Abstract
<jats:p>Ecological niche modeling (ENM) is an invaluable tool in biodiversity science and many algorithms were developed for or incorporated into this area over the years. Back in 2007, Araújo and New outlined a framework to integrate different sources of variation into robust model projections. Almost 20 years later, surveys of real world applications show that standards are still lacking, particularly when it comes to combining different modeling strategies into a single set of predictions. In this paper we introduce a novel information theory approach to ensemble learning using the Siderastrea Atlantic Complex (SAC) as a working example. We employed nine models (BioClim, Gower, and Mahalanobis distances, GAM, GLM, MARS, ANN, MaxEnt, and SVM) to test for cross-area transferability of two species in SAC. We used Shannon's entropy to evaluate model overfitting to the data. Our method reveals that overfitting is not spatially uniform, thus, we developed a weighting strategy that reduces the influence of overfitted models on ensemble predictions on a cell-by-cell basis. Overfitting is obscured by traditional measures, such as AUC and TSS, that provided inflated estimation of real model performances across projection areas. We also introduce reliability maps, analogous to the Maps of Biogeographical Ignorance (MoBIS), combining occurrence probability and ensemble efficiency on each cell. Our approach uncovered the Yucatán peninsula as an adequate habitat for the endemic Brazilian species Siderastrea stellata and, conversely, the Northeastern Brazilian coast as a suitable area for the Caribbean species S. siderea, supporting previously published records. It also uncovered a stretch of Western African coast as highly suitable habitat for the genus, a result that was recently corroborated by citizen-science reports.</jats:p>