Biogeographic Realism in Species Distribution Models: Pseudo-Absence Strategies and Spatial Autocorrelation in Apodemus sylvaticus

Species distribution models (SDMs) often extrapolate into regions that are ecologically and geographically unsuitable for the focal species. Using the Western Palaearctic wood mouse (Apodemus sylvaticus) as a case study and data pre-filtering, we evaluated six pseudo-absence (PA) generation strategies and four variable selection protocols aimed at improving the accuracy of predicted distribution maps while minimising spatial autocorrelation of residuals. Models were fitted using an iterative Bayesian additive regression trees algorithm (bart.step) and evaluated based on AUC, TSS, false-negative rate, Moran’s I, and Biogeographic Realism – the latter defined as the absence of predicted suitability in absolute dispersal barriers such as deserts, Arctic tundra, and northern islands.

Environmental-contrast filtering (SRE) produced only limited improvements in spatial autocorrelation and did not consistently enhance the realism of predicted distributions. Across the tested pseudo-absence strategies, several models achieved very high discriminatory performance (AUC ≈ 0.97–0.99), including EuGlobal with Range and Palaearctic-based approaches. However, these models often produced spatial artefacts in barrier regions or showed reduced coverage of the species’ peripheral range. In contrast, the Global+AC strategy (random PA selection combined with autocorrelation-guided variable selection) provided the most balanced performance, achieving high predictive accuracy (test AUC ≈ 0.98; TSS ≈ 0.86) along with enhanced Biogeographic Realism and reduced spatial artefacts. All four best models were obtained using residual autocorrelation‑guided variable selection. The reproducible workflow, implemented in custom R functions, is transferable to other Palaearctic terrestrial vertebrates.