Identifying important connectivity areas for the wide‐ranging Asian elephant across conservation landscapes of Northeast India

Connectivity is increasingly important for landscape‐scale conservation programmes. Yet there are obstacles to developing reliable connectivity maps, including paucity of data on animal use of the non‐habitat matrix. Our aim was to identify important connectivity areas for the endangered Asian elephant Elephas maximus across a 21,210 km2 region using empirical data and recently developed animal movement models.


| INTRODUC TI ON
Connectivity, or functional linkages between populations or resource patches (Vasudev, Fletcher, et al., 2015), is crucial for maintaining species viability and ecosystem functioning, particularly in the light of ongoing environmental change (Doerr et al., 2011;Staddon et al., 2010). Connectivity, forged by dispersal of individuals across populations or habitats (Clobert et al., 2012), determines where species are (Peterson et al., 2011), as well as their ability to adapt to novel environments (Holderegger & Wagner, 2008) and climate change (Parmesan & Yohe, 2003). Connectivity shapes inter-species interactions such as competition, predation and seed dispersal, ultimately influencing ecosystem health and function (Orrock et al., 2006;Staddon et al., 2010). Thus, maintaining connectivity is a critical need for landscape-scale conservation programmes.
However, animal movement is severely restricted in the Anthropocene (Tucker et al., 2018). Furthermore, evidence across taxa and landscapes demonstrates the negative impacts of reduced connectivity on species and communities (Fletcher et al., 2016). Connectivity is potentially impeded by a number of factors, including linear infrastructure such as roads and rails (van der Ree et al., 2015), land uses through which animal movement is restricted, or where dispersing animals face heightened risk, and human habitation (e.g. Kramer-Schadt et al., 2004;Thatte et al., 2018). Antagonistic responses of people to dispersing animals, such as chasing of "conflict" animals, and certain mitigation measures for human-wildlife conflict, such as fences, can also pose a barrier to connectivity (Ghoddousi et al., 2021;Goswami & Vasudev, 2017;Laundré et al., 2001). Yet, for most species, we have little quantitative knowledge of the impacts of these barriers to connectivity (but see Fletcher et al., 2019;Osipova et al., 2018); this, in turn, means we are limited in our ability to successfully achieve connectivity conservation targets (Vasudev, Fletcher, et al., 2015). Clearly, we need a better understanding of animal movement across heterogeneous landscapes, and more accurate modelling of connectivity to develop effective connectivity conservation strategies and targeted interventions (e.g. overpasses to mitigate impediments due to linear infrastructure).
Connectivity is shaped by intrinsic characteristics of species, the external environment, and the spatial configuration of landscape elements (Nathan et al., 2008;Vasudev, Fletcher, et al., 2015).
Interactions between dispersing animals and the environment determine the resistance of the matrix-non-habitat sections of the landscape-to the animal movement (Taylor et al., 1993). Thus, certain landscape elements, such as rivers or roads, can pose barriers to connectivity, whilst others, such as woodlands may serve as corridors for animal movement. More recently, barriers and corridors have been seen as two sides of the same coin, representing a continuum of the role that different areas play in facilitating (or impeding) landscape connectivity (Panzacchi et al., 2016).
Whilst the concept of matrix resistance is increasingly incorporated into connectivity modelling (Fletcher et al., 2016), parameterising resistance-based models remains a non-trivial issue (Panzacchi et al., 2016;Zeller et al., 2012). Observations of the large-scale and infrequent movement events that typically comprise species dispersal have become more feasible with the advent of telemetry (Hooten et al., 2017). Nevertheless, telemetry remains inapplicable to a large number of species and regions due to high costs and logistical constraints, factors that often also limit sample sizes in those studies that do incorporate the approach. Thus, resistances are often parameterised through expert opinion or proxies, typically invalidated by field information on animal use of the matrix (Sawyer et al., 2011;Zeller et al., 2012). An alternative option for easily detectable species such as birds or large non-elusive mammals is crowd-sourced or citizen-science data (Brown et al., 2018;Frigerio et al., 2018).
Such information is useful as it allows researchers to cover large landscapes and temporal scales adequate to attain sufficient sample sizes of infrequent dispersal events. The Asian elephant Elephas maximus is one such non-elusive, detectable and easily identifiable wide-ranging species, for which connectivity is a critical conservation need (Goswami & Vasudev, 2017;Goswami et al., 2021).
The suite of available connectivity models that can incorporate matrix resistances represents two ends of a spectrum. At one end are least-cost models, which assume that dispersers move optimally, with complete knowledge of the landscape (Adriaensen et al., 2003). At the other end are random-walk models, such as those based on circuit theory, which assume that dispersers have information solely on their immediate environment (McRae et al., 2008). One advantage of the latter has been its ability to account for path redundancy, thereby providing a more accurate representation of landscape connectivity (McRae et al., 2008). The likely reality though is one that lies somewhere between the two extremes, whereby dispersing animals attempt some degree of non-optimal exploration of the landscape, have some knowledge about their surrounding landscape, and make movement decisions accordingly. The randomised shortest path (RSP) framework offers this balance, explicitly incorporating an information parameter θ, which controls the trade-off between random exploration of the Main conclusions: Fine-scale mapping of connectivity, using empirical data and realistic movement models, such as the approach we use, can provide for informed and more effective landscape-scale conservation.

