Predicting Plant Species Distribution Across an Alpine

5 Predictive models constitute an important tool in ecology. Using presence/absence data of 6 15 plant species of an alpine rangeland in northern Spain, and a set of 14 topographical and 7 geomorphological descriptors of relatively easy acquisition, we examined and compared the 8 performance of five state-of-the-art methods used in ecological modeling: Multiple Logistic 9 Regression (MLR), Artificial Neural Networks (ANN), Support Vector Machines (SVM), 10 Classification and Regression Trees (CART) and Multivariate Adaptive Regression Splines 11 (MARS). Validation of the models was carried-out computing the Area Under the ROC 12 Curve (AUC) using leave-one-out cross validation and the resolution and reliability diagrams 13 of the resulting probabilistic predictions. We also analyzed the binary presence/absence de14 terministic predictions obtained by setting two different probability thresholds: the species 15 prevalence and a ROC-optimized value, and we computed the corresponding confusion ma16 trices to calculate sensitivity, specificity, Cohen’s kappa and the True Skill Statistic (TSS). 17 The overall result of this comparison shows that the performance of each technique varies 18 depending on the target species; in general, CART exhibited a poor performance and MLR 19 was competitive with the more sophisticated ANN, MARS and SVM methods. The best 20 predictive resolution was obtained in most cases by ANN followed by SVM and CART 21 models; on the other hand, MLR and MARS were generally the best calibrated. We also 22 present an ecological interpretation of results, with emphasis in the possible ways of im23 proving our models. Most of the target species were accurately predicted evidencing that 24 geomorphological and topographical variables are suitable descriptors at the scale of analysis. 25 26