Adrian Meyer‡§, Stefan Pretali‡, Théo Reibel‡, Manisha Bhardwaj|, Mathias Kneubühler§, Denis Jordan‡Corresponding author: Adrian Meyer ( adrian.meyer@fhnw.ch ) © Adrian Meyer, Stefan Pretali, Théo Reibel, Manisha Bhardwaj, Mathias Kneubühler, Denis Jordan. This is an open access preprint distributed under the terms of the Creative Commons Attribution License (CC BY 4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Citation:
Meyer A, Pretali S, Reibel T, Bhardwaj M, Kneubühler M, Jordan D (2026) Multimodal Machine Learning to Predict Wild Boar Collision Risk on the French Railway Network. ARPHA Preprints. https://doi.org/10.3897/arphapreprints.e209840 |  |
Wild boar (Sus scrofa) are the most prevalent species involved in wildlife-vehicle collision (WVC) incidents on the French railway network, resulting in direct mortality, operational delays and safety risks. More than 1,400 wild boar-related WVCs were recorded in 2025 by SNCF Réseau, the national railway infrastructure operator. Strong recent increases are observed across all regions of metropolitan France, often following exponential growth. While railway ecology remains understudied compared to road-based WVC research, strong spatial clustering of observations and winter-dominated seasonal variation amplify the need for more targeted mitigation measures and data-driven decision-making.
We therefore present a spatially and seasonally explicit, national-scale risk prioritization workflow for wild boar WVC projection on the French railway network. Our LightGBM machine learning regression models are based on a full decade of SNCF WVC records utilized as raw and smoothed target risk scores, as well as a comprehensive dataset of environmental covariates: At square kilometer resolution, we compile rail infrastructure and operation attributes, ecologically relevant land cover and habitat suitability maps, terrain and passability variables, as well as municipality-level hunting indices. Winter and summer risk are modelled as separate endpoints to account for seasonal shifts in animal behavior, hunting regime, and resource availability. Broad-scale 12 km smoothed models yield highest test-set R2 values of 0.73 for summer and 0.81 for winter compared to other smoothing distances. By contrast, finer scale 5 km models retained higher predictive power for individual local hotspots. Spatial multiscale leave-out stress tests confirm generalization capability.
We demonstrate a novel seasonal machine-learning method to derive risk surfaces and support railway biodiversity monitoring and wildlife risk management. The models provide a validated and transferable basis for producing risk maps, identifying ranked intervention hotspots, and projecting regional incident rates into the future.