K E Y W O R D S
Assam, corridor, crowd-sourced data, dispersal, forests, fragmentation, linear infrastructure, movement models, randomised shortest path, resistance mapping landscape by dispersers, and optimisation of movement (Kivimäki et al., 2014;Panzacchi et al., 2016). In so doing, the RSP framework offers a modelling approach that can more realistically represent animal movement across heterogeneous landscapes (Panzacchi et al., 2016).
We use the RSP framework in combination with crowd-sourced data to prioritise important connectivity areas for the wide-ranging Asian elephant across a large region in Northeast India. We estimated resistances using crowd-sourced information, systematically collected via a questionnaire survey. We then modelled animal movement across the region using the RSP framework. We parameterised the movement model using estimated resistances and modelled multiple scenarios of disperser information on their environment. We compared movement models parameterised solely by land use, and those that incorporated more detailed information on the landscape. Finally, we overlaid two types of threats to connectivity on modelled animal movement: (a) linear infrastructure, including roads and railways, and (b) human-elephant conflict hotspots, as recorded from our surveys. We discuss the relevance of our findings for landscape-scale conservation of this endangered species.

| Study area
Our study area encompassed a large 21,210 km 2 region in the foothill plains of central Northeast India (Figure 1a). The region lies on the southern bank of the Brahmaputra River in the state of Assam and includes an adjoining northern portion of the state of Meghalaya. Two important Elephant Reserves (ERs) are situated fully, or in part, in our study region, namely, the 3,270-km 2 Kaziranga-Karbi Anglong ER inhabited by an estimated 1,746 elephants at the minimum (Goswami et al., 2019), and the 2,740 km 2 Dhansiri-Lumding ER (which includes the Kholahat Reserve Forests) home to an estimated 205 elephants (Rangarajan et al., 2010). Elephant Reserves in India are important elephant conservation landscapes, comprising protected areas (e.g. protected National Parks or Wildlife Sanctuaries), other forests with lower degrees of protection (e.g. Reserved Forests), and other non-protected lands (e.g. agricultural fields, plantations). ERs are declared in recognition of the need for landscape-scale conservation efforts for the wideranging Asian elephant. ERs are not fenced, but individual private lands, or forests within ERs, may have fences along their respective boundaries. Our study region also includes the forests in Garo Hills, Meghalaya, and the Rani landscape of Assam, which are elephant conservation areas of known importance. The region receives an average rainfall of about 1,500 mm annually, with the monsoon season (June-September) accounting for most of the rain. The grasslands, moist evergreen forests and deciduous forests that form elephant habitat in this region are fragmented. These elephant habitats are surrounded by a matrix comprising mostly of rice paddy Oryza sativa fields, tea Camellia sinensis and sal Shorea robusta plantations, and human habitations ( Figure 1a). Rice paddy fields are open lands with few to no trees, and they are often small landholdings and unfenced. Tea plantations, depending on management of the plantation, may offer varying degrees of tree cover, and maybe fenced with bamboo, wire, or solar fences. Dispersing elephants traverse this matrix to access resources and refuge (e.g. Goswami et al., 2021). It was this matrix that thus formed the primary focal area of our study.

