Patterns of livestock predation risk by large carnivores in India ’ s Eastern and Western Ghats

Large scale spatial patterns of livestock predation risk from multiple co-predators are fundamental to applied conservation planning. Here, we examined important ecological, social, and landscape attributes explaining spatial patterns of human-carnivore interactions. We used a systematic grid-based framework, across an area of 14,200 km 2 of sixteen Forest Divisions at the human-wildlife interface encompassing Protected Areas, Reserved Forest and Fringe Areas at the human-wildlife interface in the Eastern and Western Ghats, India. The data was collected on livestock depredation incidents from the tiger ( Panthera tigris ), leopard ( Panthera pardus ), and dhole ( Cuon alpinus ) for the past ﬁ ve years, through semi-structured interviews (n ¼ 1460) of local communities. We examined socio-ecological (i.e. livestock abundance and forest dependency) and landscape attributes (i.e. forest cover, climates and topographic) in ﬂ uencing the depredation events from each carnivore species. We found that livestock predation risk by the tiger, leopard and dhole was driven by the size of livestock species, the dependency of local people on the forest, topography, proximity to water body and the forest boundary, precipitation, and forest cover. Risk of predation from leopard and dhole exhibited high spatial overlap, and pre-dation by leopards was higher than dhole and tiger. Livestock predation by leopard and dhole was frequent in open areas of Reserved Forest and buffer zones, while that from tiger occurred in densely forested core regions of Protected Areas (PAs). Our predictive risk maps (ca. 22,525 km 2 ) showed species-speci ﬁ c predation patterns, re ﬂ ected ecological differences among large carnivores with regards to their habitat and spatial partitioning for domestic prey. Our predictive predation risk map and factors associated with livestock predations provides powerful visual guidance and tools for PA managers in developing multi-species con ﬂ ict mitigation strategies. We recommend diversifying local economic livelihoods and bene ﬁ t-sharing options for local communities to minimize their forest dependency. © 2020 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND


