Prediction of areas with high risk of roadkill wildlife applying maximum entropy approach and environmental features: East Antioquia, Colombia

Abstract

Linear infrastructures such as roads are known to cause adverse effects on the surrounding ecosystems. Wildlife–vehicle collisions (WVC) are considered to be one of the main causes of biodiversity loss. Several studies have demonstrated that WVC occurs on Colombian roads. However, studies have focused on a body count, the most affected species, and places with high mortality rates. We aim to propose a methodology for predicting WVC risk in the east of Antioquia, Colombia employing a machine learning approach to identify road segments with a high risk of WVC. Additionally, we present a novel validation technique for the "MaxEnt" approach. During this investigation, 499 reports were collected through road surveys between 2015 and 2016. We identified 160 road segments with high mortality rates with a 2D Hotspots analysis. 15 environmental descriptors were collected for each road segment. Validation of the predictive capabilities of the algorithm was performed using the area under the Receiver Operating Characteristic curve (AUC-ROC). The model achieved a good predictive ability (AUC>0.77). The response curves evidenced that features like distance to forest, land cover, resistance, and land use increase the probability of WVC, specifically, collision risk was higher in zones with high resistance values, crops, and pastures. This methodology has the potential to become an important tool for the prioritization of resources to mitigate WVC.