Community-level predictions in a megadiverse hotspot: how stacked species distribution models compare to forest inventory data?

Given the current scenario of climate change and anthropogenic impacts, spatial predictions of biodiversity are fundamental to support conservation and restoration actions. Here, we compared different stacked species distribution models (S-SDMs) to forest inventories to assess if S-SDMs capture emerging properties and geographic patterns of species richness and composition of local communities in a biodiversity hotspot. We generated SDMs for 1499 tree species sampled in 151 sites across the Atlantic Forest. We applied four model stacking approaches to reconstruct the plant communities: binary SDMs (bS-SDMs), binary SDMs cropped by minimum convex polygons (bS-SDMs-CROP), stacked SDMs constrained by the observed species richness (cS-SDMs), and minimum convex polygons of species occurrences (MCPs). We compared the stacking methods with local communities in terms of species richness, composition, community prediction metrics, and components of beta diversity - nestedness and turnover. S-SDMs captured general patterns, with bS-SDMs-CROP being the most parsimonious approach. Species composition differed between local communities and all stacking methods, in which bS-SDMs, bS-SDMs-CROP, and MCPs followed a nested pattern, whereas species turnover was most important in cS-SDMs. S-SDMs varied in terms of performance, omission and commission errors, leading to a misprediction of some vulnerable, endangered, and critically endangered species. Despite differing from forest inventory data, S-SDMs captured part of the variation from local communities, representing the potential species pool. Our results support the use of S-SDMs to endorse biodiversity synthesis and conservation planning at coarse scales and warn of potential misprediction at local scales in megadiverse regions.