to the heterogeneous landscape's complexity, the classes
497
can occasionally be difficult to distinguish; nevertheless, the integrated image classified by
498
RF method significantly enhanced the mapping of each class. The F- statistics, Producers
499
accuracy, user's accuracy, commission and omission rate show that the classes classify with
500
low accuracy scores in LISS IV and Sentinel 2A images, however, employing integration to
501
improve their accuracy. In conclusion, the integrated image showed unique findings with
502
varying degrees of improvement. For both visual and qualitative analysis of the classified
503
images, RF demonstrated the best results on integrated image and GMM classification on
504
LISS IV was the worst. From the accuracy assessment the second-best classifier algorithm
505
was KNN on integrated image. However, it is evident that Sentinel 2A image also perform
506
good to classify the ground objects but no classifier algorithm could not classify LISS IV
507
image at a significant way.
Conclusion:
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
508
Furthermore, this investigation reinforces the efficacy of SDM ensemble modelling in
509
the prioritisation of conservation strategies and management. We recommended ensemble
510
model, against using a single modelling technique, particularly for species with complex
511
habitat. Due to the geomorphological complexities of this study area, only LCLU is not
512
sufficient to predict this species suitable habitat, so the topographic and radiometric variables
513
useful for this study. However, it can conclude that how the specific topographic factor with
514
LCLU supports the habitat of the species potential distribution in this study area, because the
515
landcover classes distributed in specific altitudinal gradients. The maximum models use
516
elevation and SAVI for prediction of suitable habitats in this area. However, we can infer
517
from this study that topographical characteristics may serve as a proxy for identifying
518
suitable habitat of mountainous species when LCLU classification is not fair, hence the
519
poorly classified image of LISS IV also shows fair suitable areas of the species.
520
Himalayan Ibex is a mountain ungulate which help to shape the ecosystem by many
521
dimensions. The snow leopard, the apex predator in this Trans Himalayan environment, relies
522
on this species of mountain ungulate as one of its primary food sources. (Suryawanshi et al.
523
2017, Sharief et al. 2022). Furthermore, the indigenous inhabitants of this region have
524
several kinds of traditional beliefs regarding Himalayan Ibex.
525
We therefore draw the conclusion from this study that all types of images can be
526
helpful for species distribution models in this type of complicated terrain when they are
527
associated with other crucial and influential variables like elevation, SAVI, slope, and aspect.
528
However, a precise description of LCLU can help to know the distribution of suitable habitat
529
in space more significantly, which helps to understand where the focus for conservation
530
needs to be placed for the long-term survival of wild animals. With this finding, we suggested
531
a community conservation area in the study landscape because it is clear that there is a
532
considerable amount of habitat for this wild mountain goats in this area, despite the fact that
533
there is no such protected area. It can also be advantageous for the inhabitants and the natural
534
population of Himalayan Ibex to safeguard this area from unforeseen infrastructures because
535
it can offer various ecosystem services, such as water, sustainable tourism, and opportunities
536
for outdoor recreation.
537
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
538
References:
539
540
541
542
543
544
Abburu, S. and Golla, S.B., 2015. Satellite image classification methods and techniques: A
review. International journal of computer applications, 119(8).
545
546
Anders Karlsson, 2003. Classification of high resolution satellite images, August 2003,
available at http://infoscience.epfl.ch/record/63248/files/TPD_Karlss on.pdf.
547
548
549
Anderson, R.P., Martínez-Meyer, E., Nakamura, M., Araújo, M.B., Peterson, A.T., Soberón,
J. and Pearson, R.G., 2011. Ecological niches and geographic distributions (MPB-49).
Princeton University Press.
550
551
Araújo, M.B. and New, M., 2007. Ensemble forecasting of species distributions. Trends in
ecology & evolution, 22(1), pp.42-47.
552
553
554
Bagchi, S. and Ritchie, M.E., 2010. Herbivore effects on above-and belowground plant
production and soil nitrogen availability in the Trans-Himalayan shrubsteppes. Oecologia, 164, pp.1075-1082.
555
556
557
Balcik, F.B. and Sertel, E., 2002. Wavelet-based image fusion of Landsat ETM images: a
case study for different landscape categories of Istanbul. ITU, Civil Engineering Faculty,
Istanbul, Turkey.
558
559
560
Berger, C., Riedel, F., Rosentreter, J., Stein, E., Hese, S. and Schmullius, C., 2015. Fusion of
airborne hyperspectral and LiDAR remote sensing data to study the thermal characteristics of
urban environments. Computational Approaches for Urban Environments, pp.273-292.
561
562
563
Besnard, A.G., La Jeunesse, I., Pays, O. and Secondi, J., 2013. Topographic wetness index
predicts the occurrence of bird species in floodplains. Diversity and Distributions, 19(8),
pp.955-963.
564
565
Booth, T.H., Nix, H.A., Hutchinson, M.F. and Jovanic, T., 1988. Niche analysis and tree
species introduction. Forest Ecology and Management, 23(1), pp.47-59.
