Species distribution modeling projections

Species distribution models (SDMs) were constructed using the ‘biomod2’ package in R. We first ran a rectilinear surface range envelope model, and then, from outside the area predicted as suitable habitat, we picked 20 independent sets of 100 pseudo-absence points, each of which were combined with the same 91 presence records. Four modeling algorithms were run: artificial neural networks, generalized boosted models or boosted regression trees, random forest, and maximum entropy. We used 5 cross-validation runs per algorithm, for a total of 400 runs (4 algorithms x 5 cross-validations x 20 datasets), with 5,000 iterations per run. To assess model performance, 75% of the data were used for training, with 25% set aside as "out-of-bag" test data. To maximize the accuracy of presence/absence classification, we used the True Skill Statistic (TSS = sum of sensitivity and specificity – 1), where SDMs with mean TSS above 0.2 were retained. We then used the ensemble framework to obtain a weighted average of all SDMs, where SDMs were weighted according to TSS values. Projections: present-day SDMs were based on mean climatological data spanning 1960–1990, and historical distributions were modeled for the Mid-Holocene (~6,000 years ago), the Last Glacial Maximum (~22,000 years ago), and the Last Inter-glacial (~120,000–140,000 years ago).