Community, conict and conservation: response of mammalian fauna to ecological and anthropological correlates ฀ a critical habitat in Indo-Bhutan transboundary landscape urges multiagency cooperation

The impacts of conflict on nature are devastatingly adverse but differ widely in different socio-political regimes. Armed conflict often facilitates illegal plunder and unsustainable use of natural resources, variously by rebel groups and impoverished or displaced people challenged with limited subsistence options. We studied the response of mammals in Ripu Reserve Forest (Assam) that suffers prolonged anthropogenic pressure due to armed conflict instigated by social unrest. We used standard single-season (spatial-dependence) occupancy models using sign survey to assess the factors affecting the space use of mammals and subsequently build capacity of conservation volunteers for long-term sustenance of Ripu. Our study revealed that Ripu has a high proportion of occupied area by prey species of large carnivores. Asian elephant, barking deer, and wild pig occupied most of the habitat, whereas gaur, sambar and spotted deer restricted themselves to selected patches within the Ripu. Common leopards found to be positively associated with prey occupancy. The studied mammals responded variably to different ecological and anthropological covariates and urge for species-specific management alongside landscape scale conservation approach. Our ground effort to strengthen community patrolling and operational execution of various alternative livelihood has helped to empower the economic condition of patrolling staff. Strategic implementation of law enforcement could support dispersal of tigers from Phibsoo WLS (Bhutan), potentially linked with the larger tiger and elephant landscape far west (Buxa Tiger Reserve) in the Terai region of India. Community-based conservation initiatives required continuous support from various agencies, including national, international, and local bodies, to restore this critical habitat.

Markovian processes. In recent years, spatially replicated correlated detection occupancy surveys have primarily been applied to track surveys covering medium to large size carnivores and herbivores (Hines et al. 2010;Wibisono et al. 2011; Thorn et al. 2011;Lakshminarayanan et al. 2016;Thapa et al. 2019;Jornburom et al. 2020;Lamichhane et al. 2020). In addition, it was already well established that spatial replication in occupancy studies useful for assessments of species habitat relationship (Srivathsa et al. 2018).
We used detection and non-detection data collected from sign surveys of mammals within an occupancy modelling frame work to assess the habitat use and simultaneously build capacity of conservation volunteers further to strengthen the monitoring and protection of Ripu RF. Our specific objectives were (a) to determine the proportion of Ripu occupied and investigate factors affecting the probability of occupancy and habitat use of elephant (Elephas maximus), gaur (Bos garus), sambar (Rusa unicolor), spotted deer (Axis axis), barking deer (Muntiacus muntjak), and wild pig (Sus scrofa), leopard (Panther pardus) (b) to develop predictive habitat use maps based on the spatially explicit occupancy models (c) to build capacity of conservation volunteers for the long-term monitoring and protection of Ripu RF. Here, we examined the influence of the proportion of different forest types, water availability, and anthropogenic activities on the site use pattern of the above-mentioned mammalian species. We hypothesized that these species would differ in their site use responses given their different habitat requirements. The findings of this study would be useful in informing management activities to conserve herbivore population and support large carnivore's recovery goals in this transboundary landscape.

Study Area
The Ripu Reserve Forest (RF) is one of the three reserve forests of Kachugaon Forest Division in the Kokrajhar district of BTAD (Bodoland Territorial Area Districts) in western Assam (Fig 1). The Indo-Bhutan boundary forms the northern boundary of the Ripu from the Sonkosh river on the west to Saralbhanga river on the east comprising 605.  (Tiwari et al. 2017). Except for Sonkosh and Saralbhanga other rivers and several rivulets and streams remains waterless in the dry season. The general topography of the region is low lying and more or less flat, except few undulating stumpy foothills that slope gently towards the south. The climate is warm and humid characterized by sub-tropical with pronounced monsoon with three distinct seasons, viz., winter, summer and monsoon (Borthakur, 1986). The mean annual rainfall is 3000 mm with an average temperature range between 06°C to 37°C. The study area falls under the Bhabar zone and the forest type is dominated by Moist-mixed deciduous and semi evergreen forest.
A total of 136 trees, 53 shrubs and 61 herb species recorded from Ripu-Chirang area (Menon et al. 2008). Albizia amara, Dillenia pentagyna and Terminalia arjuna are found to be associated with the sal dominated areas with in the study area. Moist deciduous miscellaneous forest represented by Dillenia pentagyna, Lagerstroemia parvilora, Castanopsis indica. The evergreen forests primarily include Mesua ferea and Moringa angustifolia. Besides, 270 species of birds have been reported (Menon et al. 2008) from the area which include the critically endangered White Bellied Heron (Ardea insignis). The forest division is also rich in mammalian fauna, previous study reported 24 species of mammals which includes threatened species like elephant, tiger, Asiatic wild dog, Golden langur, gaur, leopard, Chinese pangolin (Hilaluddin and Sharma 2008). The hunting and poaching of wildlife along with smuggling of timber in the study area have been reported to be frequent. The tree species targeted in the region are as follows-Shorea robusta, Pterospermum personatum, Terminalia sp, Schliochera oleosa, Diptereocarpus sp. Apart from timber, locals also heavily extracted fuel wood and other NTFP on regular basis (Menon et al. 2008).