| Field survey
Our aim was to sample large parts of our study region, which has high human population density, to assess elephant use-however infrequent-of the matrix. For a detectable and identifiable species like the elephant, we decided to use crowd-sourced information, systematically collected via a questionnaire survey, so as to sample a large area at a relatively fine scale, whilst obtaining adequate reports on elephant use of the matrix.
To sample the region in a systematic fashion and attain adequate spatial coverage, we divided the geographical extent of our study area into a grid network of 286 cells, each of size 25-km 2 . Thus, our sampled grid effectively covered an area of 7,150 km 2 . This grid was not used for analyses, but guided our sampling such that we achieved spatial spread and even coverage of our entire sampled area, thereby minimising any potential sampling bias. For our assessment of resistance, we restricted ourselves to respondents from Assam, as the social-ecological context of this state and the neighbouring state of Meghalaya are quite different. Assam also comprised most of our study region and housed the majority of our respondents. We overlaid a buffer of width 10-km around survey locations, resulting in our total study area of 21,210 km 2 (Figure 1a). We chose a buffer of 10 km, as it is comparable to elephant home range and space-use radii reported in previous studies (Goswami, Vasudev, et al., 2014;Goswami et al., 2019). Our survey was part of a larger effort to obtain information relevant to landscape-scale elephant conservation in Northeast India.
We systematically interviewed respondents, primarily farmers, from the survey area between June 2019 and March 2020. We used the grid network as a spatial guide, and interviewed 4-8 respondents per cell, chosen opportunistically. Interviews were conducted at locations separated by at least 400 m, to ensure the independence of responses and adequate spatial coverage within grid cells.
As a further precaution against pseudo-replication, we confined our analysis to a subset of respondents who were separated by 1 km or more, aligning with our resolution of analysis. Each respondent represented a sample point in our analysis.
The questionnaire surveys were conducted by trained field teams in the local languages, after obtaining consent from respondents. We undertook a structured interview, instrumented using the Open Data Kit (Hartung et al., 2010). We questioned each respondent on elephant presence over the past year within their neighbourhood, which we defined as a buffer of 1 km around the respondent. Where possible, we validated the presence of elephants by documenting signs such as footprints or dung at the exact sites of movement. If a respondent sighted elephants multiple times in a year, we collapsed these into a single sighting to represent whether the sampled location was used, or not, by elephants. A note on the presence of captive elephants was also kept, so as to not mistake these elephants as those from wild populations. We recorded local land use, which we used for groundtruthing land use/land cover classification, described further below.
Lastly, we asked respondents if they faced conflict with elephants, in the form of crop loss, property damage, human injury or loss of life.