Introduction
Human-wildlife conflict is a precarious threat to the survival of many endangered large mammal species (Madhusudan and Mishra, 2003;Treves and Karanth, 2003;Woodroffe et al., 2005;Karanth et al., 2013a).Livestock predations are globally recognized as the most common form of human-carnivore conflict (HCC) (Inskip and Zimmerman, 2009;Treves et al., 2004;McInturff et al., 2020).In India, conflict frequently brings major economic loss to the forest-dependent and the forest-adjacent communities, threatening local livelihoods and economic security (Lyngdoh et al., 2014;Bargali and Ahmed, 2018).Protected Areas (PAs) and surrounding Reserved Forests (RFs) are currently exposed to the expanding human settlements, compromised habitat connectivity, increased forest patchiness, and urbanization which result in large mammalian carnivores to occupy spaces amidst or adjacent to human habitations (Wikramanayake et al., 2004;Inskip and Zimmermann, 2009;Mascia and Pailler, 2011;Athreya et al., 2016).Most PAs in India are ca.< 300 km 2 , surrounded by dense human population (>300 people/km 2 ), and host high livestock densities, 62 heads per km 2 global average (10 heads per km 2 ) (Karanth and Madhusudan, 2002;Robinson et al., 2014;Athreya et al., 2016).Since the country's PA management frequently permits grazing of both native herbivores and livestock in some human-wildlife interface areas, the resulting wildlifelivestock interactions can have negative consequences on ecosystem functions (DeFries et al., 2007;Thinley et al., 2018).Although forest resource collection and grazing by local people are not allowed within the PAs of India, the resistance from people's traditional belief system often over-powers law enforcement by the government sector (Ramesh et al., 2019).Today over half of India's human population is engaged in some form of agriculture and forestry, including livestock grazing, also the land in and around PAs are frequently exposed to these diverse traditional livelihood practices (Census of India, 2011; Meiyappan et al., 2017).
Livestock grazing is a major contributor to wild prey population decline through reduced forage availability (Madhusudan and Mishra, 2003), increased rate of disease transmission to wild herbivores (Rahmani, 2003), thereby switching dietary preferences of large carnivores from wild prey to domestic prey.Thus, the overlapping resources and shared spaces between humans and large carnivores often lead to livestock predation, people's retaliation by snaring or poisoning of conflict individuals (Madhusudan, 2003;Kalaivanan et al., 2010;Gubbi, 2012;Khorozyan et al., 2015), human injury or threat to livelihood security (Ripple et al., 2014;Wildlife Protection Society of India, 2017).If sustained over time, many of these areas will eventually turn into "conflict hotspots" (Broekhuis et al., 2017).Although some of India's PAs are experiencing an increase in large carnivore populations due to the implementation of consistent and effective conservation measures (Jhala et al., 2019), human-carnivore interactions are becoming more frequent, when high densities of large carnivores occur in fragmented small PAs (Wikramanayake et al., 2004;Kshettry et al., 2017;Sidhu et al., 2017), thereby making multispecies carnivore conservation goals even more challenging (Woodroffe et al., 2005).
Along the human-wildlife interface, different large carnivore species can vary in their response to anthropogenic stressors (Harihar et al., 2007;Bhattacharjee et al., 2015;Miller et al., 2016a;Tyagi et al., 2019) and might even select relatively more domestic prey in proportion to native wild prey (Miller et al., 2016b;Chetri et al., 2019).Although both tiger (Panthera tigris) and leopard (Panthera pardus) are stalking hunters and prefer structurally complex heterogeneous habitat with adequate ambush cover (Ramesh, 2010) the former requires larger contiguous and less disturbed forested landscapes (Jhala et al., 2010(Jhala et al., , 2019)), making them less resilient to human modifications and activity (Harihar et al., 2007;Bhattacharjee et al., 2015;Tyagi et al., 2019).On the contrary, leopard is highly adaptable, more tolerant of human-wildlife interface areas across more landuse contexts (Athreya et al., 2013;Odden et al., 2014;Kshettry et al., 2017).The pack-living dhole (Cuon alpinus) hunt in relatively more open habitats (Ramesh et al., 2009;Ramesh, 2010) and can utilize secondary forests, multi-use forest fragments, and agro-forest plantations (Majumder et al., 2011;Srivathsa et al., 2014Srivathsa et al., , 2019a;;Punjabi et al., 2017;Thinley et al., 2018).Because India's PA managers have little jurisdiction over areas outside PA boundaries, along with the humanwildlife interface areas, there is a constant struggle in the implementation of forest conservation and community land-use policies to achieve broad conservation goals (Macura et al., 2011).Therefore, in landscapes where multiple large carnivore species overlap with human habitations, HCC mitigation strategies and management approaches will need multi-predator HCC management approach.
Many previous studies on livestock predations have focused only on single species of carnivores, being highly localized, or limited to a few PAs (Athreya et al., 2013;Harihar et al., 2014;Miller, 2015;Kshettry et al., 2017;Naha et al., 2018).Thus, we require further detailed research on multi-predator HCC studies focusing on a larger landscape.HCC surveys on a larger spatial scale or encompassing multiple landscapes for wide-ranging large carnivores can provide detailed insights into differences in the frequency and distribution of predations across various management systems and socio-cultural gradients (Karanth et al., 2013a;Srivathsa et al., 2019b).Suitable predation risk maps are needed to guide wildlife managers and locals to minimize this multi-predator HCC.As spatial predation risk model is a quantitative tool used to predict the probability of livestock predation based on existing location-specific predation event and relating it to landscape attributes (Karanth et al., 2012;Miller et al., 2016b;Naha et al., 2020).The outcome of predation risk models helps in the prioritization of management interventions and implementation of HCC mitigation measures including minimising retaliation across landscapes (Karanth et al., 2012;Miller et al., 2016b;Mpakairi et al., 2018).
Effective carnivore conservation integrates livestock management strategies as a means to reduce livestock predation and improve coexistence with people.It is particularly important because retaliatory killings can threaten the survival of the local carnivore populations (Inskip and Zimmermann, 2009;Dhanwatey et al., 2013).It is also important to recognize the role of large carnivores in maintaining intact ecosystems by regulating herbivore and meso-predator populations (O'Bryan et al., 2018).Our study uses predation risk model and addresses the knowledge gap regarding large carnivore (tiger, leopard, and dhole) conflicts across a very large southern Indian landscape, the biodiverse Western and Eastern Ghats of the Tamil Nadu state.The landscapes support a relatively high density of tigers, their co-predators, and mega-herbivore (Ramesh, 2010;Jhala et al., 2010Jhala et al., , 2019)).Despite consisting chiefly of lower-income local communities, and varying land-uses within and surrounding buffer areas of PAs, and experience intense HCC (Silori and Mishra, 2001;Davidar et al., 2008;Ramesh et al., 2019), this region has never been the subject of detailed HCC surveys.We conducted extensive household-level interviews across the human-wildlife interface and diverse land management systems in this region, including PAs (Tiger Reserves and Sanctuaries), RFs, and Fringe Areas (FAs).We collected reports of the tiger, leopard and dhole predation incidents through household-level interviews and mapped the number of livestock killed by large carnivores, at the household-level.We recorded the household characteristics and husbandry practices, along with ecologically important social and landscape attributes, that might potentially influence HCC.
The principal objectives of our study were: (1) to determine which important biotic and abiotic ecological variables best explain the spatial patterns of livestock predation; and (2) to identify the spatial extent and map conflict hotspot areas for each large predator species and all livestock predators combined.We predicted that households near the forest and those located inside forested areas would experience more conflict.We also expected that the loss of forest cover would make households more vulnerable to conflict, as a result of depletion of wild prey and increased availability of the domestic prey species (Treves et al., 2004;Khorozyan et al., 2017;Puri et al., 2020).We expected tigers to choose large-sized prey (e.g., cows, buffalo) whereas, leopards and dhole to go for smaller ones (e.g., goats, sheep) more frequently.We also hypothesized that the HCC would increase with an increased dependence of a household on the forest, as such livelihoods are linked to forest degradation in the human-wildlife interface areas.Finally, because the forest divisions of the Eastern Ghats are comparatively more fragmented than those of the Western Ghats in Tamil Nadu (AWIFS LULC data, Bhuvan, 2017), we predicted different patterns in the spatial extent, frequency, and species involved in HCC, and expected the former would be subject to more conflict.

Study area
The study area ranged from Cauvery North Wildlife Sanctuary of Tamil Nadu State in the north to Kanyakumari Wildlife Sanctuary in the south, which included 16 forest divisions, covering both interface areas of PAs, and RFs in the Western and Eastern Ghats (Fig. 1).The study area also includes FAs which covers up to 5 km from the forest boundary.The major large carnivores and omnivores found across the study area includes tiger (Panthera tigris), leopard (Panthera pardus), dhole (Cuon alpinus), striped hyaena (Hyaena hyaena), sloth bear (Melursus ursinus); their potential wild ungulate prey includes gaur (Bos gaurus), chital (Axis axis), sambar (Rusa unicolor), wild boar (Sus scrofa), muntjac (Muntiacus muntjak), and mouse deer (Moschiola indica).Common langur (Semnopithecus entellus), bonnet macaque (Macaca radiata), black-naped hare (Lepus nigricollis), porcupine (Hystrix indica), Indian giant squirrel (Ratufa indica) and peafowl (Pavo cristatus) are also among the other prey species found across the region.Major forest types encompass evergreen, semi-evergreen, moist mixed deciduous, dry deciduous, dry mixed deciduous, dry thorn, and riparian forests; the region also contains grasslands and wet shola grasslands (Champion and Seth, 1968).Average annual rainfall ranges from 600 mm in low altitude areas, to 2000 mm in the high-altitude areas, and average human population density varied spatially by a factor of ten, from 100 to over 1000 persons per km 2 (CIESIN, 2016).Elevation ranged from 200 m to >2600 m, and annual mean temperature ranged from 5 to 40 C. The size of forest divisions ranged in area from as small as 322 km 2 , to the largest at 1409 km 2 .The study area contains largely forest-dwelling communities, who depend on agriculture, harvesting non-timber forest products (NTFP; e.g., roots, tubers, honey and firewood), and the rearing of livestock (e.g., cattle, goat, and sheep), for their livelihoods (Davidar et al., 2007(Davidar et al., , 2008)).The 16 forest divisions are subjected to various developmental pressures, including the construction of hydro-electrical power stations, and the conversion of natural forests for the cultivation of tea, coffee, eucalyptus and wattle.Besides, the killing of large mammals by electrocution, snares, and the poisoning of livestock carcass are some of the other major threats to native wild fauna in these landscapes (Kalaivanan et al., 2010;Arumugam, 2012).Although herding of livestock within PAs are prohibited, illegal grazing in the forest patches, particularly near water bodies are prevalent.Herders frequently leave livestock unattended to graze during the day, sometimes guarded by dogs, and herds them back to the villages in the evenings.

