부산대학교 도서관

Species distribution model (SDMs) is increasingly used to determine the distribution range of individual species and identify biodiversity hotspots. Of many technical issues, model over-fitting or over-parameterization is a major concern, which can lead to severe under-prediction. However, under-fitting and over-prediction may also occur if species requirements for environment are inadequately modeled. We used the collection data of stoneflies (Plecoptera, Insecta) from Illinois, USA to examine how often and severely maximum entropy (Maxent) over- or under-predicts species richness and species-occurrence frequency. A recently proposed AICc-based method (Warren and Seifert, 2011) was used for model-complexity control or regularization. Twenty-nine historically well-sampled watersheds were used to validate the predictions. The standard models, which used the default regularization (β=1), over- or under-predicted, depending on the watershed, species, and threshold used for converting suitability score into species presence–absence. The AICc-selected models (β=7–40) used 77% less parameters, but often strongly and consistently over-predicted. Three thresholds, equal training sensitivity and specificity, maximizing training sensitivity and specificity (MTSS) and minimum training presence, yielded most accurate estimates. Accordingly, we developed standard models for 41 species and identified the historically species-rich watersheds in Illinois. Our results offer new insight into the effects of regularization and choices of thresholds on Maxent performances.