566
Breiman, L., 2001. Random forests machine learning, vol. 45. pp. 5–32.
567
568
569
Bremner, D., Demaine, E., Erickson, J., Iacono, J., Langerman, S., Morin, P. and Toussaint,
G., 2005. Output-sensitive algorithms for computing nearest-neighbour decision
boundaries. Discrete & Computational Geometry, 33, pp.593-604.
570
571
572
Brodu, N., 2017. Super-resolving multiresolution images with band-independent geometry of
multispectral pixels. IEEE Transactions on Geoscience and Remote Sensing, 55(8), pp.46104617.
573
574
575
Brotons, L., Thuiller, W., Araújo, M.B. and Hirzel, A.H., 2004. Presenceabsence versus
presenceonly modelling methods for predicting bird habitat suitability. Ecography, 27(4),
pp.437-448.
Allouche, O., Tsoar, A. and Kadmon, R., 2006. Assessing the accuracy of species distribution
models: prevalence, kappa and the true skill statistic (TSS). Journal of applied
ecology, 43(6), pp.1223-1232.
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
576
577
Brown L. G. 1992. A survey of image Co-registration Techniques. ACM Computing surveys,
24: 325-376.
578
579
580
581
582
Buisson, L., Thuiller, W., Casajus, N., Lek, S. and Grenouillet, G., 2010. Uncertainty in
ensemble forecasting of species distribution. Global Change Biology, 16(4), pp.1145-1157.
Vaiphasa, C., Piamduaytham, S., Vaiphasa, T. and Skidmore, A.K., 2011. A Normalized
Difference Vegetation index (NDVI) Time-series of idle agriculture lands: A preliminary
study. Engineering Journal, 15(1), pp.9-16.
583
584
Burkhard, B., Kroll, F., Nedkov, S. and Müller, F., 2012. Mapping ecosystem service supply,
demand and budgets. Ecological indicators, 21, pp.17-29.
585
586
Büyüksalih, İ., 2016. Landsat images classification and change analysis of land cover/use in
Istanbul. International Journal of Environment and Geoinformatics, 3(2), pp.56-65.
587
588
589
590
Camargo, F.F., Sano, E.E., Almeida, C.M., Mura, J.C. and Almeida, T., 2019. A comparative
assessment of machine-learning techniques for land use and land cover classification of the
Brazilian tropical savanna using ALOS-2/PALSAR-2 polarimetric images. Remote
Sensing, 11(13), p.1600.
591
592
593
Chellappa, R., Veeraraghavan, A., Ramanathan, N., Yam, C.Y., Nixon, M.S., Elgammal, A.,
Boyd, J.E., Little, J.J., Lynnerup, N., Larsen, P.K. and Reynolds, D., 2009. Gaussian mixture
models. Encyclopedia of Biometrics, 2009(2), pp.659-663.
594
595
Cheng, G., Han, J. and Lu, X., 2017. Remote sensing image scene classification: Benchmark
and state of the art. Proceedings of the IEEE, 105(10), pp.1865-1883.
596
597
598
Chitade, A.Z. and Katiyar, S.K., 2012. Multiresolution and multispectral data fusion using
discrete wavelet transform with IRS images: Cartosat-1, IRS LISS III and LISS IV. Journal
of the Indian Society of Remote Sensing, 40, pp.121-128.
599
600
601
602
Clark, E.L., Munkhbat, J., Dulamtseren, S., Baillie, J.E.M., Batsaikhan, N., King, S.R.B.,
Samiya, R. and Stubbe, M., 2006. Summary Conservation Action Plans for Mongolian
Mammals. Regional Red List Series Vol. 2. Zoological Society of London, London. English
and Mongolian, 165.
603
604
Cohen, J., 1968. Weighted kappa: nominal scale agreement provision for scaled disagreement
or partial credit. Psychological bulletin, 70(4), p.213.
605
606
607
Colditz, R.R., Schmidt, M., Conrad, C., Hansen, M.C. and Dech, S., 2011. Land cover
classification with coarse spatial resolution data to derive continuous and discrete maps for
complex regions. Remote Sensing of Environment, 115(12), pp.3264-3275.
608
609
Congalton, R.G. and Green, K., 2019. Assessing the accuracy of remotely sensed data:
principles and practices. CRC press.
610
611
Congalton, R.G., 1991. A review of assessing the accuracy of classifications of remotely
sensed data. Remote sensing of environment, 37(1), pp.35-46.
612
613
614
Congalton, R.G., Oderwald, R.G. and Mead, R.A., 1983. Assessing Landsat classification
accuracy using discrete multivariate analysis statistical techniques. Photogrammetric
engineering and remote sensing, 49(12), pp.1671-1678.
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
615
616
617
Congedo, L., 2021. Semi-Automatic Classification Plugin: A Python tool for the download
and processing of remote sensing images in QGIS. Journal of Open Source Software, 6(64),
p.3172.
