The CTL ( Figure 1 ) spans 5400 km² abutting the lower Himalayan Ranges in the states of Uttarakhand and Uttar Pradesh in India and southwest Nepal. The CTL is recognized as a globally important eco-region and supports numerous endemic and endangered plant and animal species (Olson and Dinerstein, 1998), including tigers, hog deer (Axis porcinus), swamp deer (Rucervus duvaucelii), rhinoceros (Rhinoceros unicornis), and elephants (Elephas maximus). Key vegetation communities are deciduous forests dominated by Sal and grasslands characterized by Saccharum spontaneum, Themeda arundinacea, Narenga porphyrocoma, Imperata cylindrica and associated species (Dinerstein, 1980; Johnsingh et al., 2004). The regions grasslands and associated wetlands, important habitats for large mammals, have been greatly reduced in extent by agricultural expansion from historic to more recent times (Peet et al., 1999a, Peet et al., 1999b; Johnsingh et al., 2004). Today, the CTL is one of the most densely populated regions across the tiger’s range with human population density of about 600 persons/km² and population growth rates of about 30% (in the 2001–2011 decade). Aside from the region’s forests, the only areas that have escaped extensive land use change lie in immediate proximity of the Sharda and Ghaghra Rivers that experience extensive annual floods during the monsoons.

Our study area included three PAs and three MUFs within the CTL ( Figure 1 ; Table 1 ). The PAs, including Dudhwa National Park and Kishanpur and Katerniaghat Wildlife Sanctuaries, were established between 1972 and 1977 and are collectively known as the Dudhwa Tiger Reserve. Management actions in Tiger Reserves like Dudhwa prioritize the conservation of habitats for tigers and other wildlife. In contrast, MUFs in the CTL were managed for timber production, via selective felling, for approximately 150 years. In addition, these MUFs are extensively used by local communities to extract fuel-wood, fodder and other resources (Strahorn, 2009).

The three MUFs, collectively called the Pilibhit Forest Complex, (PCF; 1400 km² in extent) lie within the largest area of contiguous primary forest and grassland habitat in the CTL. The largest MUF within the PFC,was re-designated as Pilibhit Tiger Reserve in 2014, based on the results of our study. Within the CTL, Dudhwa National Park has limited connectivity to other tiger habitats, while the other PAs and MUFs are connected, via conservation corridors, with tiger habitats in Nepal ( Figure 1 ; Chanchani et al., 2016).

We sampled tiger populations using camera traps from November 2011 to June 2012 and November 2012 to June 2013. Each sampled area (PA or MUF) was referred to as a “site” ( Table 1 , Figure 1 ). Each site was sampled for ≤ 60 days to meet the assumption of demographic closure required by the closed-population, capture–recapture analyses we conducted (Karanth and Nichols, 2002; Royle et al., 2009; Table 1 ).

Each sampling station consisted of a pair of cameras (Cuddeback Attack, Cuddeback INC. WI, USA) housed in metal security cases and secured to a post or tree 2–5 m from the edges of trails or unpaved roads, about 50–60 cm above the ground. Camera trap stations within the survey grid were spaced 1–2 km apart. To maximize the probability of capture, camera stations were installed in carefully selected locations, for example, at wildlife trail intersections and along drainages. Camera trap data were downloaded every 5–7 days.

We had too few cameras to survey a site in its entirety using a single survey block (camera traps deployed across a site simultaneously). Therefore, we employed an incomplete trap layout design (Karanth and Nichols, 2002; O’Brien and Kinnaird, 2011), and sampled large study sites using 2–3 survey blocks. Each block of camera trap stations was active for 15–30 days, after which cameras were moved to an adjacent block ( Table 1 ). In Katerniaghat WLS, we sampled approximately 30% more area in 2013, than in 2012. In DNP, a similar area was sampled in 2012 and 2013 ( Figure 1 ).

Photographs of adult tigers (post-dispersal, ≥ 2 years of age) were visually analyzed by three independent observers to identify unique individuals. In addition, we used pattern recognition software (Extract Compare; Hiby et al., 2009) to corroborate identities of all usable images in the dataset (a library of tiger images is available online; Chanchani et al., 2014a; Appendix S1 ). Detection history matrices recorded tiger capture histories, referenced by individuals, trap station (location) and sampling occasion (24-hr period, within the sequence of days over which cameras were active).

By estimating the locations of individual activity centers based on animal movement within the trap array, spatially explicit capture recapture (SECR) models provide model-based estimates of density (D) (Borchers and Efford, 2008; Royle et al., 2013; Efford et al., 2015).

