Tracking Forest Loss and Fragmentation During 1930–2020 in Asian Elephant (Elephas Maximus) Habitats in Nepal

Forest cover is the primary determinant of elephant distribution, thus, understanding forest loss and fragmentation is crucial for elephant conservation. We assessed deforestation and patterns of forest fragmentation during 1930–2020 in Chure Terai Madhesh Lanscape (CTML) which covers the entire elephant range in Nepal. Forest cover maps and fragmentation matrices were generated using multi-source data (Topographic maps and Landsat images of 1930, 1975, 2000, and 2020) and spatiotemporal changes was quantied. Forest cover within the elephant range was 19,069 km 2 . Overall, 21.5% of elephant habitat was lost between 1930 to 2020, with a larger (12.3%) forest cover loss between 1930 & 1975. Area of the large forests (Core 3) in CTML has decreased by 43.08% whereas smaller patches (Core 2, Core 1, edge and patch forests) has increased multifold during 1930–2020. The continued habitat loss and fragmentation probably fragmented elephant populations during the last century and made them insular with long-term ramications for elephant conservation and human-elephant conict. Given the substantial loss in forest cover and high levels of fragmentation, improving the resilience of elephant habitats in Nepal would urgently require habitat and corridor restoration to enable the movement of elephants.


Tracking Forest Loss and Fragmentation During 1930-2020 in Asian Elephant (Elephas Maximus) Habitats in Nepal
people depend on subsistence agriculture and are involved in farm and off-farm-based livelihood activities (Chaudhary and Subedi, 2019). Paddy (Oryza sativa), maize (Zea mays), wheat (Triticum aestivum), lentils (Lens culinaris) are some major food crops, where jackfruit (Artocarpus heterophyllus), mangoes (Mangifera indica), bananas (Musa acuminata) are some fruit crops farmed in the area 44 . Large-scale linear infrastructure projects and mining activities are the major drivers of deforestation and habitat fragmentation in the landscape.
We divided CTML into four regions (Eastern, Central, Western and Far-western) of similar size to assess the extent of forest loss (Table 1). Thus, elephants are distributed in four population clusters with limited connectivity viz. eastern population (Mechi River to Kamala River), (b) a central population (Kamala River to Narayani River), (c) western population (Narayani River to Western boundary of Dang district), and (d) a far-western population (Eastern boundary of Banke district to Mahakali River) 45,46 . Derivation of forest cover We analyzed forest cover change and fragmentation using both the patch and landscape metrics and considered forest fragmentation as habitat fragmentation 47 . We categorized forested areas as natural and plantations with a tree canopy cover of more than 10 percent and an area of more than 0.5 ha 48 . We used the hybrid classi cation techniques to combine high-resolution images, medium resolution images, and digitization of topographic maps. First of all, we prepared a forest cover map of the 1930s by digitizing greenwash areas shown on topographical maps prepared by Army Map Service, U.S. Army, Washington, surveyed during 1920-1940 (http://legacy.lib.utexas.edu/maps/ams/india/) at 1:250,000 scale. Due to the unavailability of multi-spectral satellite images of the study area before the 1970s, we relied on the existing topographic maps to obtain forest cover of 1930 49,50 51,52 found 5-10% inherent errors at various stages of land cover change analysis; while using historical data and topographic maps. The inaccuracy of forest cover mapping was minimized by visual interpretation and overlay analysis in the topographic maps. In addition, we resampled all the digital images at a 30-meter resolution to improve the mapping errors. 51,52 reported the reliability of topographical maps to reconstruct forest cover. We also obtained the forest cover map of 1975 by on-screen digitization of Landsat 1 TM level 1 satellite images.
We produced the forest cover maps of 2000 and 2020 from Landsat imagery scenes of respective years (Table 2; Figure. 2). All the Landsat data processing was conducted using the cloud-computing technology in the Google Earth Engine (GEE) platform (https://earthengine. google.org/). The GEE platform carried out a fast analysis using Google's computing infrastructure 53,54 . We used the pre-processed Landsat imagery available through GEE to assess forest cover change across the study area 55 . We used a cloud screening algorithm to remove cloud contaminated pixels from each Landsat image by applying quality assessment (QA) bands for 2000 and 2020. Then, we produced an annual composite by taking the median value from images from the target year 56 . We delineated > 1000 reference points for each period 2000 and 2020 respectively. We used supervised machine learning classi ers, i.e., Random Forest (RF), to classify remotely sensed data 57 .
Random Forest Classi er creates a set of decision trees from a randomly selected subset of the training set and aggregates the votes from different decision trees to classify the image 58 . The classi ed image was downloaded as raster tiff les. The raster was converted into vector polygons and overlaid with highresolution google earth images of respective years. The nal forest cover map was obtained with the highest accuracy by post-processing (validating) the We have estimated the conversion of forests into the non-forest area on a grid overly basis and grids were generated on the basis of minimum home range size of elephants i.e., 18 km 2 45,59 . We generated 5× 5 km 2 grids for the time series assessment and analyzed spatial distribution trends of forest cover in these grids from 1930-1975, 1975-2000, and 2000-2020 49,60 . We computed the forest cover area (distribution of transitions and persistence of forest) of four different periods in each grid using the zonal statistics tool of ArcGIS software 61 . Overall, forest cover change was calculated by combining all the grids and calculating the annual deforestation rate (percentage) using a compound-interest-rate formula 62 .
where a1 and a2 are the area covered by forest at times t1 and t2. The region wise rate of deforestation was computed and presented.

