Multiscale spatially explicit modelling of livestock depredation by reintroduced tiger ( Panthera tigris ) to predict conflict risk probability

2 Understanding the causal factors associated with human/livestock-large carnivore conflict and distribution of 3 conflict risk is key to designing effective preventative and mitigation strategies. Spatial modelling of human-4 carnivore conflict has recently gained traction, and predictive maps have become a great tool to understand the 5 distribution of present and future conflict risk. However, very few such studies consider scale and use 6 appropriate spatial modelling tools. We aimed to understand the ecological correlates of human-tiger ( Panthera 7 tigris ) conflict, predict livestock predation risk by reintroduced tigers in Panna Tiger Reserve, Central India and 8 understand the prey-predator dynamics behind the conflict. We modelled livestock kill as a function of various 9 tiger relevant ecological variables at multiple scales employing spatially explicit statistical tools. As a first step, 10 we used geostatistical modelling to create raster layers of covariates (prey, cover, human activities), following 11 which we did univariate scaling. We then modelled livestock loss by tiger using a geoadditive model. 12 Employing this model, we predicted and mapped conflict risk probabilities within our study site. It was found 13 that prey and shrub cover both selected at a fine scale, were key ecological determinants of human-tiger conflict. 14 Prey showed an inverse relationship while shrub showed non-linear relationship with livestock predation. Which 15 lead us to conclude that in habitats where optimum ambush cover is available but prey presence is low at fine-16

the intensity of HTC within the reserve. But the compensation data was not geotagged, hence we also obtained 153 the livestock kill data (that has GPS locations), which is collected by tiger monitoring teams in the reserve, for predation risk probability modelling, we treated livestock kill data as presence and generated equal number of 158 random pseudo-absence points using the "create random points" tool in ArcGIS 10.4 (ESRI 2016a) within the 159 study site (discarding the absence points falling within a buffer of 42m around each presence point). We used

NDWI uses green and near infrared bands to show presence
Additionally, we obtained drainage and water source data from the forest department and created Euclidean 190 distance raster using "Euclidean Distance" tool in the Spatial Analyst toolbox in ArcGIS 10.4 (ESRI 2016a). We 191 also created Euclidean distance rasters for Ken River and its tributaries, and water sources tagged perennial. 197 TRI = SquareRoot (Abs((Square("3x3max")-Square ("3x3min")))).

261
We checked spatial autocorrelation in these variables using Moran"s I statistic using ArcGIS 10.4, and it was

271
Since the response variable, in this case, is binary, Bernoulli distribution is assumed, and logit link used

300
The effective degree of freedom (edf) is higher than 3 for most of the smooth terms, indicating that the 301 wiggliness is high and relationships are nonlinear (Table 1). Even more is revealed by examining the partial 302 effect plots of smooth terms, also called rug plots. A partial effect plot shows the effect of an explanatory 303 variable on the response variable after accounting for the effects of all the other variables included in the model.

304
Upon examining the partial effect plot for all prey encounter rate, we found that it has an inverse relationship 305 with log odds of livestock kill i.e., the odds of livestock kill by tiger are higher when prey is low (Figure 3 a). In 306 case of shrub abundance, we observed a unique trend, log odds of livestock kill increase with shrub abundance 307 but only till it reaches a certain mark, after which increase in shrub abundance seems to reduce the odds of 308 livestock kill (Figure 3 b). NDVI, human encounter rate and elevation, as also indicated by their chi-square p 309 values, do not seem to have a significant relationship with the odds of livestock kill (Figure 3 c, d, e).
the model accuracy was calculated to be 0.65 (Table 2), and AUC was found to be 0.70, indicating that the 312 model had fair amount of prediction capability.