618
619
620
de Sherbinin, A., 2002. Land-use and land-cover change, a CIESIN thematic guide. Center
for International Earth Science Information Network, Columbia University, Palisades, NY.
pp. 67.
621
622
623
Dobson, M.C., Ulaby, F.T. and Pierce, L.E., 1995. Land-cover classification and estimation
of terrain attributes using synthetic aperture radar. Remote sensing of Environment, 51(1),
pp.199-214.
624
625
626
627
Drusch, M., Del Bello, U., Carlier, S., Colin, O., Fernandez, V., Gascon, F., Hoersch, B.,
Isola, C., Laberinti, P., Martimort, P. and Meygret, A., 2012. Sentinel-2: ESA's optical highresolution mission for GMES operational services. Remote sensing of Environment, 120,
pp.25-36.
628
629
630
631
Dutta, R., Mukherjee, T., Sharief, A., Singh, H., Kumar, V., Joshi, B.D., Banerjee, D.,
Thakur, M. and Sharma, L.K., 2022. Climate change may plunder the facultative top predator
Yellow-throated Martin from the Hindu-Kush Himalayan Region. Ecological Informatics, 69,
p.101622.
632
633
634
Dzięciołowski, R., Krupka, J., Bajandelger, X. and Dziedzic, R., 1980. Argali and Siberian
ibex populations in the Khuhsyrh Reserve in Mongolian Altai. Acta Theriologica, 25(16),
pp.213-219.
635
636
637
Elith*, J., H. Graham*, C., P. Anderson, R., Dudík, M., Ferrier, S., Guisan, A., J. Hijmans,
R., Huettmann, F., R. Leathwick, J., Lehmann, A. and Li, J., 2006. Novel methods improve
prediction of species’ distributions from occurrence data. Ecography, 29(2), pp.129-151.
638
639
640
Elith, J. and Graham, C.H., 2009. Do they? How do they? WHY do they differ? On finding
reasons for differing performances of species distribution models. Ecography, 32(1), pp.6677.
641
ESRI. (2018). ArcGIS Desktop: Release 10.6. Redlands, CA: Environmental Systems
642
Research Institute.
643
644
Fedosenko, A.K. and Blank, D.A., 2001. Capra sibirica. Mammalian species, 2001(675),
pp.1-13.
645
646
647
648
Fonteh, M.L., Theophile, F., Cornelius, M.L., Main, R., Ramoelo, A. and Cho, M.A., 2016.
Assessing the utility of sentinel-1 c band synthetic aperture radar imagery for land use land
cover classification in a tropical coastal systems when compared with landsat 8. Journal of
Geographic Information System, 8(4), pp.495-505.
649
650
Foody, G.M., 2002. Status of land cover classification accuracy assessment. Remote sensing
of environment, 80(1), pp.185-201.
651
652
Fox, J.L., Nurbu, C. and Chundawat, R.S., 1991. The mountain ungulates of Ladakh,
India. Biological Conservation, 58(2), pp.167-190.
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
653
654
Franklin, J., 2010. Mapping species distributions: spatial inference and prediction.
Cambridge University Press.
655
656
657
658
Fry, J.A., Xian, G., Jin, S.M., Dewitz, J.A., Homer, C.G., Yang, L.M., Barnes, C.A., Herold,
N.D. and Wickham, J.D., 2011. Completion of the 2006 national land cover database for the
conterminous United States. PE&RS, Photogrammetric Engineering & Remote
Sensing, 77(9), pp.858-864.
659
660
661
Garcia, R.A., Burgess, N.D., Cabeza, M., Rahbek, C. and Araújo, M.B., 2012. Exploring
consensus in 21st century projections of climatically suitable areas for African
vertebrates. Global Change Biology, 18(4), pp.1253-1269.
662
663
664
665
Gebhardt, S., Wehrmann, T., Ruiz, M.A.M., Maeda, P., Bishop, J., Schramm, M., Kopeinig,
R., Cartus, O., Kellndorfer, J., Ressl, R. and Santos, L.A., 2014. MAD-MEX: Automatic
wall-to-wall land cover monitoring for the Mexican REDD-MRV program using all Landsat
data. Remote Sensing, 6(5), pp.3923-3943.
666
667
668
Grenouillet, G., Buisson, L., Casajus, N. and Lek, S., 2011. Ensemble modelling of species
distribution: the effects of geographical and environmental ranges. Ecography, 34(1), pp.917.
669
670
671
672
Gu, J., Chen, J., Zhou, Q. and Zhang, H., 2007. Gaussian mixture model of texture for
extracting residential area from high-resolution remotely sensed imagery. In ISPRS Workshop
on Updating Geo-spatial Databases with Imagery & The 5th ISPRS Workshop on DMGISs,
China, Agustus (pp. 157-162).
673
674
675
Guidici, D. and Clark, M.L., 2017. One-Dimensional convolutional neural network landcover classification of multi-seasonal hyperspectral imagery in the San Francisco Bay Area,
California. Remote Sensing, 9(6), p.629.
