Zooming In: Assessing the Transferability of SDMs From Large to Small Spatial Extents

Species Distribution Models (SDMs) represent a common method for predicting biodiversity responses to environmental changes. SDMs are typically established for and applied to large areas, while little is known about the transferability of large-scale SDMs to smaller spatial extents. This study aims to fill this gap by fitting country-level ensemble SDMs for the seven native reptile species occurring in the Netherlands, and testing their predictive ability against independent occurrence data in 10 municipalities. For each of the species, we trained an ensemble SDM using bioclimatic variables, land cover fractions and soil type as predictors. We then evaluated model performance in 10 municipalities that were not included in model training. Despite performing well at national extent, with a mean cross-validated True Skill Statistic (TSS) value of 0.85 (range of 0.76-0.94) across the species, the ensemble SDMs generally yielded poor to moderate performance scores for the municipalities. Averaged per species, TSS values ranged from 0.05 to 0.57, while mean TSS values per municipality ranged from -0.07 to 0.61. The general decrease in model performance from national to local extent was mainly due to a decline in model specificity (i.e., the ability to predict absence). A two-way ANOVA showed model performance varied significantly among both municipalities and species (<i>p</i> < 0.001). Our results suggest that while ensemble SDMs fitted across large spatial extents can provide a rough indication of the potential distribution of a species, caution should be applied when informing local conservation decisions based on broad-scale SDM outcomes.