Aim: The non-stationarity in habitat selection of expanding populations poses a significant challenge for spatial forecasting. Focusing on the grey wolf (Canis lupus) natural recolonization of Germany, we compared the performance of different distribution modelling approaches for predicting habitat suitability in unoccupied areas. Furthermore, we analysed whether grey wolf showed non-stationarity in habitat selection in newly colonized areas, which will impact the predictions for potential habitat.

Location: Germany

Methods: Using telemetry data as presence points, we compared the predictive performance of five modelling approaches based on combinations of distribution modelling algorithms –GLMM, MaxEnt, and ensemble modelling– and two background point selection strategies. We used a homogeneous Poisson point process to draw background points from either the minimum convex polygons derived from telemetry or the whole area known to be occupied by wolves. Models were fit to the data of the first years and validated against independent data representing the expansion of the species. The best-performing approach was then used to further investigate non-stationarity in the species’ response in spatiotemporal restricted datasets that represented different colonization steps.

Results: Whilst all approaches performed similarly when evaluated against a subset of the data used to fit the models, the ensemble model based on integrated data performed best when predicting range expansion. Models for subsequent colonization steps differed substantially from the global model, highlighting the non-stationarity of wolf habitat selection towards human disturbance during the colonization process.

Main conclusions: While telemetry-only data overfitted the models, using all available datasets increased the reliability of the range expansion forecasts. The non-stationarity in habitat selection pointed to wolves settling in the best areas first, and filling in nearby lower-quality habitat as the population increases. Our results caution against spatial extrapolation and space-for-time substitutions in habitat models, at least with expanding species.

Our wolf data comprised two complementary datasets. Our first dataset consisted of GPS telemetry locations of 20 collared resident wolves from 2009 to 2018. This data is provided at a 50km resolution due to species conservation concerns. Our final telemetry dataset consisted of 3,841 locations from 21 home ranges (183 ± 104 locations per home range). Our second dataset consisted of centroids of known wolf territories monitored annually since 2000 (www.dbb-wolf.de/Wolfsvorkommen/territorien/karte-der-territorien). The centroids of the territories were assessed after the end of the monitoring year, as the central point of all activity signs (scats; camera trap images; telemetry data, when available; opportunistic sightings; hunt remains) that were assigned to the same territory. Territory areas for our study were delineated with a radius of 8 km around their centroids, resulting in territory sizes of approx. 200 km².

Datasets are specially optimised for R, but they can open with any software that reads .csv files.