Occupancy survey and sampling design
We employed a single season occupancy framework (MacKenzie et al. 2002;MacKenzie et al. 2006) to statistically assess the occupancy/habitat use of elephant, gaur, sambar, spotted deer, barking deer, wild pig and leopard using sign-survey data while explicitly addressing the concern of imperfect detection (Duangchantrasiri et al. 2015). We created a grids (n = 144) across the Ripu RF with each grid cell of size (2 × 2) 4 km 2 and investigated it in 75 grid cells to examine ecological and anthropogenic factors influencing site use pattern (Fig. 2). Our sampled grid cells covered 300 km 2 (~50%) of the available 605 km 2 of the study area. To infer the true occupancy, the area of a sampling unit should generally be larger than the average home range size of species so that assumption of geographic closure and independent sampling sites can be maintained (MacKenzie et al. 2002). Except for elephant, gaur, and leopard, the studied species in our study have small home ranges (less than 4 km 2 ). Additionally, this grid cell size circumscribes the expected daily ungulate movement based on home range size and movement rates reported at other field sites (see McShea et al. 2011;Thapa and Kelly 2017). For elephant, gaur, and leopard with larger home ranges than our 4 km 2 grid cells, occupancy models can be used to describe habitat-use at finer spatial scales rather than true occupancy [e.g. within home range habitat use, see tigers: Sunarto et al. (2012) and dholes: Srivathsa et al. (2014)].
At this finer scale, "probability of occupancy" is a measure of habitat use or the habitat that an animal selects relative to the habitat available within a home range (Johnson 1980). For mammals, modeling habitat use is key because it helps categorize the positive and negative ecological and anthropomorphic correlates that determine their home range or third-order habitat selection (Johnson 1980). This information is vital for achieving straightforward understanding of habitat use and making management decisions.
The sign survey was carried out by five different teams (each team consisting of 4-6 conservation volunteers and forest guards) simultaneously in different areas (Sanfan, Athiabari and Raimona Range) within the Ripu RF.
Skilled personnel among the conservation volunteers were mixed in each team since sign surveys are prone to observation error (Royle and Lin 2006;Shea et al. 2011;Thapa and Kelly 2017). We searched for signs along the trails following probable animal travel routes, including all types of forest paths, stream beds, and riverbanks, which were considered to have a high likelihood of finding signs (Barber-Meyer et al. 2013). We assume that sign availability was similar across the study area, though studies have previously noted that occupancy estimation based on signs is subject to the decay rates of signs (Rhodes et al. 2011). Only fresh dung, pellet, track, other signs, and direct sightings of animals were recorded to avoid other biases that may arise from decay signs (Rhodes et al. 2011;Hedges 2012).
We conducted the surveys in winter/dry seasons (January to April 2014) to make sure that sign persistence was consistent (Barnes 2001;Hedges et al. 2012) and to reduce heterogeneity in detection probability induced by rainfall variation (Royle and Nichols 2003). We spent ample time to verify signs and only recorded signs that were unambiguously identified. Specifically, for leopard, to avoid misidentification of signs, tracks and scrapes were identified using a combination of size and shape, while scats (tiger scat usually is less coiled compared to leopard scat; Wang and Macdonald 2009) were identified based on their size and the presence of secondary evidence at the site (Harihar 2012). On several occasions, we photographed tracks and signs and verified them later with experts. To further minimize the bias from the movement of animals from the surrounding grid cells, each grid cell was surveyed between 4-5 hours. The starting point of the trail was randomly selected within the grid for the sign survey. The effort of the sign survey ranged from 0.9 km of walk-in grid cells to 3.6 km with a mean of 2.37 km. The variation in the effort occurred due to several factors, including percent forest cover in the grid cell, the encounter of major illegal activities (on-field encounter with hunter, substantial logging), and the movement of rebels and armed forces. The sign detection data were collected for each 100-m segment on the trail as either detected (1) or not detected (0).

Selection of ecological and anthropogenic variables
We selected plausible ecological and anthropogenic variables relevant to the ecology of the species and management intervention of the study area (Table 1). We collected variables both at the landscape scale using remote sensing & GIS and during the field survey. For the landscape-scale variables, we obtained Landsat image (30 m resolution; https://earthexplorer.usgs.gov/) of the previous year (i.e., 2013) for the classification forest types and land use land cover of Ripu and its adjoining area. We classified seven land cover classes -(i) semi-evergreen mixed with moist deciduous forest (SE-MMD), (ii) secondary degraded forest (SDF), (iii) scrub (mostly degraded areas dominate by invasive like Chromolaena odoratum, Perthenium sp. and Ageratum conyzoides) (SC) (iv) water body (v) agriculture (vi) settlement (built-up areas, all types of road and home gardens) and (vii) sand (river banks & stream beds) (Fig. 2, Table S1 and Fig. S1). The landscape-level variables to determine the probability of habitat use across the grid cells includedelevation (ELE), distance to human settlement (DHS), distance to water sources (DWS), and remotely sensed vegetation types mentioned above. The elevation data was computed using a 90 m SRTM dataset (Jarvis et al. 2008), and we assumed that the different species might respond differently with varied elevation gradients. However, the study site is comparatively low-lying areas except the northern part bordering Bhutan hills.
The land use class agriculture and settlement were used to generate variable-distance to human settlement (DHS) through the Euclidean distance tool in ArcGIS 10.4. We hypothesized that distance to human settlement would act as major anthropogenic factor that would influence the habitat use of species . This variable is an alternative measure of prolonged disturbance faced by Ripu RF in the last two decades, including cattle grazing, hunting, and all kinds of resource extraction (majorly fuel wood) (Gray and Phan 2011). In contrast, distance to water sources is expected to positively impact elephant habitat use (Lakshminarayanan et al. 2016) and the abundance of wild ungulates (Lamichhane et al. 2020). Subsequently, the proportion of different vegetation types (SE-MMD, SDF and SC) may influence the habitat use of animals. All these continuous variables were extracted for each pixel within the 4 km 2 grid cell using Spatial Analyst Tools-Zonal Statistics in ArcGIS 10.4. Additionally, for leopard we used summed occupancy probability of prey (sambar, spotted deer, barking deer and wild pig) species in the study area along with other covariates mentioned above.
We also recorded on-field evidence of logs and/carts, tree felling, looping and cutting brunches, hunting signs (hunters seen, snares, axes, and wildlife carcasses), and fresh fire signs along 100 m segments. In addition, detection of the sign was believed to be associated with the substrate (SUB) condition, livestock (LS) presence, and effort (EF) in terms of trail length in each 4 km 2 grid cell. Substrate conditions noted were recorded as a categorical variable: soft soil (muddy areas, river & stream beds= 1), hard soil (forest paths with dry loose soil) = 2, and 3 for leaf litter. The presence of domestic animals, too, could influence detection since domestic animals can obliterate tracks of wild ungulates (pers. obs; Jornburom et al. 2020). The presence/absence of domestic livestock was noted as a binary variable coded as "0" for absence and "1" for presence in every 100 m replicate.