| Modelling elephant space use
We assessed the influence of environmental and anthropogenic variables on the probability of elephants using sampled locations in the landscape, that is, locations corresponding to the field questionnaire survey described above. This provided us with a basis to parameterise landscape resistance (see Section 2.4). Each sampled location was associated with information on whether respondents had detected elephants or not, during the past year. We additionally extracted a set of ecologically relevant covariates, as described below, within a 1-km 2 square window centred around each point location; this buffer aligned with the dimensions of the elephant detection data, as well as with our final predicted resistance maps.
We used the following remotely sensed covariates. We derived the Enhanced Vegetation Index (Huete et al., 2002) which represents   vegetation cover, or greenness of locations, at a 20-m resolution, using Sentinel-2 Multi-Spectral Instrument Level-2A images of resolution 10 m to 60 m (Drusch et al., 2012;Gorelick et al., 2017) from the Google Earth Engine (Gorelick et al., 2017). The data covered the period Feb 2019 to Jan 2020, which largely overlapped with our field survey period, and had a revisit time of 5 days. We excluded imagery with greater than 10% cloud cover from the resultant single median composite image. We aggregated the derived EVI data to a resolution of 1 km using the zonal statistics function of Google Earth Engine. We obtained information on elevation and slope from Advanced Land Observing Satellite ALOS World 3D -30 m Version 3.1, a global digital surface model dataset (Tadono et al., 2016). We extracted spatial data on human population density from the WorldPop data set (www.world pop.org). We extracted a road (comprising national and state highways) and railways network from the online information portal India Under Construction (Nayak et al., 2020), and validated the layers visually with satellite imagery. We then calculated the linear infrastructure density (including both roads and rail) at the same 1-km 2 pixel level across the landscape using QGIS 3.12.0 (QGIS.org, 2019). We also calculated the Euclidean distance to elephant habitat (forests and grasslands), and to human habitation or settlements.
Land-management practices and regional policies are often broad-scale, and dependent on the predominant land uses or devised and implemented based on specific land uses. To do this, we further classified our study region into userdefined LULC classes, namely: (i) cropland (predominantly rice paddy), (ii) closed-canopy plantations (including rubber, teak and sal), (iii) tea plantations, (iv) forests, (v) grasslands, (vi) human habitation, (vii) water and (viii) barren land (predominantly river sand islands), at a 20-m resolution. We collected a total of 4,547 training points from field surveys and photo interpretation of satellite imagery. We used spectral bands from the remotely sensed imagery described above, as well as the derived Enhanced Vegetation Index. To improve classification, especially in regions of high cloud cover, we additionally used data from the Sentinel-1 SAR GRD C-band (Erinjery et al., 2018;Torres et al., 2012). We also derived a series of textural indices, including the Index-based Built-up Index (Xu, 2008), and the Modified normalized difference water index (Xu, 2006), which have been shown to provide for more accurate classification of specific LULC classes (Poortinga et al., 2019). We used the Random Forest algorithm (Breiman, 2001) for classification, trained with 500 decision trees, including variables listed above, after checking for collinearity and variable importance.
We extracted the modal or most represented LULC within the same 1-km 2 square window described above. Sampled locations with modal LULCs other than closed-canopy plantations, tea plantations or croplands were removed from the analysis due to low sample size; these represented 4.6% of sampled points.
We used logistic regression to model the effect of the abovedescribed covariates on elephant presence at sampled locations. We constructed biologically plausible models, which included individual covariates and all permutations of additive effects of the same (Table S1). We tested for collinearity amongst covariates before including them in our models. Since LULC was correlated with the calculated Enhanced Vegetation Index, we did not include these two covariates in the same model; closed-canopy plantations had higher EVI than tea plantations, followed by croplands (rice paddy fields).

| Parameterising landscape resistance
For each 1-km 2 pixel in our landscape, we extracted values of the covariates used in our models, as described above. We then predicted the probability of elephant use of these pixels by model averaging the estimated probability of use parameter, using model weights w i , across all models in our model set (Burnham & Anderson, 2002). We considered this probability of elephant use of a grid cell to represent its conductance; resistance was calculated as the inverse of the conductance (McRae et al., 2008). We thus obtained a model-averaged resistance map of our study region.
To assess resistance based solely on LULC, we considered the probability of elephant use of a pixel-as estimated by the LULConly model based on the LULC of the pixel-as conductance.
Resistance was calculated as the inverse of this value. We scaled these resistances by a factor of 10-that is, we multiplied the calculated resistances by a factor of 10-to better highlight differences on a resistance scale of 1-100. This also placed the LULC-based resistances on a scale comparable to model-averaged resistances, and thus facilitated comparison of the two scenarios. Resistances for forests and grasslands-which represent potential habitat and are known to have negligible resistance-were set to a value of 1; resistance for habitation-which is known to be actively avoided by elephants-was set to the highest resistance value as obtained from the model-averaged resistance map. Water-in our landscape, the Brahmaputra River-and barren land (sand islands in the Brahmaputra River) were considered to represent barriers to elephant movement.

