N-mixture model-based estimate of relative abundance of sloth bear (Melursus ursinus) in response to biotic and abiotic factors in a human-dominated landscape of central India

Research permissions and ethical considerations

Madhya Pradesh Forest Department (MPFD) issued all required permissions for our field surveys (Letter No: I/3129 dated 29/05/2014). Due to the non-invasive nature of sampling, we did not require ethical clearance for this study.

Study site

Sanjay Tiger Reserve (STR) is situated between 24.125122°N 81.546009°E (northwestern boundary) and 23.822641°N 82.175139°E (southeastern boundary) in the state of Madhya Pradesh in central India (Fig. 1). STR falls under the category “Central Highlands” as per the biogeographic classification of India (Rodgers & Panwar, 1988). The core and buffer areas of STR encompass 831.25 km² and 812.58 km², respectively. Furthermore, the core area of STR consists of two administrative parts, i.e., Sanjay National Park (SNP; 466.60 km²) and Dubri Wildlife Sanctuary (DWLS; 364.60 km²).

The terrain of STR is rugged and undulating, with plateaus and gorges in the eastern part (SNP) and flat land in the western part (DWLS). STR has an elevation range between 425 and 732 m. The broad forest type of STR falls under sub-group 3C- “North Indian moist deciduous forests”, with subtype (C2) “Moist sal-bearing forest” with subdivision (2e) “Moist peninsular sal forest” (Champion & Seth, 1968). Sal (Shorea robusta) is dominant in SNP and DWLS, occupying about 80% of the entire area, followed by mixed, bamboo mixed, scrubland and grassland (Fig. S1). Apart from sal, other frequently found tree species include tendu (Diospyros melanoxylon), mahua (Madhuca longifolia), char (Buchanania cochinchinensis) and sedha (Lagerstroemia parviflora). STR receives a mean annual precipitation of 1,303 mm between June and September, and the annual temperature varies between 7.4 °C and 41.8 °C. Other than the sloth bear, tiger (Panthera tigris), leopard (Panthera pardus), striped hyena (Hyaena hyaena), Indian wolf (Canis lupus) and Asiatic wild dog (Cuon alpinus) are the major carnivores found in STR. Anthropogenic pressure is immense inside the core area of STR due to the presence of thirty-nine villages.

Camera trap survey

The camera trapping survey was conducted from December 2016 to April 2017 in the core area of STR. We divided the study area into 2 km × 2 km grids (henceforth, sites) following the All India Tiger Estimation protocol (Jhala, Qureshi & Gopal, 2015) and subsequently deployed camera traps in 143 locations (one at each site). We divided the core area into two blocks (SNP and DWLS), and camera traps were deployed in one block at a time due to the limited availability of camera traps. A pair of motion-triggered camera traps (Cuddeback C1 and Cuddeback Ambush) were deployed at each site opposite to one another to increase the detection (Pease, Nielsen & Holzmueller, 2016; O’Connor et al., 2017) at an approximate height of 30–40 cm above the ground. We placed the camera traps on forest roads, animal trails and dry riverbeds to maximize the photo-capture of sloth bears and other large carnivores (Karanth et al., 2011). The average distance between two adjacent camera trap locations was 2 km. We set 15 s time-lapse between two consecutive photographs. Cameras were active 24 h a day for an average trapping period of 42 days (range: 13–45 days) and checked once every 20–25 days to change batteries and memory cards. We collected the locations of camera traps by using Garmin e-Trex handheld GPS units (Garmin Inc., Olathe, KS, USA).

To avoid pseudoreplication, we considered two consecutive photographs of sloth bears or humans captured ≥30 min apart as independent events (O’Brien, Kinnaird & Wibisono, 2003; Xiao et al., 2018; Kafley et al., 2019). We set the temporal sampling unit (sampling occasion) as three days. We pooled the count of independent photographs of sloth bears from each site for each unit of three consecutive days, which yielded an encounter history of 15 sampling occasions. Reducing the sampling period by pooling data from a specific time (days) helped reduce overdispersion (Penjor et al., 2019).

