Predicting wildlife road-crossing probability from roadkill data using occupancy-detection models.

Wildlife-vehicle collisions (WVC) represent a major threat for wildlife and understanding how WVC spatial patterns relate to surrounding land cover can provide valuable information for deciding where to implement mitigation measures. However, these relations may be heavily biased as many casualties are undetected in roadkill surveys, e.g. due to scavenger activity, which may ultimately jeopardize conservation actions. We suggest using occupancy models to overcome imperfect detection issues, assuming that 'occupancy' represents the preference for crossing the road in a given site, i.e. is a proxy for the roadkill risk; and that the 'detectability' is the joint probability of an animal being hit in the crossing site and its carcass being detected afterwards. Our main objective was to assess the roadkill risk along roads while accounting for imperfect detection issues and relate it to land cover information. We conducted roadkill surveys over 114 km in nine different roads, biweekly, for five years (total of 484 surveys), and developed a Bayesian hierarchical occupancy model to assess the roadkill risk for the six most road-killed taxa for each road section and season (WET and DRY). Overall, we estimated a higher roadkill risk in road sections surrounded by agriculture, open habitats; and a higher detectability within the 4-lane road sections. Our modeling framework has a great potential to overcome the limitations related to imperfect detection when assessing the roadkill risk and may become an important tool to predict which road sections have a higher mortality risk.