| Modelling connectivity
We identified source patches as important elephant habitats distributed across the study region with known elephant populations, primarily from Forest Department records, available literature and secondary information (e.g. media reports of elephant presence or conflict). These were typically individual protected areas or multiple contiguous protected areas that spanned more than 100 km 2 ; we also included two smaller habitat patches in the important Kaziranga-Karbi Anglong ER, a landscape known to house a substantial elephant population (Goswami et al., 2019). In total, we identified 10 source patches, nine of which are forest habitat, and one, a forest-grassland mosaic habitat.  and hence it was included as a source. Source patches were digitised using QGIS 3.12.0 (QGIS.org, 2019). The above-described source patches also represented our set of destination patches.
The RSP model works on a network characterised by nodes, which represent each pixel in our study region, and edges, which determine the links between pixels. Each pixel in our study region was linked with its eight neighbours. The inverse of the mean resistance of the two neighbouring pixels was assigned as the edge weight, scaled to values close to 1.
We used a series of six θ values altogether-5, 1, 10 -1 , 10 -5 , 10 -8 , 10 -12 -as we had source patches that were separated by Euclidean distances ranging from as low as half a kilometre between Kaziranga National Park and the Karbi Anglong forests, to as high as 337 km between Nambor Wildlife Sanctuary and Garo Hills. The θ value determines the trade-off between optimal (when θ values are high) and random (when θ values approach 0) movement. Intermediate θ values then would represent an arguably more realistic scenario where dispersers have some (but not complete) knowledge of the landscape; that is, dispersers make movement decisions based on partial information limited to a perceptual window around them (henceforth referred to as a scenario of "partial disperser information"; see informed dispersal: Clobert et al., 2009). The absolute value of θ that corresponds to each of these scenarios (optimal movement → partial disperser information → random walk) would depend on the size of the modelled landscape and corresponding resistance values (Panzacchi et al., 2016), relative to the animal's ability to perceive and navigate the landscape (for example, its perceptual window and ranging capabilities). Our landscape had source population pairs, which were separated by small, as well as large distances; hence we considered θ values of 5 and 1 to be representative of optimal movement, at small and large spatial scales, respectively. Given the scale of our study region and the ecology of our study species, we considered a θ value of 10 -1 and 10 -5 as scenarios of partial disperser information. Finally, we considered θ values of 10 -8 and 10 -12 to represent random walks or near-random walks, with dispersers only having knowledge of their immediate surroundings. Similar θ values have been used elsewhere (Panzacchi et al., 2016).
We modelled animal movement between each sourcedestination pair, for each pixel; this value represents the expected number of passages across all paths between the source and destination node. We obtained a cumulative passage map, for each θ value, by summing across all source-destination pairs.

| Model comparison and validation
In addition to comparing model support for the LULC-only model with other models in our logistic regression model set via Akaike's Information Criterion, as described in Section 2.3, we made the following comparison to assess the utility of adding complexity to con- Ideally, we would have validated model predictions with independently collected data on elephant movement. In the absence of such data, we investigated the passage value corresponding to each of our recorded elephant presence points (Peck et al., 2017).
To do this, we first cropped the passage map from the RSP model for each θ value to our sampled area, defined as a minimum convex polygon around our sampled points. We excluded habitation and forest as our sampling framework excluded these areas. We converted passage values in the resultant map to percentile-passage values. We extracted mean percentile-passage values in a 1-km 2 buffer around each of our recorded presence points. We compared these percentile values, corresponding to elephant presence points in the matrix, across θ values. We did this for predictions from the model-averaged resistances, as well as from those based on LULC alone. We distinguished low-frequency sightings (<100 elephants) from high-frequency (≥100 elephants) sightings; the higher number here typically indicated that a large number of elephants (more than can be counted or recalled) were seen, rather than the exact number of elephants sighted. This distinguishes areas occasionally used by elephants-including for movement-and those frequented by elephants, such as frequently used movement corridors and locations adjacent to primary habitat.