Modeling Forest fragmentation
We carried out habitat fragmentation analysis in the four regions of CTML ( Fig. 1) and measured fragmentation in terms of core, perforated, edge, and patches. We used 30 m cell resolution for fragmentation analysis for four different periods. We used patch analyst 63 to obtain the patch matrix for each region viz. patch density and size (number of patches, mean patch sizes, patch size standard deviation), edge metrics (edge density, mean patch edge), and shape index (mean shape Index, mean perimeter area ratio, mean patch fractal dimension), (Supplementary table S1).
Similarly, Landscape Fragmentation Tool (LFT V2.0, http://clear.uconn. Edu/tools/lft/lft2/) was used to estimate landscape metrics 64 . The change of fragmentation during the 1930 to 2020 periods was carried out by cross-tabulating the fragmentation classes. Landscape Fragmentation Tool (LFT) classi es forests at pixel-level into fragmentation classes: core 1, core 2, core 3, perforated, edge, and patch. Core forests are located far from the forest/non-forest boundary and surrounded by other forest areas. We considered the core forest as 100 m distance from the edge 65 . The core forests include three different types -1) Core1: forest patches area < 250 acres (1.012km 2 ), 2) Core 2: Medium core (forest patches area between 250-500 acres (1.01-2.2 km 2 ), and Core 3: large core (Forest patches area > 500 acres (> 2.2 km 2 ) 49 . The peripheral forest was further classi ed into perforated (i) inner edge: forest pixels on the edge of small interior non-forest, and (ii) edge forest or outer edge: pixels that are between forest and large non-forest areas 66 .

Results
Temporal change of forest cover in CTML We estimated 24,315 km 2 of forest cover in 1930. The forest cover was reduced to 19,069 km 2 in 2020, with an annual rate of 0.27%. The deforestation rate (0.29%) was higher between 1930 and 1975. The highest rate of deforestation was documented in western region (0.33%) followed by eastern (0.29%), far western (0.28%) and central (0.16%) region between 1930 to 2020 (Table 3). In 2020, the far western region had the highest forest area (35.42%), followed by the western region (26.18%), central (19.78%), and eastern region (18.61%) of CTML (Table 3) Spatial change in forest cover Altogether 1,592 grids of 5 x 5 km 2 were used to analyze the spatial patterns of forest cover change. Deforestation was documented in most of the grids (n = 1505), and 75 grids lost entire forest area between 1930-2020. Increase in forest cover was observed in only 51 grids during the same period. The massive reduction in large forest patches (< 20 Km 2 ) was documented for 26 grids (Fig. 3a,  We calculated historical forest fragmentation for the last nine decades (1930-2020) and found that the total number of patches increased from 201 in 1930 to 28,559 in 2020. However, the highest decrease in mean patch size (121 km 2 in 1930 decreased to 0.7 km 2 in 2020) indicates that the forest has been fragmented into small patches. The mean perimeter ratio of the forest has been increased from 187 in 1930 to 1210 in 2020. The edge metrics showed that edge density increased from 548 (m/km 2 ) to 3630 (m/km 2 ), reduced mean patch edge to 2,426.63 ha from 66,271.31 ha. Similarly, the shape index suggested that the mean shape index (MSI) was decreased sharply where the mean perimeter area ratio (MPAR) increased progressively (Table 4).  whereas core 1 and core 2 increased by 488% and 162%. Finally, the far western region lost 5.51% of the forest, and the core 3 forest was reduced by 37.11%, whereas core 1 and core 2 increased by 663% and 145% simultaneously.
The overall forest fragmentation result suggests that the highest fragmentation occurred in the eastern region (in core 3), followed by the central, far western, and western region, where the core forest (core 3) was reduced by 57.34%, 46.36%, 37.11%, and 30.88% simultaneously (Table 5 and Table 6), (Fig. 4; Supplementary gure S5).

