Selection of environmental predictors for species distribution modeling in Maxent

Prior to conducting the modeling of the potential distribution of a species, it is advised to make a pre-selection of covariables because redundancy or irrelevant variables may induce errors in most modeling systems. In this study, we propose an automated method for a priori selection of covariables used in modeling. We used five typical species of the Mexican flora (Catopheria chiapensis, Liquidambar styraciflua, Quercus martinezii, Telanthopora grandifolia and Viburnum acutifolium) and 56 environmental covariables. Presence-absence matrices were generated for each species and were analyzed using logistic regression, and the resulting model of each species was evaluated via a bootstrap resampling. We modeled the distribution of five species using maximum entropy and employed three sets of environmental covariables. The precision of the models generated was evaluated with the confidence intervals for each receiver operating characteristic (ROC) curve. The confidence intervals of the resulting ROC curves showed no significant difference between (P < 0.05) the three predictive models generated; nevertheless, the most parsimonious model was obtained with the proposed method.