676
677
Guisan, A. and Thuiller, W., 2005. Predicting species distribution: offering more than simple
habitat models. Ecology letters, 8(9), pp.993-1009.
678
679
Guisan, A. and Zimmermann, N.E., 2000. Predictive habitat distribution models in
ecology. Ecological modelling, 135(2-3), pp.147-186.
680
681
682
Gumma, M.K., Thenkabail, P.S., Teluguntla, P. and Whitbread, A.M., 2019. Indo-Ganges
river basin land use/land cover (LULC) and irrigated area mapping. In Indus River Basin (pp.
203-228). Elsevier, Amsterdam, Netherlands.
683
684
685
GÜNLÜ, A., 2021. Comparison of Different Classification Approaches for Land Cover
Classification using Multispectral and Fusion Satellite Data: A Case Study in Ören Forest
Planning Unit. Bartın Orman Fakültesi Dergisi, 23(1), pp.306-322.
686
687
688
Hayes, M.A., Cryan, P.M. and Wunder, M.B., 2015. Seasonally-dynamic presence-only
species distribution models for a cryptic migratory bat impacted by wind energy
development. PLoS One, 10(7), p.e0132599.
689
690
691
Huang, X. and Zhang, L., 2012. An SVM ensemble approach combining spectral, structural,
and semantic features for the classification of high-resolution remotely sensed imagery. IEEE
transactions on geoscience and remote sensing, 51(1), pp.257-272.
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
692
693
694
695
Hussain, S., Mubeen, M., Ahmad, A., Akram, W., Hammad, H.M., Ali, M., Masood, N.,
Amin, A., Farid, H.U., Sultana, S.R. and Fahad, S., 2020. Using GIS tools to detect the land
use/land cover changes during forty years in Lodhran District of Pakistan. Environmental
Science and Pollution Research, 27, pp.39676-39692.
696
697
698
Hütt, C., Koppe, W., Miao, Y. and Bareth, G., 2016. Best accuracy land use/land cover
(LULC) classification to derive crop types using multitemporal, multisensor, and multipolarization SAR satellite images. Remote sensing, 8(8), p.684.
699
700
701
Illán, J.G., Gutiérrez, D. and Wilson, R.J., 2010. The contributions of topoclimate and land
cover to species distributions and abundance: fineresolution tests for a mountain butterfly
fauna. Global Ecology and Biogeography, 19(2), pp.159-173.
702
703
Ishwaran, H. and Kogalur, U.B., 2007. Random survival forests for R. R news, 7(2), pp.2531.
704
705
Ishwaran, H., Kogalur, U.B., Blackstone, E.H. and Lauer, M.S., Random Survival Forests.
The annals of applied statistics. 2008; 2 (3): 841–860.
706
707
708
Jia, K., Wu, B., Tian, Y., Zeng, Y. and Li, Q., 2011. Vegetation classification method with
biochemical composition estimated from remote sensing data. International journal of remote
sensing, 32(24), pp.9307-9325.
709
710
711
JiménezValverde, A., Acevedo, P., Barbosa, A.M., Lobo, J.M. and Real, R., 2013.
Discrimination capacity in species distribution models depends on the representativeness of
the environmental domain. Global Ecology and Biogeography, 22(4), pp.508-516.
712
713
714
Juliev, M., Pulatov, A., Fuchs, S. and Hübl, J., 2019. Analysis of Land Use Land Cover
Change Detection of Bostanlik District, Uzbekistan. Polish Journal of Environmental
Studies, 28(5), 3235-3242.
715
716
Karasiak,
N. 2019. Lennepkade/dzetsaka:
Dzetsaka
Zenodo. http://doi.org/10.5281/zenodo.2647723.
717
718
719
Keshtkar, H., Voigt, W. and Alizadeh, E., 2017. Land-cover classification and analysis of
change using machine-learning classifiers and multi-temporal remote sensing
imagery. Arabian Journal of Geosciences, 10, pp.1-15.
720
721
722
723
Khan, G., Khan, B., Qamer, F.M., Abbas, S., Khan, A. and Xi, C., 2016. Himalayan ibex
(Capra ibex sibirica) habitat suitability and range resource dynamics in the Central
Karakorum National Park, Pakistan. Journal of King Saud University-Science, 28(3), pp.245254.
724
725
726
727
Khatami, R., Mountrakis, G. and Stehman, S.V., 2016. A meta-analysis of remote sensing
research on supervised pixel-based land-cover image classification processes: General
guidelines for practitioners and future research. Remote Sensing of Environment, 177, pp.89100.
728
729
730
Lakshmi, T.H., Madhu, T., Rao, E.K. and Mounica, V.L., 2015. Satellite image resolution
enhancement using discrete wavelet transform and gaussian mixture model. International
Research Journal of Engineering and Technology (IRJET), 2(04).
v3.4.4
(Version
v3.4.4).
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
731
732
Landis, J.R. and Koch, G.G., 1977. The Measurement of Observer Agreement for Categorical
Data. Biometrics, 33(1), 159–174.
