Landscape predictors of human–leopard conflicts within multi-use areas of the Himalayan region

Conflict with humans is a significant source of mortality for large carnivores globally. With rapid loss of forest cover and anthropogenic impacts on their habitats, large carnivores are forced to occupy multi-use landscapes outside protected areas. We investigated 857 attacks on livestock in eastern Himalaya and 375 attacks in western Himalaya by leopards between 2015 and 2018. Multivariate analyses were conducted to identify the landscape features which increased the probability of livestock depredation by leopards. The risk of a leopard killing livestock increased within a heterogeneous landscape matrix comprising of both closed and open habitats (very dense forests, moderate dense forests, open forests, scrubland and non-forests). We used the results to map potential human–leopard conflict hotspots across parts of the Indian Himalayan region. Our spatial risk maps indicate pockets in the eastern, central and western part of eastern Himalaya and the central, northern part of western Himalaya as hotspots of human–leopard conflicts. Most of the attacks occurred when livestock were grazing freely within multi-use areas without supervision of a herder. Our results suggest that awareness about high risk areas, supervised grazing, and removing vegetation cover around human settlements should be initiated to reduce predation by leopards.

Large carnivores are apex predators and their lethal and non-lethal effects have strong implications for ecosystem structure and functioning. They also have prominent cultural reverence in several human societies and act as flagship species for global conservation campaigns 1 . In areas where their populations have suffered local extinction, cascading effects have been observed on ecosystem processes either mediated by abundance or behavior of mesocarnivores and prey 2,3 . Low population densities, high energetic requirements, social or solitary hunting strategies and wide-ranging behavior means that they often range beyond nature reserves for access to space and resources and seek out easy prey i.e. livestock and occasionally humans within shared landscapes. Species commonly involved in predation on livestock are canids (Canis spp.), felids (Panthera spp.) and ursids (Ursus spp.) across both Northern and Southern Hemispheres 4 . Government agencies, wildlife managers and conservationists spend substantial money to compensate losses to carnivores through a combination of preventative, assertive and reactive measures 5,6 . In areas where the perceived risk of large carnivores is high, attacks on humans and livestock lead to retaliation by the local communities increasing their extinction risk 3 , and a setback for conservation efforts 7 .
Though protected areas are essential for conserving large carnivores, their size is often not large enough to sustain carnivore populations or even individual home ranges 8 . Therefore, the quality of the larger multi-use landscapes in terms of the availability of suitable habitat, abundance of wild prey, extent of human presence and tolerance of local communities determine future survival of these species. Such shared landscapes represent a major proportion of the geographic distribution of large carnivores globally 9 . Increasing human populations, declines in wild prey and changes in land use patterns have resulted in fragmented, heterogeneous resource limited landscapes for carnivores 10 . Such anthropogenic impacts on ecosystems have forced carnivores to frequent areas near human settlements, kill livestock and consequently local communities have retaliated through

Influence of landscape features on leopard attacks on livestock.
Moran's I identified spatial clusters of livestock predation within both North Bengal and Pauri Garhwal. The z value (13.422), Moran's Index (0.395) and (p value < 0.01) indicate that there was a less than 1% likelihood that this pattern was due to random chance in North Bengal. The threshold distance between each neighboring livestock depredation site was estimated to be 5,000.50 m.
The z value (11.329), Moran's Index (0.499) and (p value < 0.01) indicate that there was less than 1% likelihood that this pattern was due to random chance in Pauri Garhwal. The threshold distance between each neighboring livestock depredation site was estimated to be 5,077.024 m.