In SECR, the probability model for detection histories includes parameters for the distribution of activity centers (both within and beyond the trap area), and for detection probability (assumed to be<1 and decline with increasing distance from an animal’s activity center to a trap). We estimated detection probability as a function of two parameters – a spatial scale parameter (σ) and a baseline detection probability (g₀ ; the probability of detection assuming that a camera trap location is coincident with an individual’s activity center) (Efford et al., 2009). Because individuals within a population may have vastly different space use (e.g., males have larger home ranges than females), g₀ and σ may co-vary. In such cases, estimates were obtained by replacing the conventional detection parameterization g ₀ and σ by an alternate “compensatory heterogeneity” parameterization a_0, σ (Efford and Mowat, 2014). Compensatory heterogeneity implies a negative covariance between g₀ and σ among individuals of one or both sexes. This occurs when the differences in detection probability of a given animal at specific locations are directly proportional to the amount of time it spends at different locations within its home range (Efford and Mowat, 2014). The distribution of animal activity centers in SECR models can be described as a homogenous or inhomogeneous Poisson point-process. Homogeneity implies that the expected value of density is uniform across the state-space over which activity centers are distributed (a pre-defined part of the model), whereas an inhomogeneous point process implies that density varies across space, as a function of environmental covariates (Efford and Fewster, 2013).

We used SECR models (maximum likelihood parameterization implemented in the R package secr, version 2.9.5; Efford, 2015) to test hypotheses about spatial heterogeneity in tiger density ( Table 2 ). We assumed that the distribution of tiger activity centers followed a Poisson point process (Efford and Fewster, 2013). Following Royle et al. (2009), we delineated a grid (580 × 580m) over which individual tiger activity centers could be distributed. This state space included the tigers’ primary habitats, forest and grassland areas. Each square gird cell (0.34 km²) represented a potential activity center. The region S (integration region) over which activity centers could be distributed was defined collectively by a habitat mask comprising of the camera trap array, a 15 km buffer around the trapping array that included forests, grasslands and other features identified as “habitat”.Tiger density was estimated separately for each site ( Table 1 ) using a two-step procedure in a multi-model hypothesis testing framework (Burnham and Anderson, 2002; Doherty et al., 2010). First, we assessed relative support for models with alternate parameterizations of g₀ and σ, or a₀ and σ with and without sex-specific heterogeneity. Because male home range sizes are generally larger than those of females (Smith, 1993; Sollmann et al., 2011), we allowed detection parameters to vary by sex using a two-class hybrid mixture model (Efford, 2014). The hybrid mixture model includes a mixing proportion parameter ‘pmix’, which enables class-specific modeling of detection parameters and sex ratio estimation. Further, we fit three additional models: common detection parameters for all individuals [g₀ (.), σ(.)] and sex-specific heterogeneity with respect to g₀ or _σ (models [g₀ (sex), σ(.)] and [g₀ (.), σ(sex)]. Finally, we evaluated support for the compensatory heterogeneity parameterization (a₀ , σ) by building one additional model [a₀ (.), σ(sex)]. For all five alternative detection function parameterizations, we retained a “global” structure for density, allowing it to vary as a function of four covariates: distance to large (>1km²) grasslands (‘grass’), percentage of tree canopy cover (‘treecov’), prey density (‘prey’), and distance to nearest forest edge [‘edge’; model: D(grass+treecov+prey+edge)]. Relative support for models was assessed using Akaike Information Criteria (AICc) adjusted for small sample sizes (Burnham and Anderson, 2002). The best supported model (lowest AICc score) from each model set was selected and carried forward into the second analysis step.

We retained the best supported covariate structure for the detection parameters, in subsequent models. For each of our data sets ( Table 1 ), we fit 15 models that represented our a priori hypotheses about the relative effect of prey density, habitat (vegetation cover) and proximity to edges on fine scale spatial variations in tiger density ( Appendix S3 : Table S1 ). We compiled values for five covariates (prey density, distance from grasslands, percent tree cover, distance to forest edges and distance to the international border) for each point (potential tiger activity centers), representing a 580 × 580 m (0.34 km²) area over the integration region ( Table 2 ).

We estimated tiger abundance (N) as the number of activity centers within the boundaries of PAs and MUFs (Efford and Fewster, 2013). In addition to density estimates, we N within the specific boundaries of an administrative region, such as an individual PA (Efford and Fewster, 2013). The boundaries of these regions were a subset of the integration region (S) for each site ( Figure 1 ). We compiled and compared model-averaged estimates of N for all 15 models, and for all eight model sets ( Figure 2 ; Appendix S3 : Supplementary Table S1 ).

Model selection results from the first analysis step varied between sites and across years ( Appendix S2 : Supplementary Table S1 ). Sex-specific heterogeneity in the spatial scale parameter σ(sex) was supported in six of the eight model sets. In two of the six model sets, inclusion of the “compensatory heterogeneity” detection parameter [a₀ (.), σ(sex)] had greater support than the conventional baseline encounter parameter [g₀ (.), σ(sex)]. In addition, there was evidence of sex-specific heterogeneity for both g₀ and σ[g₀ (sex), σ(sex)] in only one of the eight model sets. Lastly, the constant model [g₀ (.), σ(.)] was well supported (> 50% AIC weight) in only one of the eight model sets ( Appendix S2 : Supplementary Tables S1–S3 ).

