Methodological analysis of bioclimatic variable selection in species distribution modeling with application to agricultural pests (Metcalfa pruinosa and Spodoptera litura)

MaxEnt is a machine-learning-based species distribution modeling tool that is widely used to evaluate the occurrence possibility of a species. The characteristics of the algorithm requires that bioclimatic variables are appropriately selected in order for the model to predict species occurrence as precisely as possible without overfitting to the training data. In this study, we developed MaxEnt models with six different variable selection methods for two major pests, Metcalfa pruinosa and Spodoptera litura, which are differentially distributed worldwide. It was found that when evaluated by model performance indices that consider both the actual distribution and size of the prediction area, correlation-based screening, and principal component analysis-based variable reconstruction generally returned a better prediction accuracy. However, the optimal method was different for each species, indicating that the variable selection might depend on the target species and its distribution pattern, i.e., the presence data. In addition, we combined the two best MaxEnt models with a previously developed CLIMEX model, demonstrating the effectiveness of ensemble modeling in assessing species distribution. Although it was impossible to identify a single method for MaxEnt variable selection, this study provides insights on how to statistically select the bioclimatic variables and evaluate the performance between different models.

Automated summary

The study found that correlation-based screening and principal component analysis-based variable reconstruction generally returned better prediction accuracy for MaxEnt models, but the optimal method varied for each species, suggesting that variable selection depends on the target species and its distribution pattern.