Context. Vehicle collisions with wildlife can injure or kill animals, threaten human safety, and threaten the viability of rare species. This has led to a focus in road-ecology research on identifying the key predictors of ‘road-kill’ risk, with the goal of guiding management to mitigate its impact. However, because of the complex and context-dependent nature of the causes of risk exposure, modelling road-kill data in ways that yield consistent recommendations has proven challenging.
Aim. Here we used a multi-model machine-learning approach to identify the spatio-temporal predictors, such as traffic volume, road shape, surrounding vegetation and distance to human settlements, associated with road-kill risk.
Methods. We collected data on the location, identity and wildlife body size of each road mortality across four seasons along eight roads in southern Tasmania, a ‘road-kill hotspot’ of management concern. We focused on three large-bodied and frequently affected crepuscular Australian marsupial herbivore species, the rufous-bellied pademelon (Thylogale billardierii), Bennett’s wallaby (Macropus rufogriseus) and the bare-nosed wombat (Vombatus ursinus). We fit the point-location data using ‘lasso-regularisation’ of a logistic generalised linear model (LL-GLM) and out-of-bag optimisation of a decision-tree-based ‘random forests’ (RF) algorithm for optimised predictions.
Results. The RF model, with high-level feature interactions, yielded superior out-of-sample prediction results to the linear additive model, with a RF classification accuracy of 84.8% for the 871 road-kill observations and a true skill statistic of 0.708, compared with 61.2% and 0.205 for the LL-GLM. The lasso rejected road visibility and human density, ranking roadside vegetation type and presence of barrier fencing as the most influential predictors of road-kill locality.
Conclusions. Forested areas with no roadside barrier fence along curved sections of road posed the highest risk to animals. Seasonally, the frequency of wildlife–vehicle collisions increased notably for females during oestrus, when they were more dispersive and so had a higher encounter rate with roads.
Implications. These findings illustrate the value of using a combination of attributive and predictive modelling using machine learning to rank and interpret a complexity of possible predictors of road-kill risk, as well as offering a guide to practical management interventions that can mitigate road-related hazards.