733
734
Liaw, A. and Wiener, M., 2002. Classification and regression by randomForest. R news, 2(3),
pp.18-22.
735
736
Lillesand, T. & Kiefer, R. 1994. Remote sensing and Image Interpretation. Third edition.
John Wiley & Sons: New York, USA.
737
738
739
Lin, W., Liao, X., Deng, J. and Liu, Y., 2016. Land cover classification of RADARSAT-2
SAR data using convolutional neural network. Wuhan University Journal of Natural
Sciences, 21(2), pp.151-158.
740
741
742
Lu, D. and Weng, Q., 2007. A survey of image classification methods and techniques for
improving classification performance. International journal of Remote sensing, 28(5),
pp.823-870.
743
744
745
Luo, S., Wang, C., Xi, X., Zeng, H., Li, D., Xia, S. and Wang, P., 2015. Fusion of airborne
discrete-return LiDAR and hyperspectral data for land cover classification. Remote
Sensing, 8(1), p.3–12.
746
747
748
Luo, X., Zhang, Z. and Wu, X., 2016. A novel algorithm of remote sensing image fusion
based on shift-invariant Shearlet transform and regional selection. AEU-International Journal
of Electronics and Communications, 70(2), pp.186-197.
749
750
Ma, J., Ma, Y. and Li, C., 2019. Infrared and visible image fusion methods and applications:
A survey. Information Fusion, 45, pp.153-178.
751
752
753
Marmion, M., Luoto, M., Heikkinen, R.K. and Thuiller, W., 2009. The performance of stateof-the-art modelling techniques depends on geographical distribution of species. Ecological
Modelling, 220(24), pp.3512-3520.
754
755
756
McNaughton, S.J., 1979. Grassland–herbivore dynamics. . In “Serengeti. Dynamics of an
Ecosystem” (ARE Sinclair and M. Norton-Griffiths, Eds.). Chicago University Press,
Chicago, pp 46–81
757
758
759
Melgani, F. and Bruzzone, L., 2004. Classification of hyperspectral remote sensing images
with support vector machines. IEEE Transactions on geoscience and remote sensing, 42(8),
pp.1778-1790.
760
761
762
Miettinen, J. and Liew, S.C., 2011. Separability of insular Southeast Asian woody plantation
species in the 50 m resolution ALOS PALSAR mosaic product. Remote Sensing Letters, 2(4),
pp.299-307.
763
764
765
766
Mohajane, M., Essahlaoui, A., Oudija, F., Hafyani, M.E., Hmaidi, A.E., Ouali, A.E.,
Randazzo, G. and Teodoro, A.C., 2018. Land use/land cover (LULC) using landsat data
series (MSS, TM, ETM+ and OLI) in Azrou Forest, in the Central Middle Atlas of
Morocco. Environments, 5(12), p.131.
767
768
769
Morisette, J.T., Jarnevich, C.S., Holcombe, T.R., Talbert, C.B., Ignizio, D., Talbert, M.K.,
Silva, C., Koop, D., Swanson, A. and Young, N.E., 2013. VisTrails SAHM: visualization and
workflow management for species habitat modeling. Ecography, 36(2), pp.129-135.
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
770
771
772
Morris, D.W., Dupuch, A. and Halliday, W.D., 2012. Climate-induced habitat selection
predicts future evolutionary strategies of lemmings. Evolutionary Ecology Research, 14(6),
pp.689-705.
773
774
775
776
Nitze, I., Schulthess, U. and Asche, H., 2012. Comparison of machine learning algorithms
random forest, artificial neural network and support vector machine to maximum likelihood
for supervised crop type classification. Proceedings of the 4th GEOBIA, Rio de Janeiro,
Brazil, 79, p.35-40.
777
778
Nurwauziyah, I., UD, S., Putra, I.G.B. and Firdaus, M.I., 2018. Satellite image classification
using Decision Tree, SVM and k-Nearest Neighbor. no. July.
779
780
Okwuashi, O., Eyo, E. and Eyoh, A., 2011. Supervised Gaussian mixture model based remote
sensing image classification. Global Journal of Environmental Sciences, 10(1&2), pp.57-65.
781
782
Pal, M. and Foody, G.M., 2010. Feature selection for classification of hyperspectral data by
SVM. IEEE Transactions on Geoscience and Remote Sensing, 48(5), pp.2297-2307.
783
784
785
Patenaude, G., Milne, R. and Dawson, T.P., 2005. Synthesis of remote sensing approaches
for forest carbon estimation: reporting to the Kyoto Protocol. Environmental Science &
Policy, 8(2), pp.161-178.
786
787
Pearson, R.G., 2007. Species’ distribution modeling for conservation educators and
practitioners. Synthesis. American Museum of Natural History, 50, pp.54-89.
788
789
790
Pearson, R.G., Thuiller, W., Araújo, M.B., MartinezMeyer, E., Brotons, L., McClean, C.,
Miles, L., Segurado, P., Dawson, T.P. and Lees, D.C., 2006. Modelbased uncertainty in
species range prediction. Journal of biogeography, 33(10), pp.1704-1711.