Capacity building of conservation volunteers and community awareness
The study began with a series of community consultations in the Kachugaon Forest Division in coordination with Green Forest Conservation NGO. This organization is one of the Community Forest Protection Forces in Bodoland Territorial Council to boost the forest protection forces in the Ripu RF. In recent years, the impact of community forces in reducing illegal logging and hunting has been encouraging in BTAD (Howrich et al. 2010). The Green Forest Conservation has around 100 conservation volunteers, of which few receive a small salary from the Bodoland Territorial Council (Howrich et al. 2010). To further strengthen the capacity of the conservation volunteers and their families, we provided training on biodiversity monitoring, supported with alternate livelihood options including hands-on training, and exposure-trip to give them ideas on the role of community protection, the scope of ecotourism followed by formation of village development committee (VDC) in the region. Apart from that, we encouraged conservation volunteers to participate actively in patrolling along with the forest department to regulate the anthropogenic activity in the study area.

Data Analysis
We used standard single-season occupancy models (MacKenzie et al. 2006) to estimate the occupancy and detectability of species using a sign survey. For the species with home range <4 km 2 , spatial autocorrelation models (Hines et al. 2010) were used to determine the smallest segment length at which spatial dependence could be considered negligible following Throne et al. (2011) andMcHenry et al. (2016). To allow estimation of detection probabilities for trail segments of different length, a suite of 12 sets of detection histories were constructed from the sign survey data, with each set using a different segment length (0.1-1.2 km). Detection histories were imported into RPresence for analysis of occupancy models. The single-season occupancy (null) models assuming constant detection probability, p(.), and occupancy, ψ (.), (MacKenzie et al. 2002) were generated for each set (0.1-1.2 km) of detection histories to determine the effect of segment length on the detection probability of species of interest (for details see McHenry et al. 2016). Parameters estimated from standard occupancy models using adjacent spatial sampling units may be biased by autocorrelation in availability for detection between replicated observations-this spatial dependence model recompenses for lack of independence by modeling the spatial autocorrelation among segments. The spatial dependence is captured by parameters θ, representing the probability that the species is present locally, given the species was not present in the previous segment; θ′ representing the probability that a species is present locally, given it was present in the previous segment, and θ π representing the probability of local presence in the first segment where no prior information on local occupancy is available. Bias due to autocorrelation between adjacent observations can be considered negligible when model selection by AIC (at the level of ΔAIC<2) favours the null model over the spatial autocorrelation model and θ′ ≈ θ (Thorn et al. 2011;McHenry et al. 2016;Lamichhane et al. 2020). We applied this model to the sets of standard detection histories for each segment length to test for betweensegment spatial dependence.
For Asian elephant and leopard (having larger home range >4 km 2 ), we used a spatial, first-order Markovian model (parameterized by θ, θ′, and θ π , Hines et al., 2010) to check the spatial dependence and then estimate the intensive habitat use probability. We modeled probabilities of habitat use (ψ) at 600 m (Thapa et al. 2019) as a function of the above-mentioned covariates using logit link function (MacKenzie et al. 2002). For these models, we used a constant model for θ, θ′ and θ π (Sunarto et al. 2012;Duangchantrasiri et al. 2015). Furthermore, detection histories were collapsed to 600 m segment length, generating replicates suitable for use with standard PRESENCE models. We aggregated substrate quality values using the mode of values in the collapsed segments. The field measured covariate and landscape-level covariates (extracted using GIS) were used to investigate the detection and occupancy/habitat use probabilities across the Ripu RF. All the covariates were screened for multi-collinearity (Table S2). Covariates with r ≤0.75 were either removed or not used in combination within the same model. For example: distance from the human settlement and elevation were not used in combination due to high correlation between the variables (r= 0.90).
Similarly, the proportion of SE-MMD and SC forest type were not used in the combination during model building.
We retained the variables that better explained the parameter of interest based on: ecological relevance, correspondence with earlier studies, ease in data collection across the study areas, and simplest in explanation of the model outcomes (parsimony). Prior to modelling, all continuous site covariates were scaled, (x -x ̅ )/SDx, to facilitate estimation of parameters using numerical optimization techniques and to make estimates comparable between parameters (Sunarto et al. 2012;Nath et al. 2019b;Jornburom et al. 2020).