Patterns of forest cover change and fragmentation
Our study provides comprehensive information on forest cover (habitat) change and fragmentation within the primary elephant habitat in Nepal between 1930 and 2020. We documented the loss of more than one-fth of the forest area and extensive fragmentation during this period. Our results suggest that the elephant habitat remained intact during the 1930s. However, the rate of deforestation was higher between 1930 and 1975 due to the conversion of forests into agricultural land. The results indicate that Government should prioritize conservation efforts to restore elephants' movement within the human-dominated landscapes outside protected areas.
Forest fragmentation results suggested that large forest patches have decreased rapidly, whereas forests in the medium and small core have increased massively. Similarly, the area of forests in the patch, perforated, and edge category has also increased during the last nine decades (1930-2020), which indicated the high rate of forest fragmentation in the CTML. Landscape metrics analysis also reveals the massive fragmentation of forests between 1930 and 2020, increasing the patch number and decreasing patch size (Table 5,6 and Fig. 4) 68 . Similar to Nepal, a massive decline in extensive core forests and increase in fragmented patches has been documented in other elephant range countries India, Myanmar, Bangladesh, and Sri Lanka 60,62,72,73 . Fragmented forest patches should be connected through a combination of the week-and high-quality habitat to enable elephant connectivity throughout the landscape.
The human pressure (illegal grazing, resources extraction) and risks of invasive species (Lantana camera, Chromolaena odorata, Parthenium hysterophorus, Mikania micrantha, etc.) spread are high in smaller and perforated forest patches as well as forest edges 74 .
Elephant habitat is more fragmented outside protected areas due to the high pressure of encroachment and developmental activities 75 . These forests are also used more frequently by the local communities to meet their subsistence needs of livestock grazing and dependence on forest products 76,77 . With increasing forest fragmentation, the elephants and other wildlife are also forced to live in smaller forest patches with spatial overlap with human activities 78 . This situation increases the chances of confrontation between humans and elephants, often leading to fatal attacks 30,79 .The eastern region had the highest forest fragmentation (57.3% of large core forest lost) within our study, where HEC incidents were also the highest (DNPWC 2020). The eastern region also bears a long migratory route of a large herd of elephants (> 100) and provides habitat for some residential elephants. Although the forest cover is not signi cant within Koshitappu Wildlife Reserve (KTWR), it still provides refugia and a corridor for elephants in the eastern region. While navigating through the highly fragmented forests, there is always a threat of elephants getting de ected due to haphazard drives and another form of human resistance resulting in elephants ending up in human-use areas off the forests, as corroborated by telemetry studies on elephants in the landscape.

Drivers of deforestation and fragmentation
Several studies indicate loss of elephant habitats and fragmentation due to a combination of multiple factors such as (i.e., agriculture and settlement expansion, encroachments, irrigation, infrastructure development hydropower projects, illegal logging, mining, commercial plantations) 60, 80-82 . Additionally, expansion of oil palm plantation in Indonesia 83 and tea, paddy cultivation in north-east India has also contributed to habitat loss 84 . However, in Nepal, forest conversion into farmlands through government policy was responsible for forest loss and fragmentation in the initial years  Large-ranging species like elephants are affected by this as they come into frequent clash with humans while navigating seasonally through these highly fragmented forests in CTML.
This study indicates that the conservation of large-ranging species like elephants and tigers in CTML has been challenging as most of the remaining forests are highly fragmented, especially outside the protected areas. With planned and ongoing infrastructure development activities in CTML, forest fragmentation continues to increase. It shows the importance of the landscape-level conservation approach and helps policymakers restore corridor and connectivity by the implementing metapopulation management of large mammals in Nepal and around the Globe.

Conclusion
Forest loss and fragmentation induced a severe threat to elephant conservation in Nepal. Such fragmentation brought both the elephants and humans along the forest's edge, where they interact with each other, often resulting in severe human-elephant con ict (HEC). Increasing the number of forest patches also increases the transparency in the migratory routes, increasing the poaching threats. Our research ndings have implications for devising appropriate policies for conserving large mammals and their habitats in human-dominated landscapes in Nepal and beyond. Further understanding of the relationship between forest loss/fragmentation and human-elephant con ict is necessary. The particular focus of elephant conservation is necessary outside the protected areas and migration corridors where habitat is highly fragmented.
Declarations the ground veri cations.

Data accessibility
Upon publication of the article, all the supporting data for obtaining the results will be made available via the online data services such as dryad. Overall methodological framework adopted for this study.