Results of the GLM model with binomial structure suggests that landscape features such as area of moderate dense forests (β = 2.95E−07; 95% CI = 1.21E−07-4.70E−07), open forests (β = 4.62E−07; 95% CI = 2.27E−07-6.97E−07), scrub (β = 5.58E−07; 95% CI = 8.30E−08-1.03E−06), non-forests (β = 3.38E−07; 95% CI = 1.92E−07-4.85E−07) and the distance to protected areas (β = 3.11E−05; 95% CI = 7.83E−065.44E−05) were the best predictors of livestock predation by leopard in Pauri Garhwal ( Table 2, Supplementary Table S3). The probability of livestock killing increased within moderate to open habitats and with distance to protected areas. Results of the GLM model with poisson structure suggests that landscape features such as area of moderate dense forests (+), open forests (+), scrubland (+), non-forest (+), water (−), distance to protected areas (+) and night light (+) were the best predictors for leopard attacks on livestock (Supplementary Table S4). The probability of livestock killing increased within moderate to open habitats. In addition, livestock killing was most likely to occur in areas with less water, more human presence and increasing distance from protected areas. Table 1. Second-order Akaike Information criterion scores (AIC), (AICc), ΔAICc of generalized linear models with binomial structure predicting livestock depredation by common leopards in North Bengal landscape, using landscape predictors. PAs protected areas. . For Pauri Garhwal, predation risk was high in the central and north eastern part of the landscape. Conflict probabilities were higher near protected areas in North Bengal whereas they were lower in Pauri Garhwal.
Wilcoxon rank sum test results indicate that majority of the predictor variables (area of very dense forests, area of moderate dense forests, area of open forests, area of scrub, area of water/riverine patches, area of non-forest, length of roads, length of river, nightlight, p < 0.05) differed significantly between the two study sites (Supplementary Table S5). Only distance to protected areas between the two study sites did not exhibit a statistically significant difference (p > 0.05). was greater than 7,430,000 sq. meters i.e. 30% for North Bengal. For Pauri Garhwal, predation risk was highest in areas (> 0.86) when area of non-forests was less than 6,820,000 m 2 i.e. 27% and area of open forests was greater than 3,630,000 i.e. 15%. These results indicate a first priority for foothills, a distance of 6.3 km from protected areas, dense forests (dense vegetation cover) and open forests (tea plantations) for predation risk by leopards in North Bengal. For Pauri Garhwal landscape, risk of predation was highest in areas within a matrix of non-forests (human settlements) and open forests (low vegetation cover).

Discussion
Our results suggest that landscape features are major predictors of livestock predation by leopards in IHR. Our study also highlights that there is significant spatial clustering of livestock kills by leopards within multi-use landscapes of IHR. At a coarser spatial scale, livestock depredation risk was higher within both closed and open habitats in North Bengal (very dense forests, open forests, non-forests, scrub). In Pauri Garhwal risk was higher within a matrix of moderate and open habitats (moderate dense forests, open forests, scrub and non-forests). Generally, predation risk probabilities by large carnivores are reported to increase with dense vegetation cover 16, 24, 41 but our results suggest that risk of livestock predation by leopards are higher within a multi-use matrix of open and closed habitats. Our results also indicate that the hunting behavior of leopards and choice of livestock kill sites are different compared to the other large carnivores who rely on dense vegetation cover. Such hunting strategies are probably an artifact of the adaptive nature of leopards, their ability to survive within close proximity of human settlements and consume a wide range of prey 30,32,42 . Human-carnivore conflicts have been generally reported along the periphery of protected areas 8,10 , and probability of livestock killing by leopards increased with proximity to protected areas in North Bengal and decreased with increasing distance from such reserves in Pauri Garhwal. Through this study we also mapped human-leopard conflict hotspots within the IHR which should be prioritized by protected area managers, district administration, local communities and multi-interest groups.