Covariates

Relative abundance model of sloth bear

We developed the N-mixture model (Royle, 2004; Kéry & Royle, 2015) to estimate the relative abundance of sloth bears in STR within 3.52 km × 3.52 km grids, as this represents the minimum home range (i.e., 12.40 km²) of sloth bear in central India (Yoganand, 2005). However, little is known about the home range size of the sloth bear in a human-dominated landscape, though studies elsewhere (Joshi, Garshelis & Smith, 1995; Ratnayeke, Manen & Padmalal, 2007) have indicated much smaller average home range estimates compared to those of the central Indian landscape (Yoganand, 2005). Since we deployed camera traps within smaller grids (2 km × 2 km), we used the data from all 1–4 camera traps falling inside each 3.52 km × 3.52 km grid and considered these as spatial replicates, following Xiao et al. (2018). We used the N-mixture model to estimate the relative abundance of sloth bears from independent photographic captures as a function of environmental and anthropogenic covariates (Table 1).

N-mixture models assume population closure (i.e., the number of individuals residing within a site is constant as no emigration or immigration occurs between sites). Due to the individually unidentifiable aspect and long-ranging behavior of the sloth bear, we believe that the closure assumption was likely to be violated. Hence, we relaxed the closure assumption by changing the interpretation from absolute abundance at a site to the number of individuals ever associated with a site during a given period (Kéry & Royle, 2015), otherwise known as relative abundance. We pooled independent detections of sloth bear for each site “i” from “R” sites where, i = 1, 2, 3,..., R during a sampling occasion of “T”, where t = 1, 2, 3,…, T. The observed count (y_it) at “i” sites during “t” sampling occasions followed a binomial distribution, and the site abundance N_i followed a Poisson distribution (Kéry & Royle, 2015).

Due to the greater flexibility in modeling abundance with covariates and the ability to incorporate other effects (e.g., zero inflation, latent state), Poisson distribution is an integral part of the state process, i.e., abundance (Royle & Dorazio, 2008). On the other hand, binomial distribution accounts for imperfect detection or false-negative errors. The variation of observed counts (y_it) is an effect of imperfect detection of the actual (unknown) abundance N_i and its variability among sites. In simple algebraic form, we expressed the two processes as:

State process: N_i ~ Poisson (λ_i); log (λ_i) = β₀ + effects of covariates

Observation process: y_it | N_i ~ Binomial (N_i, p_it); logit (p_it) = α₀ + effects of covariates

where N_i is the latent abundance of site “i” (i = 1,…., R), λ_i is the mean abundance of site “i”, y_it is the count of the species at site “i” and occasion “t” (t = 1,…., T), and p_it is detection probability at site “i” and occasion “t”.

Studies have indicated that N-mixture models with Poisson and zero-inflated Poisson (ZIP) distributions can produce reliable relative abundance estimates if detection probability is modelled effectively with appropriate covariates (Barker et al., 2018; Kéry, 2018). We modelled the detection probability with the trapping period as we suspected that a more extended trapping period would positively influence the detection probability of sloth bears (O’Connor et al., 2017). We also used local-scale spatial covariates (and no GIS-based covariates) likely to affect the detection probability, following Hofmeester et al. (2019). The detection probability would be constant if none of the covariates influenced it (Kafley et al., 2019).

