Evaluating forest refugial models using species distribution models, model filling and inclusion: a case study with 14 Brazilian species

Abstract

Aim We aimed to assess the generality of existing models of late Quaternary biodiversity refugia in the Brazilian Atlantic forest by testing whether taxonomic identity and range descriptors influence the extent by which previously proposed models of forest (habitat) refugia successfully predict species' inferred refugial areas. Location Brazilian rain forest. Methods We compiled and filtered records of 14 animal species that belong to distantly related groups (spiders, harvestmen, scorpions, amphibians, birds, lizards and mammals) and show distinct distribution patterns within the Atlantic rain forest. Using MAXENT, we generated three distribution models for each species under different climatic scenarios (current, 6000 and 21,000 years ago). Species-specific historically stable areas (refugia) were identified through the intersection of the three models. We then measured the amount of ‘inclusion’ of species-specific refugia within published forest refugia, and quantified ‘filling’ of the biome refugia by species-specific refugia. The influence of taxonomic distance between species and range descriptors were analysed. Results Current distribution models generated for the 14 species had high accuracy (AUC > 0.9). Inclusion and filling, two uncorrelated metrics, varied among species and were not influenced by taxonomic identity. Species range characteristics influenced forest model filling only, with higher values found in widely distributed species (i.e. occurring from Northeastern to Southeastern Brazil). Main conclusions Species-specific and forest refugial areas are not necessarily the same. The power of forest refugial models to predict species-specific refugial area differs among species and may be influenced by range attributes. Species data suggests the existence of a large refugium in Southeast of Brazil, a result at odds with the currently available forest-wide models. The predictive power of forest refugial models is narrowed; we now better understand their applicability limits.