Large carnivore species e.g., hyenas 43 , brown bears 44 , lions 45 and tigers 29 use protective vegetation cover to hunt both wild and domestic prey. In contrast, leopards are reported to kill wild prey within moderate vegetation cover such as open mixed woodlands 21 in Africa whereas in India availability of water was a crucial spatial driver of leopard kills 46 . Our result is similar to a study conducted in the Himalayan region of Bhutan which documented that livestock predation by leopards was higher within a matrix of forest and agriculture 47 . Leopards Table 2. Second-order Akaike Information criterion scores (AIC), (AICc), ΔAICc of generalized linear models with binomial structure predicting livestock depredation by common leopards in Pauri Garhwal landscape, using landscape predictors. PAs protected area.  35,36 . As reported by an earlier study, leopard attacks on humans in Pauri Garhwal were negatively associated with the presence of dense forests 36 but more influenced by a matrix of agriculture lands, shrub cover and medium dense forest patches. Within a fine scale, availability of water has often been identified as a major driver of human-carnivore conflicts in arid ecosystems of Africa 25, 41, 49 but within productive ecosystems of South Asia availability of water is probably not a limiting factor. The threshold value of 6.3 km adjoining protected areas in North Bengal is similar to previous studies which documented livestock predation risk to be high within 5-12 km of the park boundary in Nepal and Cameroon 43,44 . In South Africa risk of a leopard killing livestock was highest within the periphery of protected reserves 50 . Though the probability of livestock killing increased at the edge of protected reserves, we document contrasting www.nature.com/scientificreports/ results from the two landscapes. Such results could be due to interspecific competition between two large cats 25, 39 within the same landscape i.e. tiger and leopard. Pauri Garhwal has two protected areas in the foothill region of the district with a sizeable population of Bengal tigers 51 and due to competitive exclusion leopards probably are more widely distributed in the hills away from the reserve boundaries and hence conflicts with leopards are higher as we move farther away from the reserves. North Bengal region has no other apex predators to compete with leopards (due to recent extinction of tigers) and thus human-leopard conflicts are confined to the edge of protected areas and decrease with increasing distance from such reserves. Thus, patterns of conflicts with the species show a different spatial pattern between the two sites.
Leopards are reported to kill livestock in rugged areas of Bhutan 47 but our results suggest that livestock kills occurred within flat low lying and mid elevated zones. In North Bengal the average elevation of sites where livestock were killed was low i.e. 270 m probably as a consequence of abundant livestock availability in the foothills compared to the hills. In Pauri Garhwal, the average elevation of livestock kill sites was 1,200 m indicating higher livestock availability in the middle Himalayas compared to the foothills 52 .
There was a strong effect of seasonality on the number of attacks with the majority occurring in the dry season (winter and summer). These results are in accordance with studies conducted in the Terai region of Nepal 53 where www.nature.com/scientificreports/ leopard attacks on livestock were also reported to be higher in the dry season (winter and summer months). Our results are contrary to studies conducted in East Africa where predation on livestock by lions, hyenas and leopards were much higher in the wet season compared to the dry seasons 54,55 . In Africa, wild prey availability is low during the wet season 55 whereas in South Asia there is probably no significant association of wild prey and seasonality. The seasonal pattern of predation events further coincides with harvesting and planting of major agricultural crops such as maize, wheat and paddy. During these dry months, local community members are often not available to guard livestock. Livestock grazing is higher during winter within agricultural fields, tea plantations and scrubland due to unavailability of forage within the forest interiors. Livestock kills within both sites were both diurnal and nocturnal which differs with similar studies conducted in the Himalayan region i.e. Pakistan 56 and Bhutan 57 where leopard attacks were nocturnal in nature. Though leopards have been reported to be nocturnal in areas with and without humans 48, 58 our results suggest diurnal activity peaks within human dominated landscapes of the Himalayan region. Similar to leopards, cheetahs also exhibit diurnal activity peaks within human-dominated landscapes of eastern Africa 59 . Big cats have generally been reported to kill livestock at night as with jaguars 60 in the Pantanal, lions, leopards in Africa 61 and tigers in India 62 . Leopards probably hunt wild prey at night but livestock killing is more pronounced during the day due to availability and ease of prey catchability.
