2.2.3. Species distribution modeling

The potential distribution of each species was predicted through an ensemble modeling approach performed with the “biomod2” package (BIOdiversity MODelling‐Biomod2) version 3.3‐7.1 (Thuiller et al., 2020), implemented in R version 3.6.0 (R Development Core Team, 2020). Biomod2 is a computer platform for ensemble forecasting of species distributions, which allows maximizing the predictive accuracy of SDMs by combining different modeling methods (Araújo & New, 2007; Hao et al., 2019). We fitted the SDMs using an ensemble of five different modeling algorithms: two regression techniques, namely generalized linear models (GLMs) and multivariate adaptive regression splines (MARS); and three machine learning methods: generalized boosted models (GBMs), random forest (RF), and MaxEnt (detailed description of the model techniques in Elith et al., 2011; Franklin, 2010; Phillips et al., 2006). We selected these algorithms based on their superior performance during an exploratory modeling exercise.

All the selected modeling techniques require records of presence and absence, except MaxEnt, which is a presence background modeling tool. As our data are presence‐only, we generated three different sets of pseudo‐absences, where one‐third of the available background modeling cells were randomly sampled and used as pseudo‐absences, following the recommendations of Elith et al. (2006). The models were processed with the default settings for each modeling technique and the following options: equal weight of background absences and occurrences; 10,000 maximum interactions; 100 replicate runs for each set of “pseudo‐absences” and for each model technique; the occurrence data were randomly split into 70% training data and 30% test data to evaluate the predictive performance of the models; and 10 permutations to estimate variable importance. For each species were produced 1,500 individual models. Following the recommendations of Zurell et al. (2020), we present the ODMAP protocol with details of the modeling process in Data S1.