Semi-structured interview sampling
To systematically sample households as to their recent experiences with HCC, we overlaid a grid of 5 km Â 5 km squares (25 km2) across a map of our entire study region.Between November 2017 and March 2019, we conducted a semi-structured questionnaire (Fig. 2; Appendix S3) of at least three independent households per grid, maintaining a distance of approximately 0.5e1 km between the households with varying distance from the forest.In total, we sampled 485 5 km 2 grids.More number of grids were sampled depending on the more villages and proportion interface area present in each forest division.Also, the villages were spatially spread across the entire study area.Since the analysis was done at a household-level and that each household reported multiple cases of conflict incidences, the GPS point of a representative household also was noted.We targeted an adult household member (>18 years old) who has lived there for more than five years and is aware of or remember the incidents of carnivore attack on their livestock in each forest divisions.He/she served as the household representative for our interview, which was in the regional language of Tamil.All interviewees agreed to be interviewed, representing their household.Before beginning the interview, we searched for available information on livestock predation in the local media and interviewed forest officials to select the household areas of the village prone to livestock predation for maintaining the authenticity of information represented and reported correctly.Firstly, we went through the compensation data on livestock predation and injury/death to human records available at the forest range office in Tamil Nadu.However, we found that most of those data were incomplete and that the reports of ground-level conflict incidences were left unrecorded.Most of the local villagers were not interested in reporting their livestock loss to the Forest Department due to the lengthy, time-consuming procedures, in addition to the severe delay in receiving compensation.Besides, most of their livestock grazing were illegal, and hence they were afraid to report their loss to the Forest Department because of the fear of penalty.Insufficient data availability from the Forest Department prevented us from further validation.
Each interview began with an introduction, explaining the background of our study.We then recorded the age and sex of the respondent's family members, education of household members and their literacy rate (derived based on of the number of people in a household who read and write in Tamil/English), livelihood dependence, the total number of household livestock, and GPS location of the household during the survey.We ultimately conducted 1460 questionnaire surveys across an area of 14,200 km 2 to collect detailed records and data on large carnivore (e.g., tiger, leopard and dhole) conflict incidents and records over the past five years.In the initial stages of data collection, we visited different forest divisions, tested whether locals were able to recollect the number of conflict incidences and identify the conflict species.Since the study area contains largely rural and forest-dwelling communities, most of them had experienced wildlife conflict incidents sometime in their life.Therefore, recollecting the conflict incidents and identifying the conflict species was an easier task for them.Additionally, we showed them the images of the carnivore species, carnivore kills and other indirect signs (track, scat, scratch, rake, and signs on the carcass) from the identification material we carried, to make the identification easy and accurate.We explained the distinct differences in evidence associated with the behaviour of each predator while attacking their natural prey and livestock.For example, felids primarily direct their initial attacks on to the head, face, and neck, whereas, dholes attack the flanks and hind legs in packs (Ramesh, 2010).For distinguishing the age of the livestock predated, we considered any cattle less than a year old as calves and greater than that as an adult cattle, as such distinction made it easier for the locals to understand.All records were based on the available evidence and witness accounts.It included: the time of attack, number and species of livestock killed or injured, location of the attack and nature of the interaction between community and predator, and the large carnivore species involved.As actual timing for livestock predation can be tough to recollect, we used coarser intervals to assign the time of the predation events (Morning (6 a.m.-12 p.m.), Afternoon (12 p.m.e4 PM), Evening (4 p.m.e8 PM), and Night (8 p.m.-12 a.m.).Data of less than 2% of cases were excluded for further analysis as those interviewees were uncertain about the species identity.However, we acknowledge the limitations of the interview-based surveys and predator identification that utilize recollected information.