Leopards in general are reported to kill cattle, goats and sheep when wild prey biomass falls below 540 kg/ km 263 . Limited studies on leopard diet within anthropogenic landscapes of India highlight the major contribution of livestock 32,33 . Prey abundance and availability is a major driver of human-carnivore conflicts 63 but landscape level data on prey abundance was not available for the study sites to confirm such a relationship. However, our results indicate that wild prey biomass could be low in the Indian Himalayan region leading to increased attacks on livestock. Cattle loss to leopard attacks was higher in Pauri Garhwal whereas predation on goats were higher in North Bengal. Cattle are largely left unsupervised due to the low economic benefits compared to goats in Pauri Garhwal and hence predation on cows were higher. In North Bengal, cattle provide greater economic benefits and hence are generally supervised by an experienced herder. Cattle density (indigenous and cross breed) was also high i.e. 66 animals per km 2 whereas goat density was 27 animals per km 2 in Pauri Garhwal (Livestock Census 2012). Cattle density in North Bengal was 92 animals per km 2 whereas goat density was 80 animals per km 2 (Livestock Census 2012). Leopards show preference for wild prey species with an estimated prey weight range (10-40) kg 11,54,55 , and hence exhibit similar size preference when preying on livestock. Thus, goats were killed in higher abundance in North Bengal compared to Pauri Garhwal.
Previous studies in the mountainous regions of South Asia have documented leopard attacks on livestock to be geographically widespread without much significant association with human presence 57,64 . In Bhutan livestock predation probability at a fine scale by leopards and tigers were positively associated with the density of human settlements 47 . Our results suggest that at a wider landscape scale, human presence (indicated by night light) is negatively and positively related to probability of conflicts in North Bengal and Pauri Garhwal, respectively. North Bengal is a densely populated region with human densities ranging from 200-700 individuals per km 2 whereas in Pauri Garhwal, human density is much lower i.e. 110 individuals per km 2 . Leopards probably are adapted to cooccur close to humans but there exists a threshold beyond which there is an underlying avoidance and could be a result of carnivore's perception of fear (persecution risk) within multi-use landscapes 65 . Our results are similar to space use by lions and cheetahs who avoided humans at a landscape scale due to the fear of persecution 59,66 .
Output of the risk models is generally used globally to focus interventions and plan implementation of mitigation measures for minimizing damage caused by human-carnivore conflicts 15,16 . The conflict risk maps generated through this study will be helpful to prioritize mitigation measures and reduce livestock depredation by leopards within IHR. Such measures will help reduce the present extent of retaliation by local communities and ensure survival of leopards outside of protected reserves. Policy makers will be able to better allocate resources to compensate livestock losses, forest and wildlife administration will be able to concentrate mitigation measures within specific sites and local communities will be able to avoid grazing of livestock within high risk zones. Livestock depredation can be reduced by improving existing animal husbandry practices such as using trained livestock guardian dogs and professional herders while grazing. To reduce economic damage to leopard attacks, agro-pastoralist societies of the IHR should be provided with monetary incentives through community based nature tourism initiatives. Awareness programs should be organized to educate livestock owners about biology of leopards, high risk areas and patterns of human-leopard conflicts. Radio-telemetry studies should be initiated to understand fine scale habitat utilization by leopards within anthropogenic landscapes. Finally, collaborations between managers, conservationists, local communities and understanding cultural complexities of human-leopard relations will be crucial to ensure future coexistence within heterogeneous mountainous landscapes of South Asia.

Methods
Study area. The study was conducted across two landscapes (North Bengal and Pauri Garhwal) located in eastern and western Himalaya.