After determining the effect of covariates on the detection process, we modelled the relative abundance of sloth bears as a function of environmental (forest types, LULC types, RI, distance to nearest water source, availability of fruiting trees, shrub density and number of termite mounds) and anthropogenic (human photographic capture rate, distance to nearest village and metal road) covariates with three probability distributions, i.e., Poisson, ZIP and negative binomial (NB). Mean values of GIS-based environmental and anthropogenic covariates were calculated and extracted from each of the 12.40 km² grids. For local-scale spatial covariates where the number of camera trap locations was more than one, we calculated the average values of each covariate for the 12.40 km² grids. The selection of the most parsimonious models from the candidate model set followed the ascending Akaike Information Criterion (AIC) values (Burnham & Anderson, 2002). We reported uninformative parameters but drew final inferences only from the models within ≤2 ΔAIC units (Burnham & Anderson, 2002) with statistically significant (P < 0.05) parameters (Arnold, 2010). Model averaging was not considered in the case of a single top-ranked model or the presence of uninformative parameters in multiple top-ranked models (Arnold, 2010). However, due to the “good fit bad prediction dilemma” as mentioned by Kéry & Royle (2015), we further performed Goodness of Fit (GoF) tests with 1,000 iterations of bootstrapping for each of the most parsimonious models of corresponding distributions. We also calculated overdispersion parameters (ĉ) for the same models (Kéry & Royle, 2015). We proceeded with further residual diagnostics and mapping of residuals if none of them passed the GoF test (when P < 0.05) for the final checking of model adequacy (Kéry & Royle, 2015). The result was interpreted as an outcome of unstructured noise if no specific pattern of lack of fit was found (Kéry & Royle, 2015). Subsequently, we corrected it by inflating the predicted abundance’s range (confidence interval) (Johnson, Laake & Ver Hoef, 2010; Kéry & Royle, 2015). We standardized all covariates using Z-transformation before analyses, and subsequently, Pearson’s correlation tests were performed to avoid multi-collinearity. A pairwise Pearson’s correlation coefficient value of r≥ ± 0.7 was considered strongly correlated (Dormann et al., 2013) and was not modeled together, as we were interested in seeing the effect of each covariate on the relative abundance of sloth bears. All analyses were carried out using the packages “Unmarked” (Fiske & Chandler, 2011) and “AICmodavg” (Mazerolle, 2015) implemented in R (R Core Team, 2019).

We compared the mean site abundance estimates with estimates obtained from another frequently used site-structured model, the RN model (Royle & Nichols, 2003), to understand the performance of the N-mixture model. Since modeling the relative abundance of the sloth bear with both local-scale and GIS-based covariates by applying the RN model and subsequent comparison of the model performance (of both N-mixture and RN models) was beyond the scope of this paper, we restricted the comparison between the null models of the two frameworks mentioned above. We chose the same probability distribution for the RN model which was the most suitable for the N-mixture model. We used the program Presence (Version 2.13.11) (Hines, 2006) to estimate the mean site abundance of sloth bears by applying the RN model.

Camera trap

A total of 191 and 5,292 independent photographs of sloth bears and humans, respectively, were obtained from a total effort of 5,950 trap-nights (3,870 and 2,080 trap-nights in SNP and DWLS, respectively). Sloth bears were detected at 76 out of 143 (53.10%) camera trap locations (sites), and independent detections of sloth bear varied from 1 to 15 per site (Fig. 2). However, in the larger grids (12.40 km²), the presence of sloth bears was recorded at 51 out of 73 (70.0%) grids. In SNP, a higher number (120) of sloth bear detections was observed compared to 71 detections in DWLS.

Detection process

The density of fruiting trees significantly influenced the detection probability of sloth bears for Poisson, NB and ZIP distributions (Table 2). The estimated detection probability (p) of the best detection model (modelled with the density of fruiting trees) was found to be higher in Poisson (0.08 ± 0.01) than in ZIP (0.06 ± 0.02) and NB (0.01 ± 0.004) distributions (Table 2).

