An investigation modeling risk of wildlife-vehicle collisions in Montana, USA

Abstract

Road ecologists and transportation engineers have been exploring new methods to adapt to the environmental and motorist safety concerns involving wildlife-vehicle collisions. There are over one-million crashes with large-bodied wildlife every year in the U.S. that result in substantial property damage and personal injuries. Recent studies modeling these collisions identify where they cluster, and the landscape, road, and driver characteristics that influence the likelihood of a collision along short road corridors and small geographic regions. This research expands on current knowledge and attempts to model the risk of wildlife-vehicle collisions on a large geographic scale. This research investigates different analysis methods and creates predictive models that will estimate the risk of a wildlife-vehicle collision as drivers travel across multiple ecosystems. Different analysis units were created to extract two similar datasets that are modeled against two different response variables -- reported collisions and roadkill locations. Regularization is used to help with feature selection. Negative binomial regression models are built to predict risk. Random forest machine learning helps better understand the percent of variance explained by the variables in each model. A range of statistical measurements were taken to compare the non-nested models. The best performing model is applied to the seasonal division of data. Yearlong and seasonal risk is mapped onto the road network and color-coded to show the differences in risk on Montana's road network. The maps capture the changes in risk throughout the year, they generally match where wildlife-vehicle collisions actually happen, and even coincides with published work on the locations of collision hotspots in Montana. This research is the basis for future complex real-time risk-mapping models that can be integrated into smart technology and developed into on-board driver alert systems. With the advancements of autonomous vehicle, it is possible to incorporate real-time driving data into models that will analyze wildlife-vehicle collision risk based on vehicle location, season, time of day and driving habits. This can increase driver safety by informing them when they are traveling in areas where wildlife-vehicle collisions are more likely to happen, and can be especially helpful while driving on unfamiliar roads.