How well do SDMs calibrated at large extents predict distribution in sub-areas: A case study

Accurately capturing the realized niches of species is essential for applying species distribution models (SDMs), for example in conservation planning. Therefore, SDMs are typically calibrated over large spatial extents to avoid niche truncation but subsequently applied to distinguish suitable from unsuitable habitats within much smaller areas. However, model accuracy is commonly only assessed at the full calibration range, and whether the reduction of extent between calibration and projection areas reduces model accuracy has rarely been systematically evaluated. In this case study, we calibrated SDMs for 16 alpine plant species by relating occurrence records from across the European Alps to six topo-climatic predictors at a spatial resolution of 100 × 100 m. We then projected the species’ distributions across the Alps and compared the accuracy achieved at the extent of the Alps to the one achieved within three individual mountain landscapes. Projection accuracy for individual mountains differed strongly, ranging from projections even slightly more accurate than for the entire Alps to those much less accurate. The drop in projection accuracy between the extent of the Alps and the individual mountains increased with the dissimilarity of the niche realized by a species on a particular individual mountain as compared to the one realized at the extent of the Alps. Thus, full-extent accuracy metrics can be strongly misleading for smaller-extent applications. We recommend that such applications should be accompanied by a careful evaluation of the niche realized by species at both extents. If sufficient data are available at both extents, combining models calibrated at both scales, as recently suggested, appears a particularly promising approach.