Relative abundance of sloth bears and effects of covariates

In top-ranked models, the relative abundance of sloth bears was significantly influenced by forest types (mixed and sal forest; environmental), distance to the nearest village (anthropogenic) and human photo-capture rate (anthropogenic) (Tables 3 and 4). In addition to these four, the most parsimonious models also included the LULC types (agricultural land, scrubland and water body); however, none of these covariates significantly influenced the relative abundance of sloth bears (Table 4). Models with NB distributions were the most parsimonious in terms of lower overall AIC values, followed by ZIP and Poisson models (Table 3). Due to uninformative parameters, we did not carry out model averaging for any top-ranked models. Areas of mixed and sal forest predicted a higher relative abundance of sloth bears (Table 4 and Fig. 3; also see Figs. S2 and S3). The distance to the nearest village and human photo-capture rate also had significant positive effects (Table 4 and Fig. 3; also see Figs. S2 and S3) on sloth bear abundance.

However, the models within ≤2 ΔAIC units for Poisson, ZIP and NB distributions did not pass the GoF test. Overdispersion parameter estimates (ĉ) revealed a marginal lack of fit, which did not appear to differ among Poisson (ĉ = 1.31), NB (ĉ = 1.19) and ZIP (ĉ = 1.28) distributions. Residual diagnostics and mapping of residuals did not show any specific pattern of lack of fit. Hence, this finding could be attributed to unstructured noise or overdispersion rather than a true lack of fit. The NB model (top-ranked) produced unusually high mean site abundance (λ = 10.80 ± 3.83) in comparison to mean site abundance estimates of ZIP (λ = 4.07 ± 1.56) and Poisson models (λ = 2.60 ± 0.64) (Table 5). Spatial variation of relative abundance produced by the Poisson model was less (1–11; see also Fig. 4) than those of the ZIP (1–14) and NB models (4–37).

Comparing estimates with the RN model

The RN model (null model with Poisson distribution) produced a mean site abundance estimate of 1.62 ± 0.34, which was slightly lower than the estimate produced by the N-mixture model (2.05 ± 0.35). However, the mean site abundance estimate obtained from the N-mixture model was within the 95% confidence interval range (1.08–2.44) of the estimate produced by the RN model.

Efficacy of the N-mixture model

We reported the first-ever estimates of the relative abundance of the sloth bear using the N-mixture model. We also identified associated factors governing the relative abundance by using count data obtained from the camera trap survey in STR. The N-mixture model can produce reliable abundance estimates if the data meet the assumptions of the N-mixture model while accounting for imperfect detection (Royle, 2004). Also, its ability to model spatial variation of abundance and detection probability as a function of relevant covariates makes its inference more robust and reliable than inferences drawn from conventional index/encounter rate-based surveys (Gilbert et al., 2021). A growing number of studies have assessed the population abundance (relative and absolute) of unmarked individuals by using the N-mixture model from camera trap data (Brodie & Giordano, 2013; Keever et al., 2017; Xiao et al., 2018; Kafley et al., 2019; Penjor et al., 2019) as well as employing other field methods (Belant et al., 2016; Ficetola et al., 2018; Kidwai et al., 2019; Searle et al., 2020). Estimating the abundance of unmarked animals from camera trap data has always been challenging for ecologists. For the past two decades, very few studies have actually assessed the population abundance of Asian bears employing statistically rigorous methods; instead, many of the studies have heavily relied upon sign index or interview-based distributions due to the relatively less expensive nature and ease of application in the field (Garshelis et al., 2022; Proctor et al., 2022). Information on sloth bear abundance and its governing factors is lacking throughout the species’ range (Dharaiya, Bargali & Sharp, 2016). Most of the previous studies (Eisenberg & Lockhart, 1972; Laurie & Seidensticker, 1977; Santiapillai & Santiapillai, 1990; Ratnayeke, Manen & Padmalal, 2007; Akhtar, Bargali & Chauhan, 2008) regarding the abundance of sloth bears produced empirical estimates, except for Garshelis, Joshi & Smith (1999), where the authors first estimated abundance from sound analytical methods in Chitwan. However, in the case of this wide-ranging bear species, site-structured models (i.e., the RN model and the N-mixture model) may be reliable for estimating the population trend with careful considerations of model assumptions and study design (Morin et al., 2022).