Predictor variables
Our goal was to prepare a possible fine scale general livestock predation map for a large landscape across the state to identify hotspots for policy level decision making and management purposes, for which we analyzed the conflict frequency data at the household level.Due to restriction in the forested areas, most of the livestock are left grazing at closer distances (ca.around 1e2 km) from household, therefore we selected 1 km 2 for spatial variable extraction.Also, the respondents reported livestock predation largely happens within 1e2 km around the household where livestock are kept.Accordingly, we extracted only information on spatial variables at 1 km spatial resolution as most of the spatial variables such as slope, elevation, annual mean precipitation, temperature and human density available at that scale and then to each of the household sampling point.
Using ArcMap 10.3, we overlaid subsampling grids (1 km Â 1 km) across our sampling area.We obtained land cover layers for 2015e2016 (AWIFS LULC data, Bhuvan, 2017) that had 18 classes: "built-up" or developed, kharif crop, rabi crop, zaid crop, double/triple crop, currently fallow, plantation, evergreen forest, deciduous forest, degraded/scrub forest, littoral swamp, grassland, wasteland, water bodies, mining, and others.For our study, we reclassified them into the following 11 classes: "built-up", crop, fallow, plantation, evergreen forest, deciduous forest, degraded/scrub forest, littoral swamp, grassland, wasteland, and water bodies, using the reclassify tool in ArcMap 10.3.We then integrated the following natural land-cover vegetation classes: evergreen forest, deciduous forest, and degraded/scrub forest, collectively defined as "natural forest cover".We extracted only information on spatial variables at 1 km resolution and then to each of the household sampling points.We calculated the percentage of forest cover and crop cover in each 1 km Â 1 km grid separately by multiplying the no. of pixels with forest and crop to total no. of pixels in a grid.Before calculating the differences, we calculated the percentage of forest cover and crop cover separately from both 2006 and 2016 layers.We measured natural forest cover loss & gain, and crop loss & gain, by calculating the difference between 2006 and 2016 layers at 62 m spatial resolution using a 1 Â 1 km grid (AWIFS LULC data, Bhuvan, 2017), using ArcMap 10.3 (ESRI, 2014).We then assigned these variables to the household sampling points falling within the respective grid.
We further extracted other spatial variables such as slope, elevation, annual mean precipitation and temperature from WorldClim (Hijmans et al., 2005) and human population density (CIESIN, 2016) at 1 km spatial resolution and then to each of the household sampling points falling within respective 1 Â 1 km grid.Using the Euclidean distance tool in ArcMap 10.3, we measured the distance to forest area boundary (Tiger Reserve, Sanctuary and RF), natural forest cover, water body and road, for each household.We characterized forest dependency by assigning higher scores to households dependent on more categories of forest use, including NTFPs, firewood collection, livestock grazing, and miscellaneous.The forest dependency categories weren't quantified values.If the person is dependent on one source of income then that family is most likely to get revenue from various other activities apart from forest resources.Hence, the score was assigned based on the number of representations which reflect the ground reality of sociological variable and people's dependency across a vast landscape.We also assigned a score of "3" for villages located inside PAs, "2" for those located in RFs, and "1" for villages located in the FAs; higher score represents greater exposure of the village to HCC, partially a reflection of an expected higher relative predator abundance or occupancy in these habitats.Potential domestic preys were classified into three major groups based on body size: cow-buffalo, goat-sheep and domestic dog.Livestock abundance was calculated as the number of livestock per household.

Data analysis
We conducted chi-square (c 2 ) tests to compare overall frequencies of livestock loss (i.e., no. of livestock predation cases) at household-level among different carnivore species, management regimes or land protection status, and for each time (Morning, Afternoon, Evening and Night).To identify spatial patterns of large carnivore conflict, we considered the number of livestock loss from each carnivore species as the response variable at household-level.Using a Generalized Linear Model (GLM) approach, we examined the importance of the following potential predictor variables in a GLM (Table 1).To assess the likely predictors of predation risk for each carnivore (tiger, leopard and dhole) and develop a predation risk probability map, we fitted GLMs assuming a Poisson distribution in program R version, 3.0 (R Development Core Team, 2018).There was evidence of overdispersion for leopards and dhole data.So, a GLM with a negative binomial family (theta ¼ 1) (Hilbe, 2011;Cameron and Trivedi, 2013) was used to control for potential overdispersion of leopard and dhole data.
Before model construction, we first checked for multiple correlations among predictor variables using a hierarchical Pearson correlation co-efficient test (Graham, 2003) conducted in the 'corrplot' package of program R, version 3.0 (Wei and Simko, 2017).We removed highly correlated (r > 0.50) variables (in this case -crop cover, crop change, distance to natural forest cover, and temperature) for further statistical analyses (Figure S1.1).Null deviance and residual deviance values were used to assess model fit.We used an information-theoretic approach to evaluate model strength of evidence and fitness based on Akaike's Information Criterion (AIC c ) corrected for relatively small sample sizes; we evaluated relative AIC differences (D) and weights (wi) (Burnham and Anderson, 2002) to identify models best explaining predation predictor variables.Using for final inferences (Burnham and Anderson, 2002).All analyses were performed using Programme R Package, version 3.0 (R Development Core Team, 2018), other packages like MASS (Venables and Ripley, 2002), rJava (Urbanek, 2010), glmulti (Calcagno and de Mazancourt, 2010) and MuMIn (Barto n, 2013).Finally, we used the R package 'effects' (Fox et al., 2014) to plot the response curves of the top model covariates for the dependent variable.

Predictive spatial patterns of conflict
Using the package 'gstat' (Pebesma and Graeler, 2019) in Programme R Package version 3.0 (R Development Core Team, 2018), we used a spatial interpolation (Kriging) to infer or "predict" incidents of livestock predation in unsampled areas based on data from our top model for the probability of livestock loss.This approach accounted for uncertainty as distance increased between spatial locations and yielded a semivariogram describing the spatial correlation between the points (Appendix Figure S1.2;S1.3; S1.4).Several statistical distribution models, including spherical, circular, exponential, Gaussian, Matern, and M. Stein's parameterization, were fit to the data to test for their suitability for each species.We applied Matern and linear family distribution models to the overall data set for all species combined.Employing an automatic interpolation method, we found that Gaussian, M. Stein's parameterization, and spherical distribution models, explained the patterns of conflict for tigers, leopards and dholes, respectively.Using the package 'automap' (Hiemstra et al., 2009), we then generated maps of conflict patterns of livestock predation risk for respective predators across the entire landscape, including unsampled grids.We calculated all weighted sum values predicting livestock predation risk for all three carnivore species using the Spatial Analyst tool in ArcMap 10.3 to estimate cumulative predation risk.To compare predation risk overlap metrics for predation risk maps, we used the package 'spaa' (Zhang, 2015) to quantify Schoener's niche overlap metric (D), which ranges from 0 (i.e., niche models have no overlap) to 1 (i.e., niche models are identical).The total predicted frequency of predation risk for each predator were divided into three categories in equal proportion (low, medium and high), based on which areas were identified and reported accordingly.