We used package RPresence (that explicitly took spatial autocorrelation into account) and unmarked in R v. 4.0.2 to carry out the analysis. The spatial dependence models estimate four parameters, even without covariates (Hines et al. 2010), and given our sample size, we considered it likely that models containing covariates might fail to converge. Therefore, we fitted the spatial-autocorrelation model without the covariates and used the results to determine the cumulative segment length at which detection of the focal species ceased to be dependent on the preceding segment as mentioned above. In model building, first, we modeled detection probability (p) by varying candidate detection models. We explored covariate (substrate, presence of livestock, and effort) effects on detectability while holding the site covariates for occupancy constant. Using the best-fitted detection probability model, we modeled site use probability based on the information-theoretic approach (or a priori hypothesis) by varying combinations of covariates. We only used combinations of covariates as additive effects in the models. Interaction models were not used due to the complexity of general covariates to model different species without prior information on their distribution (Thapa and Kelly 2017). Finally, we ranked all models using Akaike's Information Criterion (AIC) and selected the best model based on the lowest AIC values. We considered all models with ΔAIC <2 as competing models (Burnham and Anderson 2002) and examined the β coefficient and model averaged β estimates of the covariates to test the significance of their effect on detection and occupancy probability.
The low detection of gaur (600 m replicates following Thapa and Kelly 2017) in our studied sampling (naïve occupancy = 0.13; presence noted on 10 grids out of 75 samples) grids has resulted in numerically unstable measures.
Previously this issue was well addressed for detection/occupancy probability is small or when the number of sites and number of visits per site is small (Moreno and Lele 2010). Hence, for gaur specifically, we used the Bayesian approach to estimate occupied habitat since the results using maximum likelihood estimators were numerically unstable. We conducted all computationally intensive model fitting necessary for Bayesian inference using MCMC sampling via packages R2jags and rjags in R statistical software to run the occupancy models (Plummer 2003, Su andYajima 2012). We used the uniform priors for probability of detection p~β (1,1) and occupancy probability ψ~β (1,1).
Uninformative-uniform priors are defined by the log-odds interval [-5,5] for all parameter distributions with three chains of 10,000 iterations each. As we are working on the logit scale and we have standardized the predictors (site covariates) to have SD = 1, therefore it would be highly unlikely if coefficients were outside the range ±5. Further, based on the calculation, we assessed model convergence using the " value (closer to 1.0, indicating a more plausible convergence) and from a visual inspection of chain trace plots (Link et al. 2012;see Fig. S2). We selected the final model based on the lowest DIC value since DIC has been widely used to rank hierarchical models based on their anticipated predictive performance, and it remains one of the most commonly used approaches for ranking models in the Bayesian context (Gelman et al. 2014;Barker and Link 2015;Hooten and Hobbs 2015). In addition, comparisons of Bayesian model selection methods for hierarchical models often include DIC, despite its limitations (e.g., Hooten and Hobbs 2015).
Finally, we estimated (projected for 144 grids= 576 km 2 ) occupancy for all the ungulate species and intensive habitat use maps of elephant, gaur, and leopard based on inferences made from the model (with lowest AIC and DIC value) outputs in ArcGIS 10.4. To estimate the overall occupancy and habitat use of studied species within the Ripu RF (144 grids), we weighed the cell-specific occupancy estimates within each grid cell (4 km 2 ) (Thapa and Kelly 2017). Further, to evaluate the status of illegal activities, we compared the number of confiscated materials and reports of hunting and poaching activities taken place in the previous years with the present effort.
On the other hand, for the large-ranging mammals (elephant and leopard) we carried out a spatial dependence model (Hines et al. 2010) at 600 m spatial replicates and compared it with the null model (MacKenzie et al. 2002).
The estimates of the Hines et al. (2010) model demonstrated a high degree of spatial dependence between replicates (Table S3)

Detection Probability
Sign detection probabilities were found to vary for leopard, spotted deer, barking deer and to some extent for elephant either due to the presence of livestock or substrate condition. Whereas, varying efforts (in km) made on each grid to record signs had no influence on detectability on the studied species (Table S4). Spotted deer sign detection probability was found to be significantly affected by the presence of livestock (Fig. 3). The probability of detection drops to ̂ = 0.0468 ± 0.0465 (SE) in the presence of livestock compared to its absence = 0.29 ± 0.05 (SE).. Similarly, the presence of cattle also had negative impact on the detection of elephant [presence of livestock, ̂ = 0.75 ± 0.08 (SE); absence of livestock, ̂ = 0.86 ± 0.05 (SE)] and leopard [presence of livestock, ̂ = 0.09 ± 0.05 (SE); absence of livestock, ̂ = 0.29 ± 0.05 (SE)]. We reasoned a priori that sign detection probabilities could be low in grid cells with more livestock presence, as opposed to grid cells with minimal livestock activity. In contrast, the probability of sign detection of barking deer [̂ = 0.76 ± 0.12 (SE)] was high on substrate bearing leaf litters and significantly low for leopard signs [̂ = 0.073 ± 0.0.074 (SE)] compared to other substrates (Fig. 3). Therefore, we used the effect of livestock presence on detection probability to examine habitat occupancy of spotted deer and elephant and substrate type on barking deer.
Whereas, the additive effect of the sign detectability on the presence of livestock and substrate type for leopard. For all other studied species viz., sambar, wild pig and gaur, we used constant detection p(.) probability while building habitat occupancy models.

