Evaluation of species distribution models for estimating animal dark diversity

Aim Applying wide and effective sampling of animal communities is rarely possible due to the associated costs and the use of techniques that are not always efficient. Thus, many areas have a faunistic hidden diversity we denote Animal Dark Diversity (ADD), defined as the diversity that is present but not yet detected plus the diversity defined by Pärtel et al. (2011) that is not (yet) present despite the area’s favourable habitat conditions. We evaluated different species distribution model types (SDM techniques) on the basis of three requirements for ADD estimate reliability: 1) estimated spatial patterns of ADD do not differ significantly from other SDM techniques; 2) good predictive performances; and 3) low overfitting. Location Iberian Peninsula. Taxon Chiroptera and Noctuoidea (Lepidoptera) Methods We used distribution data for 25 species of bats and 352 species of moths. We evaluated eleven SDM techniques using biomod2 package implemented in the R software environment. We fitted the various SDM techniques to the data for each species and compared the resulting ADD estimates for the two animal groups under three threshold types. Results The results demonstrated that estimated ADD spatial patterns vary significantly between SDM techniques and depend on the threshold type. They also showed that SDM techniques with overfitting tend to generate smaller ADD sizes, thus reducing the possible species presence estimates. Among the SDMs studied, the ensemble models delivered ADD geographic patterns more like the other techniques while also presenting a high predictive performance for both faunal groups. However, the Ensemble Model Committee Average (ECA) performed much better on the sensitivity metric than all other techniques under any of the thresholds tested. In addition, ECA stood out clearly from the other ensemble model techniques in displaying low-medium overfitting. Main conclusions SDM techniques should no differ among each other in their ADD estimations, have good predictive performances and exhibit low overfitting. Furthermore, to reduce estimate uncertainty it is suggested that the threshold type be one that transforms high values of presences probabilities into binary information and furthermore that the SDM technique have a sensitivity bias, as otherwise the estimates will perform better for species absence in cases where it is not in fact known whether a species is truly absent.