Households, livestock loss, and predation risk
Forty-eight percent (n ¼ 694) of households reported experiencing HCC, or the loss of livestock due to large carnivores.Of these, 78% suffered more than one incident of livestock predation; we recorded 866 predation events, of which two-thirds were by leopards (n ¼ 576), approximately 19% by dholes (n ¼ 165), and 14.4% by tigers (n ¼ 125).We found that households inside RFs were the most likely to report livestock losses (71%), followed by PAs (58%) and lastly, FAs (42%).The percent of livestock depredated by each carnivore species varied significantly across management systems (c 2 ¼ 15.39, df ¼ 4, P < 0.005).Among carnivores, those households reporting livestock loss by tigers were highest inside PAs (25%); RFs (14%) and FAs (5%) experienced less human-tiger conflict.Households reporting the highest rates of livestock loss due to the predation of leopard and dhole, occurred in RFs (62% and 29% respectively), followed by PAs (44% and 11% respectively) and FAs (36% and 9% respectively).Seven hundred and ninety-one families (54.2%) were engaged in at least one source of forest-dependent practice for their livelihood (e.g., NTFP, firewood collection, livestock grazing & miscellaneous).Among them, people living inside RFs (91%) and PAs (70.1%) were much more dependent on forest resources for their livelihood compared to people living amidst the FAs (47%).Households from RFs reported keeping more goat/sheep (58%) than those from PAs (40%) and FAs (42%).Similarly, more households in RFs reportedly kept cow/buffalo (77.5%) than those from PAs and FAs (51% each).The Human population density ---9 Annual mean precipitation ---10 Distance to forest boundary - proportion of households owning domestic dogs were also highest in RFs (60%), followed by FAs (44.6%) and PAs (42%).Finally, the respondent reported only four fatal attacks on humans by large felids in our study area.
We found that the following major variables influence were associated with risk of livestock predation by the three large carnivores in our study: (1) forest dependence, (2) village location, (3) natural forest cover loss, (4) number of cow & buffalo per household, (5) number of goat & sheep per household, (6) the number of domestic dogs per household, (7) distance to water, (8) distance to forest boundary, (9) elevation, (10) annual mean precipitation and ( 11) slope (Table 2; Table 3 3).Livestock predation risk probability by tiger alone was best explained by a model < D2AIC c (wt ¼ 0.63) that included six variables: (1) higher relative abundance of cow/buffalo, (2) higher dependence of people on forests, (3) greater proximity to water sources, (4) shorter distance to forest boundaries, (5) relatively higher elevations and (6) villages located within the forest areas, particularly in core areas of PA's (Appendix Figure S1.5).The best model predicting livestock predation risk by a leopard (<D2AICc; wt ¼ 0.83) included (1) relatively higher abundance of goat/sheep, (2) relatively high abundance of domestic dog, (3) higher dependence on the forest by households, (4) an increase in topographical slope (Appendix Figure S1.6).Finally, for the top model predicting livestock predation risk from dholes (wt ¼ 0.71), we found that (1) greater forest cover loss, (2) increase in the relative abundance of goat/sheep, (3) greater dependence on the forest by households, and (4) relatively lower precipitation, were all important (Appendix Figure S1.7).The cumulative predation risk showed that the HCC was very localized to a few regions (Fig. 4).

Identified areas of predation risk
We identified several areas of relatively high tiger predation risk within the Nilgiri Biosphere Reserve, including large parts of the Nilgiri FD, Mudumalai and Sathyamangalam covering approximately 125 km 2 ; we also identified another cluster of predations across an area of 50 km 2 in the Megamalai Wildlife Sanctuary (Fig. 3a; Appendix Figure S1.8).For leopards, we identified many more high predation risk hotspots, including parts of the Sathyamangalam Tiger Reserve (60 km 2 ), and the    T. Ramesh, R. Kalle, D. Milda et al. Global Ecology and Conservation 24 (2020) e01366 Nilgiri eastern slopes of Mudumalai Tiger Reserve (50 km 2 ), parts of Srivilliputhur Wildlife Sanctuary (180 km 2 ), Megamalai Wildlife Sanctuary (90 km 2 ) and a few areas Nellai Wildlife Sanctuary (40 km 2 ) (Fig. 3b).For dhole, we identified high predation risk areas in the Nilgiri eastern slopes of Mudumalai (30 km 2 ), parts of Sathyamangalam (60 km 2 ), and large areas in Srivilliputhur (200 km 2 ) and Megamalai (140 km 2 ) (Fig. 3c).Interestingly, areas of predation risk by tiger had less spatial overlap with leopard (0.53) and dhole (0.51), whereas spatial overlap between leopards and dholes was high (0.89).Our cumulative predation risk map showed that certain sections of the Nilgiris, eastern slope of Mudumalai, parts of Sathyamangalam, Megamalai and Srivilliputhur, are the most important hotspot areas for all large carnivores combined (Fig. 4).Predation risk modelling (PRM) suggested that households located closer to and within forested areas have a higher risk of conflict with large predators, with the highest risk levels inside and adjacent to PAs, and the lowest in agricultural areas along the fringes.Also, our study clearly showed that there is variation in the spatial pattern of conflict clusters across the landscape according to the "predation niche" of each species.