Overall, tiger densities were higher in sites with higher prey densities, independent of management designation and degree of human disturbance. Site- and year-specific estimates of tiger density (individuals/100 km²) were highest for Kishanpur Wildlife Sanctuary   KWLS  D ^ 2012 = 5.45 (SE=1.29), KWLS D ^ 2013 = 4.97 (SE=1.27). Tiger density was the lowest for Dudhwa National Park, [DNP D ^ 2012 = 2.05 (SE = 0.56) D ^ 2013 = 1.88 (SE = 0.52)]. The other PAs and MUFs had intermediate tiger densities: Katerniaghat Wildlife Sanctuary [KGWLS D ^ 2012 = 4.76 (SE=1.14), KGWLS D ^ 2013 = 2.78 (SE=0.72)]; Pilibhit Reserved Forest [PRF D ^ 2013 = 3.028 (SE=0.71)]; and Pilibhit Forest Complex [PFC D ^ 2013 = 3.4 (SE=0.51)] ( Figure 2 ; Appendix S3 : Table S1 ). In the three sites surveyed in 2012 and 2013, estimates were marginally higher in 2012 than in 2013, but less precise ( Figure 3 ; Appendix S3 : Table S1 ). Corresponding estimates of prey density (combined for all ungulates) were 29.81 (SE=5.69) animals/km² in KWLS, 13.64 (5.28) in DNP, 4.4 (1.55) and 22.4 (8.10) in KGWLS (in forests and grasslands respectively) and 40.05 (SE=0.23) in PFD (Bista, 2011; Chanchani et al., 2014a).

Model-averaged estimates of σ were approximately two times higher for male (range: 2.08–6.931 km) than for female tigers (range: 1847–2962 m; n= 6 site/year combinations; Appendix S3 : Table S4 ). Estimates of σ were largest for both males and females in Katerniaghat Wildlife Sanctuary, a site with even sex ratios and lower tiger density. In contrast, the lowest estimates of σ were recorded in Pilibhit Forest Division and Kishanpur Wildlife Sanctuary, where sex ratios were female-biased, and tigers occurred at relatively higher densities. Both σ and g₀/a₀ estimates and their CIs were similar across all 15 models assessed in the eight model sets ( Appendix S3 : Supplementary Table S4 ). Estimates of σ were consistently higher in 2013, than in 2012, especially in Dudhwa National Park and Katerniaghat Wildlife Sanctuary where survey effort was greater in 2013 ( Table 1 ).

Similar to broad-scale patterns in tiger density, variation in prey abundance explained the greatest amount of the fine-scale variation in tiger density ( Table 4 ). Hypothesis 1, was strongly supported in Dudhwa National Park (2013), where the model (D(Prey)) had 85% overall support ( Appendix S3 : Supplementary Table S1 ). In Dudhwa, there was a steep gradient in prey density: high in the south and low in the north ( Figure 3 ; Appendix S6 ). In four of the seven additional model sets, [D(Prey)] contained >10% of the model support. The distance to grassland (hypothesis 2) model [D(Grass)] had support (w=0.48) only in the model sets pertaining to Katerniaghat Wildlife Sanctuary (2012). In Katerniaghat WLS, grassland habitats occur only in its eastern and southern reaches, with the remaining areas dominated by dense, even-age deciduous forests. In other sites where grasslands were distributed more uniformly, the distance to grassland covariate received less support.

Hypothesis 2 – vegetation, and hypothesis 3 – edges (TreeCov and Edge) had less support than models with no covariates [null model, D(.)] in seven of eight model sets. Overall, in six of the eight model sets, the null model had the most support. AICc weight for the [D(.)] model varied between 0.8 in Kishanpur Wildlife Sanctuary (2013) and 0.23 in Pilibhit Complex (2013) ( Appendix S3 : Supplementary Table S1 ). In several sites, models with one or more covariates had similar support to the null model. Because the addition of one or more covariates did not result in improved model fit, these models were considered uninformative (Arnold, 2010).

The largest tiger population in the CTL was recorded in the PFC ( N ^ ₂₀₁₃ = 50.54, SE=2.63), which at the time of our study was composed on >80% MUF ( Figure 1 ). Within the PFC, the largest tiger population was recorded in Pilibhit Forest Division, a MUF ( N ^ ₂₀₁₃ = 25.22; SE= 1.71). The 200 km² Kishanpur Wildlife Sanctuary – which is the sole PA within PFC,supported a large tiger population ( N ^ ₂₀₁₂ = 19.21; SE=1.54) and ( N ^ ₂₀₁₃ = 16.0; SE = 0.02).

