Species distribution models built with local species data perform better for current time, but suffer from niche truncation

To cope with climate change-induced alterations, forest ecosystems' conservation and restoration require models that are both capable to incorporate current local-scale dynamics but also to anticipate future changes. These requirements may be fulfilled by robust assessments of response (i.e., species data such as forest inventories) and predictor (e.g., climate) variables. The aim of this study is to predict current and future probability of occurrence for 22 tree species comparing inventory and climate data at different spatial scales and test for model performance, reliability, and niche truncation. We built species distribution models (SDMs) for 22 tree species of Piedmont, an Alpine administrative region of north-western Italy. We compared (i) a fine-scale model calibrated with a local forest inventory with a 250-m spatial resolution at the extent of Piedmont and a regional climate model calibrated on the Italian extent versus (ii) coarse-scale model calibrated with a pan-European forest inventory (EU-Forest) at 1-km resolution and a global climate dataset (CHELSA v1.2). Moreover, (iii) we developed a data pooling method by combining the species data and using CHELSA. We evaluated models using spatial-block cross-validation and external validation through several metrics. We predicted the probability of occurrence for current and future under RCP4.5 and RCP8.5 climate scenarios. Models built with local species data performed better for the future than those incorporating broad species data and their current predictions reflected the realized distribution of species but they suffered from niche truncation while extrapolated to the future. Indeed, models calibrated at the local scale predicted greater magnitude of changes for future scenarios compared to coarse-scale models. Integrating species data at different extents and resolutions is a valid approach when both are available.