Influence of covariates on occupancy probability and habitat use
Using the best-fitted detection probability model, we modeled site use probability based on the information-theoretic approach. The best predictive models on occupancy and habitat use for each species, including a complete set of models and estimates of regression coefficients in these models, are found in tables 2-4. The influence of covariates on occupancy probability of studied species shown in figures 4 & 5. The best model for sambar revealed that distance to human settlements (DHS) was the strongest (βestimate: 1.28 ± 0.44) ecological correlate predicting habitat use; probability of occupancy > 0.5 began at > 7 km from human settlements. The proportion of semi evergreen and moistmixed deciduous (SE-MMD) had a positive [p(.)ѱ(SE-MMD) βestimate: 0.73 ± 0.36)] influence on the occurrence probability of sambar. However, the second rank model showed that grids near to human settlements even with a high proportion of SE-MMD had negative influence on the occurrence probability (βestimate: -0.21 ± 0.63).
For spotted deer, there was a single best model with 48% AIC weight. The strongest ecological correlate was a negative response to anthropogenic activity (AAβestimate: -0.88 ± 0.41) and quadratic effect (SDF 2 βestimate: -0.88 ± 0.48) proportion of secondary degraded forest. Initially, the probability of occupancy increases with the increase in the proportion of SDF, however, declines are marked once the proportion of SDF measures greater than 0.3. In case of barking deer, DHS was found to have a quadratic effect (DHS 2 βestimate: -1.15 ± 0.57) on the occurrence probability with 46% AIC weight. The second ranked model also showed positive response to the distance to human settlements.
For Wild pig, the first top two best model showed negative response towards anthropogenic activity (AAβestimate: -0.64 ± 0.39) and distance from water sources (DWSβestimate: -0.55 ± 0.33). The cumulative AIC weight of first two model was 47%. On the other hand, proportion of SE-MMD was found to positively influence the habitat use of gaur in the study area. Whereas, for elephant habitat use DHS, DWS had positive and SDF had negative response. The cumulative best top three model had an AIC weight of 46%. However, all these covariates have very weak effect on habitat use.
For leopard, additive effects of prey occupancy and proportion of secondary degraded forest was the best model followed by proportion of semi evergreen and moist-mixed deciduous forest type. The first three top-ranked model (ΔAIC>2) had cumulative AIC weight of 58%. Prey occupancy (PREYβestimate: 4.80 ± 3.20) and semi evergreen & moist-mixed deciduous forest (SE-MMDβestimate: 1.33 ± 0.67) had positive impact and proportion of secondary degraded forest (SDFβestimate: -1.19 ± 0.68) had negative impact on the habitat use probability of leopard.
In addition, grids with more than 50% of SE-MMD forest were also found to have leopard habitat probability of more than 50%.

Estimates of occupancy and habitat use probability
Ungulates with small to medium body size, in the sampled grids, Barking deer has the highest occupancy probability  (Table 5). For the non-sampled grid cells, we used covariate information to draw inference using the best model. Except for spotted deer, the projected occupancy estimates produced lower but similar occupancy measures for other three species (

Community patrol
Conservation volunteers actively participated in the patrolling alongside forest department to strengthen the protection measures. Apart from routine daily patrolling, conservation volunteers additionally carried out 1260 hrs. of night patrol, 400 hrs of day patrol when informed on particular anthropogenic activity in the study area. In addition, ~250 km of motor-bike patrolling was done in the Athiabari Range and 300 km of patrolling made using four-wheeler vehicle along with the Forest department in the Central range of Ripu. The specific outcome of the patrolling effort was as follows-a total of 225 logs (mostly Sal), 36 bicycles, 25 bullock carts, and four vehicles were seized loaded with logs; two poachers caught with guns and five hunting machan was damaged during patrolling (Table S5, Fig S3).
We also noted that the majority of the fuel-wood extraction had occurred in the western part of the Ripu R.F. and logging activities were seen to be scattered due to selective logging of sal trees.

Community development
Conservation volunteers get a nominal amount from the forest department as remuneration which was not adequate to incur their family expenses. Additional under this study, conservation volunteers (n= 100) were also provided with daily wages for a year during the study period (total of 200000 INR). As mentioned here since they were supported only for a year, for long term sustenance we decided to support their family members through sustainable income generating livelihood activities based on the skills they possess. So in order to initiate the process we constituted three Village Development Committee (VDC) in Kachugaon, Raimona and Jambuguri area in the southern boundary of the study site. Subsequently, we conducted a consultation workshop in order to assess the skill and their willingness for preferred livelihood activities of the family members of conservation volunteers. Based on the information gathered, livestock rearing was found to be the preferred livelihood activity. We could assess that since 90% of the families were indigenous tribes who traditionally have the skills to rear livestock they expressed their willingness for piggery and dairy. So, to further enhance their skills, we had conducted six training programme on piggery and dairy farming with the help of professionals. In ordered to enhance the livelihood of the conservation volunteers, 29 members were provided with Dairy (one heifer each), 71 members provided with piggery (two piglets each). At the beginning the families of conservation volunteers were earning around ~2000 INR month -1 but now after our intervention in 2014 their monthly income had gradually increased to on an average ~11000 INR month -1 (9-10 liter day -1 milking cow -1 ) beneficiaries associated with dairy as per our assessment in 2018 and it is still continuing. In case of beneficiaries associated with the piggery earned ~44000 INR year -1 . Furthermore, to boost their motivation, an exposure trips to Khonoma, Nagaland and Sijusa, Arunachal Pradesh (India) was organized for conservation volunteers to educate and sensitize them on the role of local communities in the protection and conservation of forest. As both these sites in northeastern part of India has gained appreciation for community conservation practices and sustainable tourism.
Apart from handheld GPS training, the members were also provided with camera, binocular and field guides (on butterflies, birds, reptiles and mammals) to generate awareness on biodiversity monitoring, protection and conservation.