Discussion
The identification of those factors associated with or potentially causing livestock predations by large carnivores is becoming an increasingly important tool for their management and conservation around the world (Naha et al., 2020).Our analysis indicated that variables such as human dependence on the forest, the location of a village, loss of forest cover, domestic livestock abundance, proximity to water resources and forest boundaries, topography, and the amount of precipitation in combination explained the HCC risk patterns in human-wildlife interface areas of the Western and Eastern Ghats complex of Tamil Nadu.Although some spatial risk patterns for state-wise analyses can also be obtained from secondary sources (Athreya et al., 2015), our predictive spatial risk maps showed the existence of carnivore predation risk gradients across almost 22,525 km 2 , of which 14,200 km 2 (approximately 63%) of the total area of inference, was sampled.Although we have relied on a single source of local's reporting on human-large carnivore conflict the large predictive predation risk map helps in asserting the relative conflict status of sympatric large carnivores for major PAs and RFs networks of the Western and Eastern Ghats of Tamil Nadu.Our conflict hotspot map also highlighted areas prone to high livestock loss from large predators.
A large number of households surveyed across such a large landscape also minimizes sampling bias and represents a greater sampling effort and a better coverage than other studies (Karanth et al., 2012(Karanth et al., , 2013a(Karanth et al., , 2013b)).Interestingly, relative to other parts of India (Prashant, 2004;Dhanwatey et al., 2013;Kshettry et al., 2017), there were only a very few cases of human casualties from large felid attacks in the study area indicating a general absence of direct HCC, attributed to the high abundance of domestic and wild prey species.Therefore, the large carnivore conflict is predominantly related to livestock predation in our study area.

Patterns of livestock predation and area management
Our results indicate that the patterns of livestock predations by large carnivores varied across management systems and land use type, as well as inter-species variation in resource use (e.g., habitat, livestock type).Relative to tigers, our study reported much higher livestock predation rates by leopards and dholes in RFs and associated buffer zones.Our data also suggests that the risk of livestock predation from tigers across Tamil Nadu PAs is comparatively lesser than in the parts of Central India (Karanth et al., 2012;Miller et al., 2016b) and Uttarakhand (Bargali and Ahmed, 2018).The scale of livestock predations by the leopard, however, was comparable to that endured by other PAs elsewhere (Karanth et al., 2013a(Karanth et al., , 2013b;;Athreya et al., 2015).Overall, livestock predations occurred mostly in the morning hours, consistent with regional carnivore activity patterns (Ramesh et al., 2012b) and overlapping times of peak livestock grazing activity.
Tiger predated more livestock, mainly adult cow/buffalos, from households located inside PAs compared to RFs and FAs.It is most likely to be attributed to the higher abundance of tigers in core areas of PAs, which are more densely forested and generally subject to less anthropogenic influence (Ramesh, 2010;Kalle et al., 2011;Jhala et al., 2019).Those tigers principally depredated cow/buffalos, likely due to a preference for larger prey (Ramesh et al., 2012a).The overall higher livestock loss in RFs compared to PAs and FAs may be due to the disproportionately cumulative impact of predation patterns by leopard and dhole.Other studies in India (Athreya et al., 2007;Ramesh et al., 2012aRamesh et al., , 2012b;;Kshettry et al., 2017) and elsewhere (Khorozyan et al., 2017) corroborate the selection of smaller-bodied livestock species by leopards, particularly along with the human-wildlife interface where a lower abundance of natural prey may be expected (Pillay et al., 2011).
Leopards in our study region predated dogs at a lower rate than goats and calves.It was lesser when compared to the rate at which they took dogs in more urbanized landscapes (Edgaonkar and Chellam, 2002;Athreya et al., 2016).The dogs lost to leopard predation were ironically mostly livestock herding/guarding dogs of agro-pastoral communities, a pattern consistent with other studies (Butler et al., 2013;Khorozyan et al., 2017).Every household kept at least one or two domestic livestock guarding dogs, which roamed freely when sheep or goats grazed in RFs and FAs.Livestock predation by dhole in our study was comparatively lower than around select PAs of northeast India, which appears to experience higher livestock loss by dholes relative to large felids (Lyngdoh et al., 2014); this may be due to differences with Eastern Himalayan landscapes in large felid density as compared to the Ghats landscapes of southern India (Jhala et al., 2010(Jhala et al., , 2019;;Kalle et al., 2011).
Livestock loss patterns varied among forest divisions with land-use management, local dependence on the forest, and livestock grazing regulations.For reducing livestock predation risk inside PAs, we need better measures to manage livestock grazing.The predation risk map can be shared with forest rangers to inform locals to avoid areas of high conflict.This approach might build trust and relationships between forest staff and locals.It also demands improved livestock management practices, such as the implementation of livestock stall-feeding near core areas (Goodrich, 2010;Karanth et al., 2012;Miller et al., 2016b).Our findings suggest that species-specific variation in large carnivore conflict patterns.

Landscape effects on spatial patterns of livestock predation
The integration of landscape attributes to produce large scale carnivore conflict risk maps improves the development of conflict mitigation measures (Karanth et al., 2013a(Karanth et al., , 2013b)).Our model clearly showed that the risk of attacks on cattle by a tiger was highest near water sources and forest boundary areas, and for villages located inside the forested core areas of PAs; tiger predation risk decreased towards FAs.This is consistent with predation patterns exhibited by tigers in forested areas elsewhere across the region (Wang and Macdonald, 2006;Karanth et al., 2013a;Soh et al., 2014;Miller et al., 2016b).Other authors also found that communities illegally grazing free-ranging livestock inside PAs led to an increase in livestock predations by tigers in core forested areas (Jhala et al., 2010(Jhala et al., , 2019;;Ramesh et al., 2012b).For tigers, the findings of Karanth and Kudalkar (2017) were comparable to ours in the Central Indian landscape; they found that livestock predations increased at higher elevations (>2000 m), whilst those by leopards increased in gentle slope areas.The topography is also influential in predicting large felid kill risk patterns at varying spatial scales (Rostro-García et al., 2016).
Dhole predation risk increased with increasing forest cover loss and in areas of lower precipitation characterized by open scrub-jungle habitats (Pillay et al., 2011).Forest cover loss also influenced local extinction probabilities of dhole in select PAs of the Western Ghats in Karnataka (Srivathsa et al., 2019a).Reduction in forest cover can lead to a reduction in habitat quality and low availability of wild prey (Puri et al., 2020) increases the conflict and retaliation effect, which subsequently increases the extinction probability of the species.Therefore, the reduced natural prey biomass in low precipitation areas may have contributed to increasing livestock predations by dholes as elsewhere (Kolowski and Holekamp, 2006;Goodrich, 2010;Mukeka et al., 2019).
Overall livestock risk patterns suggested cattle were more vulnerable in dense forest areas whereas, goat and sheep were more vulnerable to predation in open vegetation, a pattern also reported in Central India's PAs (Miller et al., 2016b).Smaller livestock, such as goat and sheep, often graze in large herds in the dry scrub open habitats in our landscape, which was consistent with predation patterns exhibited by both leopards and dholes.Increased livestock grazing adjacent to forest patches is likely one reason for high conflict closer to PAs.In forested FAs, livestock sharing the same water sources as wildlife may contribute to livestock predations.Also, species-specific predation patterns reflect ecological differences among large carnivore species, including habitat and spatial partitioning (Rostro-García et al., 2016), inter/intra-guild competition for domestic prey and variation in selection of domestic prey size (Ramesh et al., 2012b), and landscape-scale attributes (Puri et al., 2020).Leopard and dhole, for example, may occupy suboptimal habitats or habitats where there is less chance of encountering tigers; this spatial avoidance among co-predators may force leopards and dholes more often into forest edges, PA boundaries, and buffer zones (Seidensticker et al., 1990;Ramesh et al., 2012b).

