Abstract:

Aims The most common method to build a predictive model of species' potential distribution is to use environmental factors, because they strongly affect species distribution. Unfortunately, most predictive models suffer from the “high dimension small sample size" problem, and cannot give satisfactory results in many cases. Support vector machine (SVM), which is based on structural risk minimization principle, has proven to be especially suitable for such data by both theory and abundant applications. Our objective was to implement a new predictive system of species' potential distribution based on the SVM method.
Methods We performed a country-scale case study using 20 Chinese endemic species of Rhododendron, employing herbarium specimen data and 11 layers of 1 km×1 km digital environmental grid data. Through expert evaluation and receiver operator characteristic (ROC) curve, we compared SVM predictions with those of a commonly used modeling method, the genetic algorithm for rule-set prediction (GARP).
Important findings All scores of SVM's prediction are higher than GARP's in expert evaluation. For the statistical analysis of ROC curve, almost all the area under the curve (AUC) determinations of SVM are larger than that of GARP. Furthermore, SVM's prediction speed is much faster than GARP's. Through our experiment, comprehensive evaluation proved that SVM is much better than GARP in terms of both performance and accuracy on the “high dimension small sample size" problem.

Key words: predictive model of species distribution, support vector machine (SVM), genetic algorithm for rule-set prediction (GARP), receiver operator characteristic (ROC) curve, Rhododendron, potential distribution