Forecasting seasonal peaks in roadkill patterns for improving road management

For several species, roadkill is not spatially aggregated on hotspots, having instead a more diffuse pattern along the roads. For such species, management measures such as road passages may be insufficient for effective mitigation, since a large part of the road crossings is likely to occur outside the influence of those structures. One complementary approach could be to implement temporary mitigation actions, such as traffic calming. This requires understanding when roadkill peaks may occur. We tested the feasibility of predicting seasonal peaks of roadkill using data from a 3-year systematic monitoring (78 surveys over ca. 960 km of roads) from eight non-flying vertebrate species from Mato Grosso do Sul, Brazil, with different body size and life history traits (ca. 6400 records from focal species). We modelled the time-series of the roadkill of these species at large scale (state level) using generalized additive mixed models (GAMMs). We used the data of the first 2 years as training datasets, and the information from the third year of surveys as testing datasets to evaluate the prediction performance of models. Overall, the models of species feed with a higher number of records were able to follow reasonably well the variations of roadkill over time, although they were not able to correctly predict the number of collisions. For species with fewer observations, the models presented a poorer goodness-of-fit and prediction ability. Our results suggest that, at least for those species with higher roadkill rates, it can be possible to forecast periods of higher probability of occurring hot-moments of mortality. Such models can provide valuable information to implement seasonal management actions.