Livestock predation patterns in southern India forest networks
HCC is a recurring problem across this forested, southern India landscape for several indispensable reasons.The absence of alternate grazing areas for most households, the location of households inside or very close to forested areas, and illegal grazing actually inside restricted use landscapes, often occur alongside other land uses and livelihoods (Karanth and Nepal, 2012;Karanth et al., 2012).Although livestock predation patterns varied by carnivore species, the overall average livestock loss for all carnivores (0.71 head/household/year) we recorded was much higher than for Ranthambore National Park (0.38 head/household/year), Kanha National Park (0.16 head/household/year) and Nagarhole National Park (0.27 head/household/ year) (Karanth et al., 2013b).However, our livestock predation rates were less than Sariska Tiger Reserve (26.5% of the households affected in three years) (Sekhar, 1998), Kaziranga National Park (1.2 head/household/year; Borah et al., 2018), the Valparai Plateau in the Anamalai Hills (3.6 head/household/year; Sidhu et al., 2017), the Trans-Himalayas (1.6 head/household/year; Mishra, 1997) and select PAs of Bhutan (1.29e3.26 head/herd/year;Wang and Macdonald, 2006;Jamtsho and Katel, 2019).We recorded a higher rate of livestock loss in PAs and RFs than in FAs suggests that livestock protection measures are warranted to address HCC across the various land management systems.
We identified major tiger predation risk hotspot areas covering parts of Nilgiri FD, Mudumalai and Sathyamangalam, where tiger abundance is relatively high in Tamil Nadu and contains a major source population relevant to adjoining humanwildlife interface areas.Aside from this major cluster, we identified another smaller tiger conflict area near Megamalai Wildlife Sanctuary.All of these interface areas also had a high abundance of buffalo and cows.We, of course, identified many more leopard predation risk hotspot areas, principally where the tiger conflict reports were lower, as well as a few medium risk areas for dhole conflict in Dharmapuri and Hosur.In areas where conflict patterns of both leopard and dhole overlapped, the risk to livestock predation was even greater.Despite high predation risk, people were less likely to report compensations to the Forest Department when smaller livestock (e.g., goats, sheep) were killed by leopards and dholes.Very often there would be a lack of field evidence (e.g., carcass remains, tracks and sign, photographic proof of large carnivore species attacking the prey) to provide as proof to forest managers for verifying claims.However, households were more likely to report largebodied livestock loss (e.g., cows and buffalo), especially from tigers.Other authors observed a similar lack of reporting livestock losses in PAs of Madhya Pradesh and Karnataka states (Karanth et al., 2012(Karanth et al., , 2013a)).We also acknowledge the limitations of the interview-based surveys as recollecting ability varies among respondents which can affect reliable reporting of actual number of livestock predation incidents.Overall our maps revealed the highest risk levels in forest patches present inside and adjacent to PAs, and the lowest risks were in agricultural areas in FAs.It is evident from our findings that successful management of human-large carnivore conflict will require state-level planning which incorporates conflict issues from RFs, FAs, and PAs at the landscape level.