Discussion
Understanding species-habitat relationships and impacts of anthropogenic factors on species site use patterns is a key to prioritize conservation actions. In Ripu, for the first time, we used the statistically robust occupancy modelling approach to examine site occupancy and intensive habitat use of mammals. We were able to document that Ripu has the potential to hold primary prey for large carnivores like tigers and leopards along with mega-herbivores like gaur and elephant. As hypothesized, studied mammalian fauna showed varied but species-specific responses to ecological and anthropogenic factors. Species responded variably to different covariates, indicating a need to consider management actions according to the relative importance of species as per conservation requirements. This implies that although conservation paradigm has shifted towards landscape-level conservation (Bhattarai et al. 2017), species conservation approach is equally vital and consideration of species-specific habitat requirements is paramount to successful conservation planning.
The present study used tracks and sign detection for occupancy survey. Therefore, the underlying ecological processes that regulate distribution and habitat associations of ungulates and predator species can be unnoticed without estimating detection probability, especially when surveys are conducted using animal signs. Had we not modeled sign detection, we would have underestimated the negative impact of livestock on spotted deer, leopard and to some extent elephant on detection probability. Similarly, as expected, substrate type influenced the detectability of leopard signs and detection probability was low where substrate dominated by leaf litter. In the case of barking deer, the high probability of detection in leaf litter is associated with the preference towards dense vegetation, and this species avoids open areas like river beds for their movement. Henceforth, by estimating detection, we increased our estimate of occupancy for most of the studied species and reduced bias in predicting occupancy (Tyre et al. 2003;Rota et al. 2011;Lahoz-Monfort et al. 2014). Our occupancy estimate revealed that barking deer was the most widely distributed medium-sized prey in Ripu habitat followed by wild pig, sambar, and spotted deer. As expected, high human activities in the forest edges bordering human settlements have a low occurrence probability of sambar and barking deer. Both of these species prefer dense vegetation McCullough et al. 2000), and majority of the degraded areas are adjacent to human settlements in our study area. The forest areas adjacent to human settlements experience prolonged pressure from various anthropogenic activities (grazing, resource extraction, hunting) compared to areas bordering Phibsoo WLS in the north. Previous studies also showed that sambar increased with increasing distance to human settlements, supporting our findings that the species is sensitive to disturbance (Kushwaha et al. 2004).
Spotted deer are habitat generalists (Schaller 1967). Their distribution in Ripu is associated with the initial increase in proportion of secondary degraded forest and negative response with an increase in anthropogenic pressure within grid cells. However, it was also observed that the probability of occupancy declines in the grid cells where proportions of secondary degraded forests measure doubles. Previous studies from Terai (Chauria, Nepal) region showed mixed-moist deciduous forests limit the distribution of this species (Thapa and Kelly 2017). However, the majority of wild ungulate density in Indian forests is contributed by spotted deer (Khan 1995;Sankar 1994;Bagchi et al. 2003). The species limit its distribution in the western Assam, and last remaining population was Ripu & Chirang RF, and recently photographic evidence was also found from further east in Panbari, Manas National Park (Bhatt 2018). Though this species is a key prey for large predators like the tiger, leopard, and dhole (Schaller 1967), we witnessed frequent hunting incidents (Fig. S3). The wild pig is a generalist species (Bratton 1975) with a highly plastic diet contributing to their wide distribution (Ballari and Barrios-Garcı´a 2014). The positive association with water bodies in our study area could be associated with vegetation it feeds on. During winter, above-ground resources may get scarce, and pigs might rely on rooting, which can be more easily accessible to nearby water bodies. Hunting wild pigs using snares and other accessories are common in the region (pers. obs.), and the meat has high demand in this belt of India. Hence, wild pigs' occupancy estimate is negatively associated with the anthropogenic activity in the study area. However, studies also stated that the species is particularly resilient to hunting pressure due to its high reproduction rate (Choquenot et al. 1996;Pepin et al. 2017).
As expected, the leopard habitat use was governed by the cumulative prey occupancy and proportion of available SE-MMD forest types. Though leopard is a broadly distributed large carnivore adapted to a multitude of habitats and to some extent lenient to human-induced anthropogenic pressure (Athreya et al. 2016). In Ripu, the species showed a negative response towards degraded forest adjacent to human settlement and predictive occupancy probability showed that the species were mostly confined to areas bordering Phibsoo WLS. Besides our study, a camera trapping exercise carried out in Phibsoo WLS in the same year (2014) reported that leopard was one of the most common out of six species of felids they recorded (Anonymous 2018). The study also recorded tigers on nine occasions at six camera trapping stations of Phibsoo. Most of these captures were from bordering Ripu RF in the lower and middle foothills zone. In Terai region, mainly in the foot-hills of the Himalayas, leopards coexist with tigers in the protected and buffer zone areas. In the absence of a tiger, the positive influence of prey has indicated the importance of wild ungulates in minimizing human-leopard conflict. Despite the high adaptability of leopards, they require a comparatively large area with abundant prey for survival, thus, threatened by prey depletion, poaching, landscape fragmentation, and conflict with humans (Athreya et al. 2011;Crdillo et al. 2005;Kissui 2008).
The overall occurrence probability of gaur was low in the study area and mostly confined to areas in which proportion of semi-evergreen and moist-mixed deciduous forest was high. Compared to all other studied species, the probability of detection was also very low in our study area. Previous studies from the adjacent landscape stated that proximity to human settlement negatively influences the occurrence probability of gaur (Zangmo et al. 2018).
Observations of senior forest frontline staffs indicated that gaur was indeed widespread in the Ripu and adjacent areas before insurgency (pers. com.). Habitats in our study area vary in terms of vegetation type and intensity of anthropogenic pressure. The habitat use of elephants is much more complex than we observed. Although elephants used all the available habitats (naïve occupancy = 96%) in the study area, detection probability was low in grids, where cattle presence is high. Fine-scale habitat use parameters are vital to individual protected area managers seeking to recover habitat conditions for elephants (Buij et al. 2007;Lakshminarayanan et al. 2016). Irrespective of high anthropogenic pressure, the elephants were bound to use the forest of Ripu since the northern side (Phibsoo WLS, Bhutan) has rugged mountainous terrain, and the eastern part of the land highly encroached. Permanently bifurcates Chirang and Manas RF. Our current study and other studies elsewhere specify that Asian elephants are sensitive to human activities (Goswami et al. 2014;. However, there were reports of crop-raiding (mostly paddy) of elephants, during winter months: October to December.
Sign-based occupancy survey is the only practical assessment strategy found applicable since anthropogenic activity is too high and movement of people is not restricted in the study area. The site also lacks permanent antipoaching camps/guard camps inside, unlike the core of the Tiger Reserve (i.e., Manas National Park) to monitor cameras regularly. Although, monitoring programs are often criticised for a dearth of scientific rigour, arising out of poor design or insufficiently robust statistical analysis (Mattfeldt et al. 2009). Our current study specifically accounts for both imperfect detection and robust spatial autocorrelation measures at different segment lengths. However, we could not address false positive detection for leopard signs (only for scats) since a study (Naha et al. 2020) claimed that scat identified in the field has an error rate of 50-80%. Moreover, species identification using genetic analysis was beyond the scope of the study since resources were limited. Therefore, in future, we recommend systematic camera trap studies to identify elusive and nocturnal mammals in the study area. However, field observations and personal communications with forest frontline staff revealed 25 species mammals from Ripu RF (Table S6). We would also like to highlight a few limitations: the species-specific responses we observed could thus be different during the summer season when forage quality and quantity are generally abundant for ungulates, and water is not a limiting factor. Studies looking at seasonal responses to ecological factors would improve our understanding of habitat use of these species.
Like most other societal challenges, conserving natural resources during conflict (insurgency) is complex and case-specific. In the case of Ripu, biodiversity suffers from the general breakdown in both law and infrastructure that facilitated various anthropogenic activities. Along with the cattle grazing, fuel wood collection and poaching, selective logging of sal trees was one of the key illegal activities in the study area. Prior to 1992 no logs were sized by the Assam Forest Department. It was all started after the formation of Bodoland Protection Force, and subsequently, they confiscated all the materials associated with illegal activities. The estimated value of these materials was of USD 100,000 (for details, see Howrich et al. 2010). In the present study (Jan 2014 to June 2015), the number of confiscated materials was low compared to previously reported during Nov 2006 to 2008. Looking at this massive destruction, apart from community vigilance, forest department needs to take immediate action to set up anti-poaching camps in strategic locations to ensure security for wildlife and its habitat.

