Evaluación de la configuración de tres algoritmos para realizar modelos de distribución potencial de especies forestales

Potential distribution models are a useful tool to identify optimal environmental conditions for an organism to prevail. The objective of the study was to evaluate the parameter configuration of the Maxent, Random Forest and Generalized Additive Models (GAM) algorithms through the generation of distribution models of six forest species from Mexico. With presence records of six forest species, potential distribution models were adjusted with three algorithms, two configurations were used in their parameters (tuned and default), the models were evaluated through the area under the curve, to compare the tuned and default configuration. violin plots, plots of predicted fitness values with both configurations, and a fuzzy global matching analysis were performed. Sets of four, five, and six variables improved predictions. The best values in the regularization multiplier ranged between 0.1 and 0.4, the feature classes that best describe the potential distribution of the six species were linear, quadratic and product. Random Forest showed that with 750 and 1 000 trees and two variables in each division the fit of the models for the six species does not improve. The best smoothing values for the GAM algorithm ranged from 0.0001 for P. pseudostrobus to 1.5 for P. durangensis, however, no differences were found between fitting models with fine-tuned and default settings. The algorithms obtained good performance, however, the effect of parameter tuned on the predictive capacity was marginal and varied depending on the algorithm.