Abstract
Aim Ecological niche modeling (ENM) is an approach used to estimate species‘ presence given its environmental preferences. Model complexity in ENMs has increasingly gained relevance in the last years. In particular, in Maxent algorithm is captured using the Akaike Information Criteria (AIC) based on the number of parameters and likelihoods of continuous raw outputs. However, it is not clear whether best-selected models using AIC are the models with the highest classification rate of correct presences and absences. Here, we test for a link between model complexity and accuracy of geographical predictions of Maxent models.
Innovation We created a set of virtual species and generate true geographical predictions for each one. We build a set of Maxent models using presence data from each virtual species with different regularization and features schemes. We compared AICc values for each model with the scores of standard validation metrics (e.g., Kappa, TSS) and with the number of pixels correctly predicted as presences, absences or both.
Main Conclusions We found that binary predictions (i.e., presence-absence maps) selected as best models for AIC tend to predict incorrectly sites as presences and absences using independent datasets. We suggest that information criteria as AIC should be avoided when users are interested in binary predictions. Future applications that capture model complexity in ENM applications should be evaluated using standard validation metrics.