Management recommendations
Avoiding predation risk hotspot areas where livestock is most vulnerable to predation, to different carnivore species can be a critical initial step to reducing HCC.Our map provides a powerful visual tool for communicating patterns of carnivore predation risk at the landscape level to stakeholders.Updating predation risk maps regularly can provide up-to-date conflict scenarios, which be very useful for policymakers when allocating physical and financial resources across various conflict scenarios.Such resources might include livestock insurance or compensation schemes, land-use zoning, and grazing restrictions in high predation risk hotspots.Hence, future studies should take into account conflict information from multiple sources.
Ultimately our results are relevant to global strategies to mitigate HCC and suggests that livestock predation can be reduced by addressing ecological factors associated with multiple carnivore species, as well as multiple regional and local social factors.For instance, diverse perceptions of local communities towards the conflict animal can have a massive impact on the conservation strategies of species in a region.Hunting patterns differ among large carnivore species thereby reflecting on the species-specific conflict intensity and in some cases spatial separation in a region.However, given the increasing numbers of conflict cases, mitigation efforts designed for single conflict species might not help much in the future if conflict incidents by multiple species are high.Our spatial predictive risk maps depict the complex multi-species conflict challenges along with the human-wildlife interface areas through maps that aid PA managers to develop novel multi-species mitigation strategies across a very vast landscape for quick decision making.Due to the size of the region, our analyses primarily relied on self-reporting but, we have not monitored actual incidents of conflict, we recommend active, location-based household monitoring each year to validate predation kills sites at a fine scale, develop effective carnivore conflict mitigation measures, and evaluate their overall effectiveness.We further recommend benefit-sharing options to local communities, such as those realized through local tourism, as well as other regional income generation, such as energy sources (i.e., biogas, solar power, etc.), which can build support locally for conservation initiatives (Dhanwatey et al., 2013).We also believe that the regional and cultural affinity of the villagers towards the wildlife can be utilized as a means to minimize the negative impacts in the human-wildlife interface areas (Karanth et al., 2010).
Our study also revealed that large carnivore predation risk was related to the dependence on forest resources by locals across the landscape.Although monetary compensation by the state Forest Department is likely to minimize livestock losses from tigers or leopards, delays in processing claims and a relatively lengthy process in providing funds can compromise trust and undermine state and other conservation efforts designed to benefit species conservation.More efficient processing of claims and quicker issuance of compensation is therefore critical.We note that care should also be taken to the disposal of livestock carcasses, to prevent their misuse in poisoning predators.Deployment of improved livestock husbandry practices, as well as more effective preventative measures, tools, and strategies, could also have a transformative impact and significantly reduce household predations: predator-proof enclosures, well-trained livestock guarding dogs, livestock carcass removal and processing programs, the use of physical deterrents (electric fencing/shock, acoustic repellents and bio-fences, chemical fences, shock collars, etc.), stall feeding of calves and lactating females, and use of experienced shepherds (Inskip and Zimmermann, 2009;Rossler et al., 2012;Ravenelle and Nyhus, 2017), are among the potential solutions that can be implemented solo or in combination with others.A decentralized village-level insurance scheme, where funds kept in targeted villages to manage on their own (Mishra et al., 2003;Khan et al., 2018), may also offer benefits and foster community responsibility and stewardship.
Previous studies in Central India (Karanth et al., 2013b) and East Africa (Kolowski and Holekamp, 2006) found that trained guard animals and fencing were useful in mitigating HCC.Trained livestock-guarding dogs and other suitable guard animals can be among the most effective methods for reducing predations.They serve as shepherds, preventing flocks from fleeing and scattering; they also stop predators from directly interacting with livestock (Allen et al., 2017).However trained guard dogs should be housed indoors, and not left freely in wild areas as free-ranging dogs in forested areas predate on wild prey, thereby increasing competition for wild carnivore species and increase the risk of disease spread (Suryawanshi et al., 2013).Taste aversion has been successfully used on the bait to repel canid species from a specific prey type (Shivik et al., 2003), and could hold potential; to date, however, its effectiveness in India remains unexplored.Similarly, shock collars have been successfully used on livestock to keep away wolves (Schultz et al., 2005;Gehring et al., 2006;Hawley et al., 2009;Rossler et al., 2012).Anhalt et al. (2014) used bio-fences with scent marks or sounds, which by projecting the presence of a competing species, can deter territorial individuals, or force residents to move away from a demarcated area.
Other potential mitigating methods through indirect and direct practices might make communities more tolerant of local carnivore populations.Habitat conditions can be improved by reducing human pressure and enhancing wildlife protection measures to increase the wild prey populations, particularly in areas of low natural prey availability, to potentially reduce the effect on livestock.For instance, dynamic spatial risk maps could be generated in real-time by developing software integrated mobile technology as a user-friendly application and a participatory approach by local pastoralists to provide timely details on livestock killed/injured by large carnivores.This technological application would serve as a warning system to inform local beat guards and range officers when any livestock predation incident occurs and also alert neighbouring households through probabilistic cluster models or animal path analytics.Through the inclusion of camera-trapping and radio-collaring of the conflict individuals and the livestock, pastoralist and predator movements could be tracked thereby the potential conflict areas can be foreseen and mapped, thus allowing resource allocation (emergency vaccination, treatment, provision of feed resources, and provision of relief materials to injured people and animals).As a result, the timely pro-active site-specific conservation actions could speed-up compensation claims, minimize ambiguity in predator identification and reduce any negative attitude towards park managers/predator.Interactive, participatory mapping can also help strengthen outreach and awareness efforts regarding the unique nature of local shared spaces and regional land-use patterns, as well as indicate priorities for the location and potential type of ground-level actions needed to mitigate predation.Diversifying local economic livelihoods, and benefit-sharing options for local communities can minimize forest dependency.Further site-specific monitoring studies are necessary for those high conflict areas identified for an in-depth investigation of the current husbandry practices while dealing with conflict animals.

Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Fig. 1 .
Fig. 1.Study area showing the questionnaire survey points in the Western and Eastern Ghats parts of Tamil Nadu.

Fig. 2 .
Fig. 2. Livestock pressure in the study area: Researcher interviewing a herder (a), livestock herded to the forest (b, c) from human-wildlife interface area and livestock utilising the water source inside the forested area (d).

Fig. 3 .
Fig. 3. Predation risk maps showing predicted frequency (No. livestock killed for the last five years) of livestock depredation from tiger (a), leopard (b) and dhole (b) in the Western and Eastern Ghats parts of Tamil Nadu.

Fig. 4 .
Fig. 4. Predation risk map showing the cumulative predicted frequency (no.livestock killed for the last five years) of livestock predation in the Western and Eastern Ghats parts of Tamil Nadu.

Table 1
Variables considered for predictive risk models explaining factors influencing predation risk by tiger, leopard and dhole in the Western and Eastern Ghats parts of Tamil Nadu.

Table 2
Summary of AICc model selection Generalized Linear Models explaining factors influencing predation risk by tiger, leopard and dhole in the Western and Eastern Ghats parts of Tamil Nadu.

Table 3
Estimated beta coefficients for the top ranked models that explains the probability of explaining factors influencing predation risk by tiger, leopard and dhole in the Western and Eastern Ghats parts of Tamil Nadu.
Cow_buff -abundance of cow/buffalo, DistForestBdry -distance to forest boundary, DistWater e distance water source, Domesticdogabundance of domestic dog, ForestDepScore -local's dependence on forest, ForestCovLG e percent forest cover loss and gain, Precipprecipitation, Slope e mean slope, VillageLoc -village location.