*3.2. Model Selection and Validation*

The predictive capacities of the models for the evaluated species, assessed through AUC, Kappa, and TSS, and their standard deviation are presented in Figure 3.

**Figure 3.** S Plot of model performance (AUC, Kappa, TSS) for 100 repetitions of each technique on Uruguay including Artificial Neural Networks (ANN), Boosted Regression Trees (GBM), Classification and Regression Trees (CTA), Flexible Discriminant Analysis (FDA), Generalize Additive Models (GAM), Generalized Lineal Models (GLM), Multivariate Adaptive Regression Splines (MARS), Maximum Entropy (MaxEnt), Random Forests (RF) and Surface Range Envelop (SRE) for *Eucalyptus dunnii* (**A**) and *E. grandis* (**B**) in Uruguay.

Overall, RF presented the most accurate result, higher values of AUC, Kappa and TSS and reduce standard deviation of the values of the statistics, while ANN presented similar accurate values than RF, though with large variation on the statistic, with results that varies in AUC from 0.5 to almost 1. The lower accuracy values are given by CART, surprisingly MaxEnt and SRE. The accuracy values of these three models presented the highest variability and reduce values on accuracy. BRT and FDA presented intermediate high accuracy scores, while linear model GLM and GAM, and MARS presented intermediate low accuracy results. In general, the accuracy results between both species are similar, though the accuracy of BRT for *E. grandis* were on the range of ANN and RF.

The predicted values obtained using ensemble models were higher than those given by individual models, for both species, except with the K index with the RF model (Table 3 and Figure 3). For *E. dunnii*, the average values of the AUC, Kappa, and TSS index were 0.98, 0.88, and 0.77, respectively. For *E. grandis*, their values were 0.97, 0.86, and 0.80, respectively. With both species there was a similar precision with the ensemble model, relative to individual models, whereas for *E. grandis* the precision was increased by using the ensemble model. Ensemble model overcame in accuracy single model predictions.


**Table 3.** Statistics of the fitted values obtained with the ensemble model for the prediction of habitat for *Eucalyptus dunnii* (top) and *E. grandis* (bottom) in Uruguay.