791
792
Phillips, S.J. and Elith, J., 2010. POC plots: calibrating species distribution models with
presenceonly data. Ecology, 91(8), pp.2476-2484.
793
794
Phillips, S.J., Anderson, R.P. and Schapire, R.E., 2006. Maximum entropy modeling of
species geographic distributions. Ecological modelling, 190(3-4), pp.231-259.
795
796
Reading, R., Michel, S., Suryawanshi, K. & Bhatnagar, Y.V. 2020. Capra sibirica. The IUCN
Red List of Threatened Species 2020: e.T42398A22148720
797
798
799
800
Rodriguez-Galiano, V.F., Chica-Olmo, M., Abarca-Hernandez, F., Atkinson, P.M. and
Jeganathan, C., 2012b. Random Forest classification of Mediterranean land cover using
multi-seasonal imagery and multi-seasonal texture. Remote Sensing of Environment, 121,
pp.93-107.
801
802
803
804
Rodriguez-Galiano, V.F., Ghimire, B., Pardo-Iguzquiza, E., Chica-Olmo, M. and Congalton,
R.G., 2012a. Incorporating the downscaled Landsat TM thermal band in land-cover
classification using random forest. Photogrammetric Engineering & Remote Sensing, 78(2),
pp.129-137.
805
806
807
Rosenqvist, Å., Milne, A., Lucas, R., Imhoff, M. and Dobson, C., 2003. A review of remote
sensing technology in support of the Kyoto Protocol. Environmental Science & Policy, 6(5),
pp.441-455.
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
808
809
810
811
Saini, R., Sharma, M.C., Deswal, S., Barr, I.D., Kumar, P., Kumar, P., Kumar, P. and Chopra,
S., 2019. Glacio-archaeological evidence of permanent settlements within a glacier end
moraine complex during 980-1840 AD: The Miyar Basin, Lahaul Himalaya,
India. Anthropocene, 26, p.100197.
812
Scheffler, D., Hollstein, A., Diedrich, H., Segl, K. and Hostert, P., 2017. AROSICS: An
813
automated and robust open-source image co-registration software for multi-sensor satellite
814
data. Remote sensing, 9(7), p.676.
815
816
817
Shackleton, D.M., 1997. Wild sheep and goats and their relatives: Status survey and
conservation action plan. IUCN/SSC Caprinae Specialist Group. IUCN, Switzerland and
Cambridge, UK.
818
819
820
Shahbaz, M., Guergachi, A., Noreen, A. and Shaheen, M., 2012. Classification by object
recognition in satellite images by using data mining. In Proceedings of the World Congress
on Engineering (Vol. 1, pp. 4-6) London, U.K.
821
822
823
824
Sharief, A., Kumar, V., Singh, H., Mukherjee, T., Dutta, R., Joshi, B.D., Bhattacharjee, S.,
Ramesh, C., Chandra, K., Thakur, M. and Sharma, L.K., 2022. Landscape use and cooccurrence pattern of snow leopard (Panthera uncia) and its prey species in the fragile
ecosystem of Spiti Valley, Himachal Pradesh. Plos one, 17(7), p.e0271556.
825
826
827
Sirro, L., Häme, T., Rauste, Y., Kilpi, J., Hämäläinen, J., Gunia, K., De Jong, B. and Paz
Pellat, F., 2018. Potential of different optical and SAR data in forest and land cover
classification to support REDD+ MRV. Remote Sensing, 10(6), p.942.
828
829
830
Sonobe, R., Tani, H., Wang, X., Kobayashi, N. and Shimamura, H., 2014. Random forest
classification of crop type using multi-temporal TerraSAR-X dual-polarimetric data. Remote
Sensing Letters, 5(2), pp.157-164.
831
832
833
Sonobe, R., Yamaya, Y., Tani, H., Wang, X., Kobayashi, N. and Mochizuki, K.I., 2017.
Mapping crop cover using multi-temporal Landsat 8 OLI imagery. International Journal of
Remote Sensing, 38(15), pp.4348-4361.
834
835
836
Soria-Ruiz, J., Fernandez-Ordonez, Y. and Woodhouse, I.H., 2010. Land-cover classification
using radar and optical images: a case study in Central Mexico. International Journal of
Remote Sensing, 31(12), pp.3291-3305.
837
838
839
Srivastava, P.K., Han, D., Rico-Ramirez, M.A., Bray, M. and Islam, T., 2012. Selection of
classification techniques for land use/land cover change investigation. Advances in Space
Research, 50(9), pp.1250-1265.
840
841
842
Suryawanshi, K.R., Redpath, S.M., Bhatnagar, Y.V., Ramakrishnan, U., Chaturvedi, V.,
Smout, S.C. and Mishra, C., 2017. Impact of wild prey availability on livestock predation by
snow leopards. Royal Society Open Science, 4(6), p.170026.
