*2.3. Model Simulation and Evalution*

We employed both Maxent and GARP to predict the distribution of *E. japonica* and the impact of global warming. These two models were chosen because previous similar studies demonstrated their better performance compared to other models [14,19]. Both models use artificial intelligence to evaluate the potential geographical distribution, and require location information and pseudo-absence (for Maxent) or background (GARP) data during the construction of models [9]. However, they differ in their operating principle. GARP is a machine-learning algorithm. It uses rules to determine whether a species is present within the given area and generate models [20]. GARP uses an iteration procedure, including rule selection, testing, evaluation, rejection or incorporation to select an approach based on various options (negated range rules, range rules, atomic rules, and logistic regression) and applies it into those training data for developing or evolving one rule (see Stockwell and Noble [20] for more details). However, Maxent has recently been reclassified as a version of the generalized linear model. It generates models based on the principle of maximum entropy (see Phillips et al. [29] for more details). It generates models by finding the distribution closest to uniform distribution (i.e., maximum entropy) of each environmental variable across the study area.

In our study, Maxent models were ran using version 3.3.3k [19,29,30]. The location data were randomly separated into two parts, where 75% were adopted in model training, whereas the remaining 25% were used in model testing. Recent studies showed that the default configuration is not always appropriate. Therefore, various regularized multiplier values were analyzed, finding that the default setting had the best performance [31]. The comparison of models was done by using the corrected Akaike information criterion (AICc). The best model has the smallest AICc value (for more detail, see Merow et al. [31]). The model extrapolation was improved using a bias file layer that was created to restrict those background points within species occurrence regions [2]. Repeated split samples were processed using ten replicates to measure the variation in the model; then we averaged the results. A total of 1000 iterations was selected to give the model adequate time for convergence; 1 <sup>×</sup> 10−<sup>6</sup> was selected as a convergence threshold [31]. The internal jackknife of Maxent was also adopted for testing and assessing the significance for all environmental parameters in the prediction

of *E. japonica* distribution. Maxent employs various methods that can be used to quantify how each variable contributes to the model. The present study employed permutation importance to identify the most important bioclimatic variables used in predicting the geographical distribution of a particular group of taxa. Permutation importance measures the decrease in training AUC that results from randomly permuted values of a specific variable during training of the model. A variable that requires less training AUC is more important to the model [29].

The GARP model was implemented in desktop version 1.1. GARP uses sets of rules to determine whether a species is present within the given area [20]. It uses an iteration procedure, including rule selection, testing, evaluation, rejection or incorporation to select an approach based on various options (negated range rules, range rules, atomic rules, and logistic regression) and applies it into those training data for developing or evolving one rule [32]. Location data were divided randomly in the same equal percentage for training and testing as was implemented in the Maxent. We ran 100 models with the model iteration convergence limit at 0.01 for at most 1000 iterations. Meanwhile, the "best subsets" procedure and the internal testing feature were activated to select the 10 best models [29,33]. Omission errors were included in the selection criteria (i.e., known locations predicted areas of absence); these were set to the lowest 20% of values. The default value of 50% was used for errors of commission. The two models, GARP and Maxent, were projected into datasets of the climate change scenarios after completing the iteration phase.
