Risk Prediction of Wildlife-vehicle Collisions Comparing Machine Learning Methods and Data Use

Risk prediction of wildlife-vehicle collisions is crucial for reducing the abundant number of road accidents with wildlife worldwide. In this study, three different machine learning approaches - Gaussian Naive Bayes, Stochastic Gradient Descent, Random Forest - relying on recorded accident data, environmental, temporal and infrastructural parameters, are used to test the impurity of these parameters, and to predict the risk of wildlife-vehicle collisions. We use a dataset of one million police recorded accidents for southeastern Bavaria. Based on different approaches for feature engineering, we show the importance of data pre-processing and cleaning. With Random Forest, we receive a prediction accuracy of accidents of 86.7%. Additionally, the transferability of the three approaches is discussed, and possible ways of dynamic warnings via smartphone are presented.