Dudhwa National Park supported a relatively small population of tigers ( N ^ ₂₀₁₂ = 14.25, SE = 0.51; N ^ ₂₀₁₃ = 14.16, SE = 0.41). Dudhwa is similar in size to Pilibhit Forest Division, but only connected to other tiger habitats in India through corridors spanning agricultural land. Large areas of Dudhwa National Park support low prey densities ( Figure 5 ; Appendix S6 ). In Katerniaghat Wildlife Sanctuary, the estimated number of tigers was N ^ ₂₀₁₂ = 19.21 (SE = 1.54) and N ^ ₂₀₁₃ = 17.11 (SE =0.33) in 2012 and 2013, respectively. Katerniaghat is 40% smaller than Dudhwa National Park and the two PAs have similar prey densities.

Estimated sex ratios for adult tigers were female-biased in Kishanpur WLS, but male-biased or equal in Dudhwa NP and Katerniaghat WLS ( Table 3 ). Significantly, sex ratios varied over the two years of study. Even though density estimates were similar across years, at several sites (e.g. Katerniaghat WLS, Dudhwa NP), there was a high turnover (~30%) of adult tigers between successive surveys (for age and sex-specific details of tiger captures, see Appendix S5 : Figure S1 ). The point estimates for the 15 models within each of the eight model sets (site × year) were consistent, with similar precision ( Figure 2 ; Appendix S3 : Table S1 ).

While large carnivores have persisted in some multi-use landscapes, previous research from the terai and elsewhere indicates tiger occupancy and density can be severely depressed in areas with extensive human use (MadhuSudan, 2004; Harihar et al., 2009). Our research raises important questions about factors and “mechanisms of coexistence” that may be critical in determining whether or not, and in what densities tigers occur in human dominated landscapes. We posit two working hypotheses in the context of our study.

First, vegetation attributes that affect the distribution and density of wild herbivores may be an indirect but primary determinant of carnivore density variation, relative to the influence of human presence. In our study area, large parts of the core zone of Dudhwa National Park, with very limited human presence supported markedly lower tiger densities than areas with much higher human “disturbance” – i.e. the PA’s edges or Pilibhit Forest Division in which wild herbivore densities were several orders of magnitude higher ( Appendices S1 , S6 ). The most likely explanation for this variation in prey densities is understory composition and palatability, as Dudhwa’s sal (Shorea robusta) dominated interior and shrubby understory supports lower ungulate densities relative to PA and understory of the Pilibhit Forest Complex.

Second, tiger densities in areas with high human use, may also critically depend on presence, extent and distribution of refuge habitats (Valeix et al., 2012; Chanchani et al., 2016; Grilo et al., 2019; Oeser et al., 2022). These are habitat tracts that are secluded, relative to surrounding areas which have higher human presence, where tigers can shelter away from people, hide and feed on kills, raise cubs, and survive, even if people are in close proximity, allowing tigers to behaviorally trade-off risk and resources, including through temporal adjustments of activity (Carter et al., 2012; Oriol-Cotterill et al., 2015). In multi-use landscapes, habitat refuges may exist in the form of tracts of marshy grasslands, steep and densely vegetated terrain, dense and thorny understorey, or even sugarcane thickets. Given adequate numbers of prey, the extent and distribution of such habitat refuges in an area may well be a critical factor that sets a ceiling on how many tigers the area can support.

Breeding tiger populations are unlikely to persist in areas where human presence is so pervasive that even such refuges are routinely intruded or degraded by excessive resource use or extraction. Likewise, tiger populations are unlikely to survive in areas where protection is deficient, and there are is high hunting pressures on tigers or their prey species (Dinerstein et al., 2007). In such contexts, it will be important to tease apart potential effects of hunting from other anthropogenic pressures (eg. livestock grazing, non-timber forest produce harvesting, and silviculture operations).

In summary, we argue that multi use landscapes that support tigers, or are envisioned as tiger-recovery areas represent “zones of coexistence”. Empirically identifying refuge habitats in such zones will be a critical strategy to effect recovery and enable tiger-persistence, provided such efforts are supported by local forest users. De facto refuge habitats within multi-use landscapes can help increase tiger survival rates (Carroll and Miquelle, 2006). Decision making regarding their identification and management should include local village institutions (e.g. Velázquez et al., 2009), and be informed by the resource access and land tenure rights of communities, levels of conflict, and levels of local acceptance of dangerous wildlife, like tigers and support for their conservation. Ultimately such areas may only support stable or growing tiger populations if comprehensive conflict management systems are in place, and if the tangible and hidden costs of conflict are adequately redressed (Karanth and Gopal, 2005; Goodrich, 2010).