The North Bengal landscape is located in the north-eastern part of India (Fig. 3) with a forest cover of 46% 67 . Predominant forest types are northern tropical semi-evergreen, moist deciduous forest and the landscape is a matrix of tea gardens, protected areas, agricultural lands and urban settlements 59  There has been the recent extinction of Bengal tiger (Panthera tigris tigris) from this region with leopard (Panthera pardus fusca) being the apex predator and only large carnivore present 69 . Pauri Garhwal district, part of the lesser, middle Himalaya and located in the north-western part of India (Fig. 3) has an altitudinal range of (200-3,200 m) with a forest cover of 64% 67 . Predominant forest types are moderate dense forests, open forest and scrubland 58 . Annual rainfall range is between 1,000 and 2,500 mm. Human density is 110 persons per km 2 (Census 2011, data accessed on July 2019) and the local inhabitants are mainly agrarian with livestock farming, horticulture and cottage industries as major professions 64 . Livestock density of this region is 58 per km 2 (Livestock Census, 19th All India Livestock Census Report 2012). This district has two protected areas in the foothill region with a sizeable population of Bengal tiger (Panther tigris tigris) 51 . Other than tigers, common leopards are the second most dominant apex predator in this region with presence of other mammals such as barking deer (Muntiacus muntjak), goral (Naemorhedus goral), sambar (Rusa unicolor), wild pig (Sus scrofa), rhesus macaques (Macaca mulatta) and common langur (Semnopithecus entellus).

Data collection.
We used a two-step technique to collect data on leopard predation on livestock. First, we searched newspapers and West Bengal, Uttarakhand state forest department compensation registers for reports of incidents of leopard attacks on livestock in North Bengal and Pauri Garhwal between 2015 and 2018. Our primary aim was to avoid strong spatial bias in data and hence we checked for incidents which were spatially spread out and not confined to specific localities or regions. The compensation registers contained information such as date of incident, carnivore species involved, number and type of livestock killed and name of village/ locality. Second, we conducted structured interviews across the regions and asked the owners and community members about age and species of livestock killed by leopards (survey protocol, see Supplementary Information Appendix 1). The livestock owner and forestry officials who were present during the attack or had verified the authenticity of the incident escorted the research team during such field visits. If there was ambiguity in response for a particular incident, we refereed only to the compensation register detail. We also crosschecked every detail of the incident with the report submitted by forestry officials after the initial investigation. The forestry officials were responsible for evaluation of wildlife attacks which is the basis for payment of compensation payments. To avoid exaggeration of livestock losses, we informed the community members that there will be no incentive or compensation provided as a part of the survey and the results will be solely used to help demarcate high and low conflict risk zones. We also recorded season and a time to livestock killed by leopards and asked owners about how often shepherds or community members used to accompany the animals during grazing ( Supplementary  Information Appendix 1). Based on the information gathered, we visited 857 sites in eastern Himalaya (North The study areas were stratified into a scale of 5*5 i.e. 25 km 2 which resulted to a total of 601 cells for North Bengal and 254 for Pauri Garhwal respectively. The cell size was selected based on ecological considerations to reduce chances of autocorrelation and identify landscape drivers of human-leopard conflicts. We generated a count statistic for the exact number of predation events for each cell. Cells where there were no attacks were assigned 0. To model the spatial spread and extent of livestock depredation we prepared both binary (presence-1 cells with at least one or multiple attacks, absence-0-no attacks) and count data (presence-exact number of attacks recorded and absence 0-no attacks). The number of cells on which leopard predation on livestock was recorded was 127 and 72 for North Bengal and Pauri Garhwal, respectively. Data preparation for spatial risk analysis. We identified a total of 5 major landscape features (Habitat, Water, Human presence and infrastructure, Distance to Protected Reserves and Altitude) for North Bengal and Pauri Garhwal (Table 3) based on their ecological importance to model predation risk.