843
844
845
Szuster, B.W., Chen, Q. and Borger, M., 2011. A comparison of classification techniques to
support land cover and land use analysis in tropical coastal zones. Applied Geography, 31(2),
pp.525-532.
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
846
847
Talbert, C.B. and Talbert, M.K., 2012. User manual for SAHM package for VisTrails. Fort
Collins: US Geological Survey, Fort Collins Science Center.
848
849
850
Thanh Noi, P. and Kappas, M., 2017. Comparison of random forest, k-nearest neighbor, and
support vector machine classifiers for land cover classification using Sentinel-2
imagery. Sensors, 18(1), p.18.
851
852
Thuiller, W., Lafourcade, B., Engler, R. and Araújo, M.B., 2009. BIOMOD–a platform for
ensemble forecasting of species distributions. Ecography, 32(3), pp.369-373.
853
854
855
856
Topaloğlu, R.H., Sertel, E. and Musaoğlu, N., 2016. Assessment of classification accuracies
of sentinel-2 and landsat-8 data for land cover/use mapping. In: International archives of the
photogrammetry, remote sensing & spatial Information Sciences, 41. ISPRS: Prague, Czech
Republic, Volume XLI-B8, pp. 1055–1059.
857
858
Turner, B.L., Skole, D., Sanderson, S., Fischer, G., Fresco, L. and Leemans, R., 1995. Landuse and land-cover change: science/research plan.
859
860
861
Vohra, R. and Tiwari, K.C., 2020. Comparative analysis of SVM and ANN classifiers using
multilevel fusion of multi-sensor data in urban land classification. Sensing and Imaging, 21,
pp.1-21.
862
863
Wang, Z.J., Li, D.R. and Li, Q.Q., 2001. Application of multiple wavelet theory in SPOT and
TM image fusion. Journal of Wuhan University (Information Science Edition), 1, pp.24-28.
864
865
866
Wenbo, W., Jing, Y. and Tingjun, K., 2008. Study of remote sensing image fusion and its
application in image classification. The international archives of the photogrammetry,
Remote Sensing and Spatial Information Sciences, 37(B7), pp.1141-1146.
867
868
869
Weng, Q., 1999. Environmental impacts of land use and land cover change in the Zhujiang
Delta, China: an analysis using an integrated GIS, remote sensing, and spatial modeling
approach. Ph.D. dissertation, University of Georgia, Athens, GA.
870
871
Wulder, M.A., Coops, N.C., Roy, D.P., White, J.C. and Hermosilla, T., 2018. Land cover
2.0. International Journal of Remote Sensing, 39(12), pp.4254-4284.
872
873
Yin, L., Yang, P., Mao, K. and Liu, Q., 2021. Remote sensing image scene classification
based on fusion method. Journal of Sensors, 2021, pp.1-14.
874
875
Zafari, A., Zurita-Milla, R. and Izquierdo-Verdiguier, E., 2019. Evaluating the performance
of a random forest kernel for land cover classification. Remote sensing, 11(5), p.575.
876
877
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
878
Table 1: Data type and acquisition details for the two satellite imagery.
S.No Satellite/Sensor Product Sensing Row Path Date
type
.
orbit
of Resoluti Tile No /
acquisition
number
Sentinel 2A
L1C
105
----
----
2021-10-06
on
Product ID
(meter)
(LISS IV)
10, 20, T43SFS
60
Sentinel 2A
L1C
105
----
----
2021-10-06
10, 20, T43SGS
60
LISS IV
879
----
----
048
095
2020-09-02
5
221881311
Table 2: Accuracy assessment of LISS IV, Sentinel 2A and Integrated images classified by five supervised classifiers namely, ML (Maximum
Likelihood), GMM (Guassian mixed model), KNN (K-nearest neighbourhood), SVM (Support Vector machine) and RF (Random Forest). The
evaluation metrices are Overall accuracy, Kappa (κ) statistic, SE of Kappa (κ) statistic and Weighted Kappa (κ).
883
SE of
Weighted
Kappa
Kappa
Overall
Kappa
( )
( )
accuracy
( )
statistic
statistic
Accuracy Assessment
κ
κ
κ
884
885
Image type
LISS IV
Sentinel 2A
Integrated
LISS IV
Sentinel 2A
Integrated
LISS IV
Sentinel 2A
Integrated
LISS IV
Sentinel 2A
Integrated
ML
GMM
57.41
73.08
71.73
0.52
0.70
0.68
0.02
0.02
0.02
0.51
0.71
0.71
KNN
48.77
71.35
67.78
0.42
0.68
0.64
0.02
0.02
0.02
0.41
0.65
0.61
SVM
55.31
70.24
81.85
0.50
0.67
0.80
0.02
0.02
0.02
0.46
0.60
0.80
60.62
76.54
76.79
0.56
0.74
0.74
0.02
0.02
0.02
0.57
0.72
0.74
RF
56.42
80.24
86.17
0.51
0.78
0.84
0.02
0.02
0.01
0.50
0.74
0.85
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
880
881
882
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
Table 3: Class accuracy metrices on best classified image i.e., Random Forest classifies the
Integrated image. The class accuracy metrices evaluated by Omission rate, Commission rate,
Producer’s accuracy, User’s accuracy and F – stat.