| Threats to connectivity
The impact of linear infrastructure on connectivity is increasingly recognised. To assess the potential for an expanding network of linear infrastructure to impede connectivity in our study region, we overlaid the road and railways network on our flow map.
Human-elephant conflict can be a potential barrier to connectivity (Ghoddousi et al., 2021;Goswami & Vasudev, 2017). We compared conflict locations, as recorded during our questionnaire survey, to percentile-passage values predicted from model-averaged resistances for all six θ values, as calculated above (in Section 2.6).
We also visually represented the overlay of conflict on our flow map.
Using data on reported conflict from our questionnaire surveys, we generated a heatmap based on a kernel density estimation, with a radius of 5 km; a distance of 5 km has been previously reported to be a meaningful scale for spatial patterns of human-elephant conflict (Gubbi, 2012;Guerbois et al., 2012). As conflict values were a direct representation of what we observed-rather than a prediction-we restricted our inference on conflict to the strict confines of our survey grid. For each grid cell, we extracted the mean conflict value from the heatmap described above. We overlaid this conflict heatmap on RSP-predicted passage maps.

| Effort
In total, we interviewed 1,184 people, primarily farmers, across the sampled area of 7,150 km 2 . We excluded respondents who were un-

| Resistance as a function of ecological and anthropogenic covariates
There was strong support for the effect of vegetation, measured as the Enhanced Vegetation Index, human population density, and distance to forest, on elephant use of the landscape (Table 1 and S1). Models with EVI, human population density, slope and distance to forest also performed much better than a model that just considered LULC categories, demonstrating the benefit of incorporating greater information into resistance parameterisation (Table 1).
Environmental and anthropogenic covariates provided more finescale information on resistance in our case, as compared to just using LULC categories (Figure 1b and S1).
EVI had a strong positive effect on elephant use of a spatial location such that areas with more vegetation had a higher probability of elephant use (Table 2). There was also support for a negative effect of human population density and distance to forests on elephant use of locations: elephants used areas closer to forests and with lower human population densities (Tables 1 and 2). There was some support for a negative effect of slope, wherein steeper areas tended to be less used by elephants (Table 1 and 2). Whilst the density of linear infrastructure and distance to human habitation were included in models with ∆AIC < 2, we believe these two covariates to be uninformative parameters (Arnold, 2010) and therefore expected to have little influence on movement (Fullman et al., 2017). This expectation is supported by the fact that model deviance was not substantially reduced with the addition of these covariates (Tables 1 and S1), and that the estimated coefficients for these covariates were close to 0, with 95% confidence intervals overlapping 0 (Table 2).
We estimated model-averaged conductance for each 1-km 2 grid cell in our study region using model weights (Table S1), and pixelspecific covariate values (Table S2). We calculated resistance as the inverse of conductance. The resistances thus obtained ranged from 1 to 138.41 (Figure 1b).  Table 2). The resistances of these three land uses were set to 14, 18 and 35, respectively ( Figure S1).

| Connectivity
Passage across each of our map pixels for the different θ values, as

| Threats to connectivity
Whilst we did not find support for the effect of linear infrastructure density on elephant use of the landscape, roads and railways do cut across certain important connectivity areas (Figure 4).
Recorded conflict locations overlapped more with areas of high predicted percentile passage, as compared to areas of lower percentile passage, especially under scenarios of higher θ values ( Figure 5a). The spatial overlap between connectivity and conflict is not straightforward. Clearly, certain landscapes-the more fragmented ones towards the west-face higher conflict, as compared to more stable landscapes such as the Kaziranga-Karbi Anglong Elephant Reserve (Figure 5b). Because we only assessed conflict within the spatial domain of our survey responses, we note that our inference on the interplay between conflict and connectivity is limited.
TA B L E 1 Model selection statistics for the logistic regression analysis used to assess the effect of covariates on elephant presence, including the number of parameters K, Akaike's Information Criterion (AIC), ∆AIC, AIC weights w i , and model deviance. We show models with ∆AIC < 10, along with the LULC-only and the intercept-only model Abbreviations: EVI, Enhanced Vegetation Index; population, human population density; dist forest , distance to elephant habitat (forests and grasslands; dist habitation , distance to human settlements or habitation); dens linear , density of linear infrastructure (roads and railways); LULC, land use land cover.

