Model training and model selection
Maxent v3.4.3 (Phillips et al. 2017) was used through ENMeval v2 (Kass
et al. 2021) to train the models for this study. Before model
comparisons, I first selected the optimal model complexity for each
approach (STA, T01, T05, T10) by running and testing 500 combinations of
feature classes (linear, linear-quadratic, hinge,
linear-quadratic-hinge, linear-quadratic-hinge-product) and
regularization multipliers (0.1 to 5, steps of 0.1). The selection of
the best model in each approach was based on the lowest corrected Akaike
Information Criterion (AICc), as it provides a measure to balance model
complexity and predictability (Warren and Seifert 2011). To account for
the elongated distribution of the occurrences along the Sierra Madre
Oriental, occurrence and background data were divided into four spatial
bins for cross-validation using a latitudinal block partition.
Validation metrics, specifically the average Area under the Curve (AUC),
the continuous Boyce Index (CBI), and the 10thpercentile of training omission rate (OR), were used to quantify the
performance of the selected model in each approach. The AUC values range
from 0 to 1, with values closer to 1 indicating better model
performance. Similarly, the CBI values range from -1 to 1, with values
closer to 1 indicating a better model performance (Hirzel et al. 2006).
On the other hand, the OR ranges from 0 to 1, and a lower value
indicates a better prediction of the occurrences.