Potential species distribution modeling and the use of principal component analysis as predictor variables

Abstract

Prior to modeling the potential distribution of a species it is recommended to carry out analyses to reduce errors in the model, especially those caused by the spatial autocorrelation of presence data or the multi-collinearity of the environmental predictors used. This paper proposes statistical methods to solve drawbacks frequently disregarded when such models are built. We use spatial records of 3 species characteristic of the Mexican humid mountain forest and 2 sets of original variables. The selection of presence-only records with no autocorrelation was made by applying both randomness and pattern analyses. Through principal component analysis (PCA) the 2 sets of original variables were transformed into 4 different sets to produce the species distribution models with the modeling application in Maxent. Model precision was higher than 90% applying a binomial test and was always higher than 0.9 with the area under the curve (AUC) and with the partial receiver operating characteristic (ROC). The results show that the records selected with the randomness method proposed here and the use of the PCA to select the environmental predictors generated more parsimonious predictive models, with a precision higher than 95%, and in addition, the response variables show no spatial autocorrelation.