Moving forward, one should not interpret the abundance estimates obtained from site-structured models as a measure of absolute but relative abundance if the assumption of closure is likely to be violated (Barker et al., 2018; Link et al., 2018; Gilbert et al., 2021). In this context, mean abundance (lambda) from the N-mixture model reflects all individuals using the site (12.40 km² grids in the present study) and not necessarily all individuals restricted to the site. Site abundance estimates obtained from the N-mixture models would be consistently larger than the actual abundance of free-ranging animals like sloth bears, as each site contains multiple overlapping territories of different individual sloth bears. In reality, the actual abundance of animals comes from a considerably larger geographic extent than the actual surveyed area, leading to an overestimation of the mean site abundance and total predicted abundance (Joseph et al., 2009). In the present study, the best-selected model (with NB distribution) in terms of AIC produced substantially larger and ecologically unrealistic abundance estimates of sloth bears compared to ZIP and Poisson models. Mean abundance estimates obtained from NB and ZIP models showed that a large number (~11 and ~4 for NB and ZIP models, respectively) of bears were utilizing each site with considerably low detection probability, which, in general, is unlikely for a species like the sloth bear. In the Indian sub-continent, the lack of studies focusing specifically on the density or abundance of sloth bears and other bear species in a scientifically rigorous framework poses difficulties for comparing our findings to other studies. We expected that for a species like sloth bear, an elusive and solitary ursid, count data obtained from camera traps would be zero-inflated. The NB model performs poorly with a zero-inflated count dataset or when the focal species is not frequently found or is truly absent from a large number of sites surveyed (Joseph et al., 2009). In the present study, NB and ZIP variants of N-mixture models produced lower detection probabilities than Poisson models. Due to low detection probability, N-mixture models can produce positively biased abundance estimates with less precision than other methods (Dénes, Silveira & Beissinger, 2015; Dennis, Morgan & Ridout, 2015). On the other hand, the N-mixture model with a Poisson distribution can handle datasets with excessive true zeros (sites where the species is truly absent), given that the species naturally occurs in low densities (Joseph et al., 2009). In our study, the Poisson model produced a comparatively low mean site abundance (2.60 ± 0.64) by modeling a substantial number (~30% of total sites surveyed) of unoccupied sites. We suggest interpreting the abundance estimates of sloth bears carefully with an ecologically realistic view rather than with strict adherence to the statistical properties of the models.

The mean site abundance of sloth bears produced from the N-mixture model (Poisson distribution) was comparable to estimates obtained from the RN model. Earlier studies have indicated that the RN model would not be a good choice when the focal species is common; in that case, the N-mixture model would be preferable (Dénes, Silveira & Beissinger, 2015; Kéry & Royle, 2015; Gilbert et al., 2021). The RN model reportedly produces imprecise estimates from camera trap-based surveys (covering comparatively small sampled areas), especially if the focal species is cryptic (Gupta et al., 2012; Kalle et al., 2014; Morin et al., 2022). However, there is evidence that the RN model can produce precise abundance estimates, enabling it to detect population trends (O’Brien et al., 2020). Estimates from the RN model can approximate the absolute density of the focal species if the study is conducted on a landscape scale in conjunction with an ecologically relevant grid size for camera trap deployment (Linden et al., 2017). On the other hand, simulation studies have shown that the N-mixture (Poisson and Binomial mixture) model may produce biased estimates of mean abundance in the context of false-positive detections (Nakashima, 2020). However, our findings from both models indicated the overall utility of site-structured models in estimating the relative abundance of elusive species like sloth bears. Although the number of camera trap locations/grid varied (1–4) in our study, the presence of >1 spatial replicates (camera trap locations) within most of the grids (>50%) might have improved the detection and robustness of mean site abundance estimates (Kolowski, Oley & McShea, 2021; Morin et al., 2022).