| D ISCUSS I ON
Connectivity is a core component of most landscape-scale conservation programmes, especially those that are planned in the light of predicted land-use change, habitat fragmentation and climate change (Doerr et al., 2011;Keeley et al., 2019). However, connectivity conservation practice on the ground has been severely limited by our (in)ability to prioritise areas that are beneficial for connectivity from a functional standpoint (Keeley et al., 2018;Vasudev, Fletcher, et al., 2015). This is particularly true for tropical landscapes, where habitat is highly fragmented and interspersed with areas of high human density. In this context, our findings provide a feasible and valuable approach to model regional connectivity for a wide-ranging species like the Asian elephant, generating information that is directly relevant to region-wide conservation planning. Whilst forested corridors are important (e.g. Vasudev et al., 2021), there is increasing recognition of the utility of the matrix in facilitating animal movement and adding redundancy-which enhances resilience of connectivity-to animal pathways (Fletcher et al., 2014;Hilty et al., 2020;McRae et al., 2008).
Our findings are particularly relevant to assessing the importance of animal movement through the non-habitat, human-modified matrix.
Connectivity modelling typically suffers from data paucity (Sawyer et al., 2011;Zeller et al., 2012), though there do exist examples of connectivity models that are well-informed by animal movement data (e.g. Fletcher et al., 2014;Panzacchi et al., 2016;Revilla & Wiegand, 2008). Sampling in the non-habitat areas, where knowledge of permeability is uncertain, is difficult, especially with respect to recording infrequent events of dispersal. Crowd-sourced data are a valuable source of information in this context, and we show that such data are amenable to connectivity modelling of detectable and charismatic species like the Asian elephant. Indeed, at the scale of large regions, and for species that have captured the interest of the general public, crowd-sourced information has provided data and scientific insights that otherwise may have been logistically difficult to obtain (Brown et al., 2018;Frigerio et al., 2018).
Our survey provides us with valuable insights into environmental and anthropogenic drivers of elephant use of locations across our region and shows that there is substantive benefit to adding such information to connectivity models. EVI emerged as a strong driver of resistance, highlighting the importance of vegetation-in forests as well as on private multiple-use or plantation lands-for landscapescale conservation. This finding was reaffirmed by the additional role of distance to forests on elephant use of the matrix. As elephants are wide-ranging and capable of traversing large distances, this effect is likely to be a manifestation of high perceived risk in the matrix.
We used our survey data to perform a form of validation of our connectivity models and discriminate between different θ values.
Ideally, we would have validated modelled passage with independent empirical data, such as elephant movement routes recorded using telemetry devices, camera trapping along potential corridors or observational follows (e.g. Panzacchi et al., 2016). Such independent validation can strengthen our results, and deviations from predicted passage may provide further insights into elephant movement behaviour; in turn, TA B L E 2 Covariate effects on elephant presence as estimated by (a) models with ∆AIC  The question of how animals move has always been one of interest to scientists (Nathan et al., 2008) as it shapes fundamental aspects of species ecology, distribution and evolution (Clobert et al., 2012;Peterson et al., 2011). In more recent times, this question has taken on significant conservation value. For both reasons, the past few decades have seen vast developments in animal movement models (Fletcher et al., 2019;McRae et al., 2008;Panzacchi et al., 2016). The RSP framework used here presents a useful and flexible approach to modelling animal movement, explicitly stating assumptions on animal perception of the landscape (Kivimäki et al., 2014). It is worthwhile to note here, that other animal movement models-ranging from least-cost (Adriaensen  (Plotnik & de Waal, 2014), store and share substantive knowledge accumulated over a long lifespan with their herd via social cues (McComb et al., 2001), and use spatial memory to determine movement paths (Polansky et al., 2015). With further information on elephant movement decisions, and their perceptual windows, we hope to be able to fine-tune our estimation of appropriate θs.