1. Habitat: We hypothesized that predation risk by leopard will be higher in areas with forests, presence of riverine patches, water bodies. Area of land use types: We calculated landscape variables i.e. area under different land-use types and water sources such as area of riverine patches, water bodies from Forest type map of India (2014) 70 . 2. Water: Large predators are reported to kill prey in areas with availability of water 15, 20 , thus we hypothesized that availability of water will be a major driver of leopard predation on livestock. We calculated the length of rivers and area of river bodies from the Roads and Drainage layers obtained from Digital Chart of the World and Forest type map of India 70 . 3. Human presence and infrastructure: We hypothesized that leopards would avoid killing livestock in areas with increased human presence 71 . We extracted night light values using the 1,000-m spatial resolution nighttime visible light data of India 72 . We calculated length of roads using the Roads and Drainage layers obtained from Digital Chart of the World. 4. Distance to Protected Reserves: We hypothesized that the probability of leopard killing livestock will be higher in areas closer to protected areas 17 . We calculated distance from protected areas (Protected Area Network of India) using the Euclidean distance tool for each cell. 5. Altitude: Considering that carnivores prefer to kill livestock in areas with gradient in altitude 47 , we hypothesized that predation risk by leopards will be higher in elevated regions. Hence, we generated the mean altitude value for each cell based on 90-m spatial resolution digital elevation maps 74. www.nature.com/scientificreports/ Once the cell files were finalized and compiled, we clipped all ecological variables to 25 km 2 cells. We omitted redundant correlated variables ≥ 0.70 74 based on Pearson correlation coefficient values calculated using R version 3.4.0. Area of moderate dense forests was positively correlated with the area of very dense forests in North Bengal (value 0.76) and hence it was excluded from the analysis. Length of river was also positively correlated with length of roads (value 0.8) and hence it was also removed from the analysis. Altitude was positively correlated with distance to protected areas in Pauri Garhwal (value 0.76) and hence it was excluded from the analysis.
We used 4 analytical approaches to model probability of livestock depredation by leopard. In the 1st step we evaluated spatial autocorrelation among livestock kills within the cells using function moran.test (Moran's I) in package (spdep) 75 in R 3.4.0. In the 2nd and 3rd approach we used generalized linear models (GLMs) with binomial and poisson structures to quantify the effect of landscape features (area of habitat types, availability of water, human presence and infrastructure, distance to protected areas and altitude) on livestock predation. All the predictor variables considered for the analysis were continuous in nature. We used a priori candidate models and ranked them based on AIC, AICc values 76 . Models with the lowest AICc values were considered the best or dominant model and the output (coefficients and estimates) explained the probability of livestock predation by leopards within IHR. Based on the results of the dominant model or the model averaged coefficients, we generated conflict hotspots for both the study sites. We used coefficients of the best model (binomial structure) or averaged all candidate models (GLM with binomial structure) to estimate probability of livestock depredation for each cell (25 km 2 ) using the equation p (x) = exp (z)/ (1 + exp (z)) and generated human-leopard conflict hotspots in Arc GIS 10.3 77 . We generated ROC curve and AUC values to predict reliability of the dominant models using package ROCR 78 in R 3.4.0. Since predictor variables between the two study sites were not normally distributed, we compared the identical landscape features using nonparametric Wilcoxon Signed-Rank Test in R 3.4.0.
In the 4th step, we used the predictor variables of the dominant models to calculate conditional inference (CTREEs), as prescribed in the R-package "partykit" 79 . This method was adopted to obtain threshold values for the significant variables for conflict mitigation recommendations. Trees based on maximally selected rank statistics were fitted using the Bonferroni correction for multiple testing and a minimum sum of weights. In addition, univariate trees were fitted for variables with a significant split in the multivariate tree. The results of our two analytical approaches (regression and ctree) are similar and provide an overall understanding of landscape features prone to livestock predation in accordance with the behavior of common leopards. Both analytical methods are based on a maximum likelihood approach and when interpreted together provide meaningful results. The GLM models computes probabilities of an event based on a logistic regression framework while the CTREE uses a machine learning classification approach and assigns values to predefined categories. The decision tree approach is a non-parametric approach which helps simplify complex relationships between dependent and predictor variables.