An artificial intelligent framework for prediction of wildlife vehicle collision hotspots based on geographic information systems and multispectral imagery

Abstract Wildlife-vehicle collision - WVC is a phenomenon that arises from the fragmentation of ecosystems by roads, limiting the mobility of individuals and putting at risk the stability of populations by increasing mortality. Colombia is not unaware of the problem of the WVC, evidenced in different scientific publications that describe the WVC in the roads of the country. Although the rise of artificial intelligence has significant advances in the prediction of spatial phenomena in recent years, it has not yet been sufficiently explored by Road Ecology. For this reason, this research aimed to develop a methodology to predict the sites of accumulation of WVC in eastern Antioquia, Colombia, based on artificial intelligence algorithms, geographic information systems - GIS, and multispectral image processing. During the development of this research, it was identified that the features most related to the WVC in the study area are: Distance to Forest, Distance to Biological Corridor, Ground Resistance to Movement, Cost of Movement, the bands of the Landsat 8 satellite: 9, 10, 11 and the normalized burning index (NBRI). Different machine learning algorithms were compared (k-nearest neighbours, support vector machines (SVM), random forests (RF), and artificial neural networks). SMOTE and ADASYN balancing techniques were applied. The results allowed to identify that the RF algorithm with ADASYN yielded the best performance when subjected to spatial-wise cross-validation (AUC-ROC 0.78 ± 0.12), surpassing the results of current state-of-the-art. Finally, the methodology was validated through a transfer learning experiment, training the RF-ADASYN algorithm with three zones of the eastern Antioquia region and validating on a different section (AUC-ROC = 0.87 ± 0.09), retraining the initial model with 5% of data from the validation database.