Applicability of other methods for unmarked population estimation

Due to the scarcity of spatially explicit studies, we could not directly compare relative abundance estimates of sloth bears with previous findings. Since the primary objective of this study was to understand the applicability of the N-mixture model, we considered other potential methods relevant to the abundance estimation of unmarked animals. However, it has also been found that abundance estimation of unmarked species like Asian bears through camera traps rarely produces precise estimates for conservation planning unless supported by ancillary information obtained from statistically rigorous methods and covering a large number of sampling sites (Morin et al., 2022). Apart from the site-structured models presented here, USCR (i.e., the spatial count (SC) model; Chandler & Andrew Royle, 2013) and its extension (i.e., spatial presence-absence (SPA) model; Ramsey, Caley & Robley, 2015) can also be applied to estimate the absolute density of cryptic species. The application of the SPA model has been well-demonstrated to estimate the density of small carnivores through camera traps (Chatterjee, Nigam & Habib, 2020a, 2020b). However, these models are computationally intensive (requiring Bayesian analysis) and sensitive to the appropriate choice of priors which could be challenging for a species like the sloth bear due to limited knowledge of its ranging pattern. Distance sampling through camera traps could also be a promising approach to estimating the density of unmarked animals (Howe et al., 2017), especially if the focal species is less abundant (Palencia et al., 2021). However, this method requires the random placement of camera traps which we could not implement, as this study followed the protocol of All India Tiger Estimation (Jhala, Qureshi & Gopal, 2015). Random placement of camera traps might result in substantially low detections of sloth bears, which may not be conducive for further analysis. The N-mixture model would be particularly suitable where finance and logistics are limiting factors. Moreover, if the study intends to model the spatial variation of abundance (relative or absolute), the N-mixture model is a better choice over other models, as long as it meets the objectives and the interpretation can be plausible.

Effects of environmental and anthropogenic covariates on relative abundance

In the present study, mixed and sal forests positively influenced the relative abundance of sloth bears. In STR, we calculated the density of termite mounds and fruiting trees, utilizing a 10 m-radius plot around each camera trap location. We recorded a high density (35/hectare) of termite mounds in mixed forest, followed by sal mixed and sal forest (27/hectare and 23/hectare, respectively). Similarly, the fruiting tree density was also higher (208/hectare) in the mixed forest in comparison to sal mixed (197/hectare) and sal forest (100/hectare). Due to fewer plots, we could not calculate meaningful density estimates of termite mounds and fruiting trees for the bamboo mixed forest and grasslands. In STR, a high number of fruiting trees and termite mounds in the mixed forest may explain the positive relationship between bear abundance and mixed forest. We found a similar observation from Panna National Park in central India, where the dense forest (comprised of miscellaneous tree species) had the highest densities of fruiting trees and prey insect colonies, with sloth bears preferring this habitat significantly (Yoganand, 2005). However, the limited number of surveyed plots (only at camera trap locations) to support our findings may not be enough to explain the effects of forest types and termite mounds on sloth bear abundance, for which more rigorous sampling is needed. In the North Bilaspur Forest Division, Akhtar, Bargali & Chauhan (2004) found that signs of sloth bears were more common in sal forest than in other forest types. We also found similar results in Chitwan but with distinct seasonal patterns (Garshelis, Joshi & Smith, 1999). The occurrence of bear signs (ground holes and mound holes) in sal forest was frequent in both landscapes, perhaps due to a relatively high number of termite mounds (Garshelis, Joshi & Smith, 1999; Akhtar, Bargali & Chauhan, 2004) and possibly the underground colonies of ants and termites. In the northern part of the Indian sub-continent, termite mounds are more abundant in sal forests (Chakraborty & Singh, 2020). However, underground colonies of termites and ants are difficult to quantify. Hence, we considered the number of termite mounds that are above-ground and visible as a surrogate for the availability of ants and termites (Akhtar, Bargali & Chauhan, 2004; Rather et al., 2020). We also could not find any significant relationship between the relative abundance of sloth bears and local-scale covariates, although the availability of fruiting trees positively influenced the detection probability of sloth bears. In Nepal’s Chitwan and Churia (a lowland forest outside the PAs) forests, the presence of fruiting trees and termite mounds positively influenced the sloth bear’s occupancy and detection probability (Paudel et al., 2022; Pokharel et al., 2022). Das et al. (2014) also recorded a similar observation in southern India during their occupancy-based study. However, we believe that our camera trap-based study design may not be appropriate for drawing inferences on such local-scale covariates, despite their proven importance to the sloth bear’s ecology. The ubiquitous nature of termites and ants in a relatively small area (Redford, 1987) may explain the insignificant relationship between bear abundance and termite mounds. The availability of fruiting trees and termite mounds at a particular point (10 m-radius plot, in our study) may not necessarily represent the entire sampling unit (12.40 km² grids). Thus, it could explain the apparent insignificant contribution of fruiting tree density and termite mounds to the relative abundance of sloth bears.