Whilst there are differences between passage maps under different scenarios of θs, there is a utility for these maps, both to identify optimal movement routes that can help prioritise important forested corridors (Figure 2a and b), as well as more diffuse routes relating to movement with partial information (Figure 2c)  the future, connectivity conservation efforts must be two-fold. One, they must secure forests that clearly forge important linkages between habitat patches. Simultaneously, efforts must be invested in the matrix to retain or even improve the permeability of these lands to movement, thus securing their role in landscape-scale conservation of endangered species.
Elephants have a complex relationship with people; whilst revered, they also are prone to enter into conflict with people (Vasudev et al., 2020). It is, thus, useful to examine the overlap between conflict and connectivity (Figure 5b), so as to tailor conflict mitigation strategies that allow elephant movement. There are clear differences across landscapes in terms of conflict incidence, with the more fragmented landscapes to the west having higher intensity of conflict. We note that our conflict data are based on direct observations and restricted to our sampled grid cells, rather than being a prediction of conflict intensity across the entire study region.
We also did not explicitly account for conflict or conflict mitigation measures in our estimation of resistance. For these reasons, we are not able to emphatically state that conflict is impeding connectivity (or not) from our current models. Nonetheless, we note that conflict, and conflict mitigation measures, do have the potential to pose a barrier to connectivity (Goswami & Vasudev, 2017;Osipova et al., 2018)-elephants were found to avoid areas of high human population density in this study, corroborating earlier research that supports the same observation (Goswami, Sridhara, et al., 2014;Goswami et al., 2021). It is arguably in this context that fine-scale prioritisation is most useful in informing local-scale actions, as well as regional policy for both connectivity conservation and conflict mitigation. Such prioritisation could form the basis for participatory conservation measures on the ground that allows for animal movement and connectivity whilst minimising conflict (e.g. . We modelled the effect of linear infrastructure density on elephant use of the landscape, but the model including this covariate received weak support, and the covariate is likely to be an uninformative parameter. We note that impacts of linear infrastructure are probably best assessed through studies at finer scales than the one we used, and through methodological approaches that assess fine-scale movement behaviour of elephants (e.g. using telemetry) rather than based on their presence alone. At such a scale and with fine-grained information, we expect that movement may be predicted to be severely restricted by roads and rails. We overlaid linear infrastructure on modelled elephant passage and found roads and rails to intersect movement zones in the study region. We note that currently, linear infrastructure in Northeast India is relatively less developed than the rest of the country (Nayak et al., 2020)  The Anthropocene has demonstrably restricted animal movement (Tucker et al., 2018). With ongoing and widespread environmental change, fragmentation of habitat, and climate change, the importance of connectivity is only emphasised within conservation plans Keeley et al., 2019). Increased hostility in people-wildlife interactions can exacerbate the loss of connectivity, whilst evidence on the impacts of linear infrastructure on both animal movement and survival is mounting.
Modelling of connectivity across large scales is clearly important and useful, not just for connectivity conservation per se, but also for habitat preservation and conflict mitigation. As we improve models of animal movement, our ability to shape holistic conservation at landscape scales, and thus achieve conservation success, is enhanced. Ramji and P. Hait for administrative assistance. We thank G. Iacona and three anonymous reviewers for constructive and thoughtful feedback that helped strengthen this paper.

CO N FLI C T O F I NTE R E S T
The authors declare that they have no conflict of interest.

PE E R R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1111/ddi.13419.

DATA AVA I L A B I L I T Y S TAT E M E N T
Data used in the paper are shared as Table S2. The Supplementary information includes code to conduct the analyses described in the paper (Table S3).