Agriculture
Sparse
vegetation
Barren
Scrub
Juniper patch
Settlement
Permafrost
Water
Road
Producer's Accuracy (PA)
LCLU Class
Agriculture
Sparse
vegetation
Barren
Scrub
Juniper patch
Settlement
Permafrost
Water
Road
F- statistics
LCLU Class
889
Agriculture
Sparse
vegetation
Barren
Scrub
Juniper patch
Settlement
Permafrost
Water
Road
LISSIV
47
Sentinel
Integrated
2A
23
14
57
13
42
36
70
2
34
91
LISSIV
53
26
20
2
1
7
18
17
13
34
23
0
0
11
3
58
31
Sentinel
Integrated
2A
77
86
43
87
58
64
30
98
66
9
LISSIV
53
74
80
98
99
93
82
83
87
66
77
100
100
89
97
42
69
Sentinel
Integrated
2A
75
85
50
60
49
50
44
88
76
16
78
76
86
74
78
99
94
57
84
80
80
85
85
100
98
80
LCLU Class
Commission
Omission
LCLU Class
Agriculture
Sparse
vegetation
Barren
Scrub
Juniper patch
Settlement
Permafrost
Water
Road
LCLU Class
User's accuracy (UA)
886
887
888
Agriculture
Sparse
vegetation
Barren
Scrub
Juniper patch
Settlement
Permafrost
Water
Road
LISSIV
47
Sentinel
Integrated
2A
27
16
40
55
57
59
16
21
9
33
LISSIV
53
17
11
38
33
20
22
33
17
3
4
1
0
1
1
14
5
Sentinel
Integrated
2A
73
84
60
45
43
41
84
79
91
67
83
62
80
67
97
99
99
86
89
67
78
83
96
100
99
95
Table 4: Evaluation metrics to evaluate the efficiency of the participating distribution models for Himalayan Ibex in the study landscape.
Participating models are BRT (Boosted Regression Tree), GLM (Generalized Linear Model), MARS (Multivariate adaptive regression splines),
MAXENT (Maximum Entropy Model), RF (Random Forest) and the efficiency of the models evaluated by AUC (area under the receiver
operator curve), PCC (Proportion Correctly Classified), sensitivity, specificity, Cohen's kappa and TSS (True Skill Statistic). CV mean (Cross
Validation) data used for model evaluation and train split used for the model, which was assessed by model evaluation. (a) LISS IV derived
LCLU used for this model building, (b) Sentinel 2A derived LCLU used for this model building and (c) Integrated image derived LCLU used for
this model building.
897
(a)
Model
AUC
PCC
Sensitivity
Specificity
Kappa
TSS
BRT
0.92
84.67
0.85
0.84
0.69
0.69
GLM
0.78
68.1
0.65
0.71
0.36
0.36
MARS
0.9
80.37
0.8
0.8
0.61
0.61
MAXENT
0.9
83.95
0.84
0.84
0.68
0.68
RF
0.77
70.55
0.7
0.71
0.41
0.41
898
899
(b)
Model
900
901
(c)
AUC
PCC
Sensitivity
Specificity
Kappa
TSS
BRT
0.91
82.32
0.8
0.84
0.65
0.65
GLM
0.81
74.39
0.74
0.74
0.49
0.49
MARS
0.82
75
0.74
0.76
0.5
0.5
MAXENT
0.91
85.28
0.85
0.85
0.71
0.71
RF
0.77
70.73
0.71
0.71
0.41
0.41
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
890
891
892
893
894
895
896
902
0.4
0.4
0.71
0.69
0.77
RF
69.94
0.74
0.74
0.85
0.89
0.92
MAXENT
87.04
0.63
0.63
0.82
0.81
0.87
MARS
81.6
0.4
0.4
0.71
0.69
0.8
GLM
69.94
0.57
0.57
0.79
0.78
78.53
0.89
BRT
Model
AUC
PCC
Sensitivity
Specificity
Kappa
TSS
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
903
904
905
906
Figure 1: Depiction of present study area along with the its location in India.
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
907
908
909
910
Figure 2: Classified images from LISS IV, Sentinel 2A and Integrated images by ML, GMM,
KNN, SVM and RF classifier algorithms.
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
911
912
913
Figure 3: AUC plots of five algorithms on different sourced images to predict species
distribution model.
bioRxiv preprint doi: https://doi.org/10.1101/2023.07.18.549476; this version posted July 19, 2023. The copyright holder for this preprint (which
was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
914
915
916
Figure 4: Probability (model agreement maps) maps and binary maps from three different
sourced images generated LCLU to predict species distribution model of Himalayan Ibex.
917
918
919
920
Figure 5: Variance in suitable habitat prediction of Himalayan Ibex by three different sourced
images generated LCLU.