Sloth bears generally avoid human disturbances (Santhanakrishnan et al., 2015; Puri et al., 2015; Paudel et al., 2022; Pokharel et al., 2022). However, they can also tolerate certain degree of human pressure or get habituated to living in disturbed and fragmented forests (Akhtar, Bargali & Chauhan, 2004; Prajapati, Koli & Sundar, 2021). In the present study, we found that the photographic capture rate of humans and distance to the nearest village positively correlated with relative abundance of sloth bears. However, based on the current study design, it is not known whether a causal relationship between these variables (relative abundance of sloth bear and human photographic capture rate) exists. Moreover, to our knowledge, no other studies have shown that any such ecological relationship prevails. We believe this finding may be purely attributed to the placement of camera traps on the forest roads and trails. Apart from sloth bears (and other carnivores), forest roads and trails are intensively used daily by local people living inside STR. We were interested in this relationship between human occurrence and sloth bear abundance as a matter of concern regarding the possible negative interactions between humans and sloth bears. However, to cope with extensive human pressure in STR, sloth bears have shown a fine-scale seasonal spatio-temporal segregation (Chaudhuri et al., 2022). On the other hand, sloth bears avoided areas close to villages in STR, which is in agreement with previous studies (Santhanakrishnan et al., 2015; Puri et al., 2015; Paudel et al., 2022; Pokharel et al., 2022). Such behavior could be due to the better availability of forage and shelter (day-resting den sites) in relatively undisturbed habitats.

Nevertheless, avoidance of human settlements does not necessarily curb the possibilities of conflict between humans and sloth bears, given the prevailing extensive anthropogenic pressure in STR. Sloth bears are also known to opportunistically raid crops and forage on fruits of Ziziphus mauritiana, Mangifera indica and Syzygium cumini found in the villages but less abundant inside the forest (Bargali, Akhtar & Chauhan, 2004; Palei, Debata & Sahu, 2020). Habituation to human settlements often leads to severe conflicts between humans and sloth bears (Bargali, Akhtar & Chauhan, 2005; Dhamorikar et al., 2017; Prajapati, Koli & Sundar, 2021), so there are significant conservation implications for this species in human-dominated landscapes.

Caveats and limitations

Our study has a few limitations and caveats. Firstly, the camera trap design was more oriented towards tiger movement, but also aimed at other co-predators. Secondly, we could not deploy camera traps intensively due to villages and agricultural fields, which resulted in an unequal number of camera traps (1–4) in each of the larger grids (12.40 km²). Such design may have affected the detection of sloth bears as they are presumed to raid crops and forage on fruits in the villages situated inside the core area of STR. Thirdly, we restricted the camera trap survey to the core area of STR and therefore, could not account for overall estimates of sloth bears’ relative abundance for the entire STR. Finally, we could not assess the seasonal variation of mean site abundance of sloth bears, as studies conducted elsewhere reported a distinct seasonal variation in sloth bear density (Garshelis, Joshi & Smith, 1999).