Conservation Implications
In recent years, the government of India has encouraged community participation in the conservation process (Pathak, 2002). Therefore, community co-management can be the solution to involve communities as influential partners in protecting wildlife and its habitat, which is being searched for by various governments and NGOs (Borrini-Feyerabend et al. 2004). Prohibiting cattle grazing within the Ripu would decrease forage competition and increase habitats for wildlife. Our community-based conservation initiatives could encourage stall feeding of livestock and address illegal activities in RF, but such programs require continuous government incentives. Similarly, our approach towards alternative livelihood support (piggery and dairy farming) can only be effective if the government and other agencies continue to generate funds and encourage locals till they self-sustain. Other anthropogenic activities, such as the collection of NTFPs, especially fuelwood, are often linked to fishing and hunting. To engage more people, ecotourism was thought to be one sector that was not well explored due to ethnic violation in the region. However, we provided exposure to the youths and tried to motivate them for sustainable tourism activities in the region as an alternative livelihood opportunity. Another option would be to include more local villagers as a part of an effort to strengthen and broaden GPS-patrolling in the region. Given that Ripu-Chirang Elephant Reserve is the viable source population of elephants in mainland Southeast Asia, combining the current top-down approach of GPS patrol with bottom-up community participation appears to be a hopeful strategy for elephant, tiger, co-predator, and prey's recovery across the region. In India, National Tiger Conservation Authority (NTCA) have already implemented M-STrIPES (an appbased platform to assist effective patrolling, evaluate ecological status and mitigate human-wildlife conflict in and around tiger reserves). We recommend that Assam Forest Department actively implement the same in Ripu to ease decision-making and adaptive management. Moreover, the Indian Government Border Security Force has already established camps in Kachugaon, Ripu, and Chirang Reserve Forests. The Security Force is an armed central government paramilitary force whose command is to control illegal activities in the Indian border areas. They helped forest department to intercept illegal logging in areas around their camps and in future joint patrolling can help further regulating the illegal activities in Ripu.
Ripu is on the road to recovery, and recently (June 5, 2021), a major portion (422 km 2 ) of the reserve forest was declared as Raimona National Park. This extended favourable social and political climate and widespread support for conservation activities would help secure the biodiversity of this region, and the local community may benefit from the same. Our study revealed that Ripu has a high proportion of occupied area by prey species of large carnivores.
Recently (February 2020), a tiger has been camera trapped in Ripu bordering Phibsoo WLS (Goswami 2020). Territorial Council, regional to international organizations, educational institutes, and local communities with stakeholders at a number of levels need to bridge the gap for the long-term conservation of Ripu. In the meantime, activities like education and outreach initiatives can play a vital role in promoting positive attitudes among various communities living adjacent to the area. Livelihood is also a major issue in this conflict driven landscape. There are not much livelihood opportunities for local communities other than daily wage earning or extracting resources from the forest and agricultural land being limited. Thus it is very important to address the livelihood issue in order to boost conservation actions and enhance community participation. Hence, activities that encourage community participation and further strengthen forest officials and conservation volunteers would be the key to securing the threatened habitat of Ripu.

Competing Interests
The authors have no relevant financial or non-financial interests to disclose

Data Availability
Data and codes will be provided on request.
Consent to participate: We give our consent to participate in the publication process. Consent for publication: We give our consent for the publication of submitted manuscript ***** . Table 1. Factors hypothesized to influence patterns of occupancy (ѱ)/habitat use (4 km 2 ) and detection probability (p) of mammals in Ripu Reserve Forest, Assam, India.

Covariates Description Justification of selection of covariate
A priori hypothesis p ѱ Distance to water sources (DWS) Generated a surface by calculating the Euclidean distance from rivers, streams (both permanent and seasonal) and stagnant water bodies The presence of water bodies is expected to have positive impact on abundance and hence on the site use of elephants and other wild ungulates under study (Lakshminarayanan et al. 2016)

Elevation (ELE)
Computed using the SRTM digital elevation model-90m Elevation can both positively and negatively influence the habitat use of the studied species Kelly 2017, Phumanee et al. 2020)

Semi-evergreen and mixed-moist deciduous forest (SE-MMD)
Proportion of intact semi-evergreen and Mixed-Moist Deciduous forest extracted from LULC map prepared for the study area Different vegetation structure associated with forest types could influence the habitat use of studied mammalian species. Previous studies from western Terai stated that barking deer prefer dense vegetation , sambar prefer mixed deciduous forest (Karanth and Sunquist 1992) and spotted deer prefer open grassland intermixed with forest (Dinerstein 1980).

Secondary degraded forest (SDF)
Proportion of secondary degraded forest with canopy openings extracted from LULC map prepared for the study area Encounter rate of cattle's and their signs Livestock raising is common in villages adjacent to study area, and their presence was likely to negatively impact the ungulate populations through competition for forage (Chaiyarat & Srikosamatara 2009;Jathanna et al. 2016, Thapa andKelly 2017).

Effort (EF)
Total km of search paths walked during the sign-survey in each of 4 km 2 grid cells.
Detection of sign would be influenced by survey effort (Harihar and Pandav 2012). +

Substrate (SUB)
Subjective grading of substrate quality for detecting tracks. 1= Soft soil, 2 = Hard soil, 3 = Leaf litter Soft soil expected to increase the detection probability of animal tracks compared to hard soil and leaf litter (Lamichhane et al. 2020). Further, presence of leaf litter on the trail can influence the detectability of pellet and scat of animals.

Livestock (LS)
Presence of domestic animals (1,0) 1= Presence, 0 = Absence Presence of domestic animals was expected to influence detection probability as it would lower abundance directly through competition for forage and indirectly through human-induced impacts Nichols 2003, Jornburom et al. 2020). Furthermore, a herd of cattle can also influence the sign detected on the forest floor -

Spotted deer
Barking deer Gaur Figure 4. Association between the highly influential covariates (based on estimates of regression parameters (β) and 95 % CI from the best models) and the probability of sambar, spotted deer, barking deer and gaur occupancy in Ripu Reserve Forest, Assam, India.

Wild Boar
Leopard Fig. 5. Association between the highly influential covariates (based on estimates of regression parameters (β) and 95 % CI from the best models) and the probability of wild boar and leopard occupancy in Ripu Reserve Forest of Assam, India.   Table S3. Comparison of parameter estimates for spatial dependence occupancy models based on detection histories using different segment lengths of trail. θ is the probability that the species is present locally, given the species was not present in the previous segment and θ′ is the probability that a species is present locally, given it was present in the previous segment. Δ AIC is the difference between spatial autocorrelation occupancy models [Hines et al. 2010 (ѱ(.)