Evaluating the Efficacy of Wrappers in Predicting Species Distribution

Abstract

Species distribution models (SDMs), also called habitat suitability models, are generally utilized in ecology for various tasks such as habitat evaluation, species preservation, and developing evolutionary perspectives by predicting species distributions. These models provide important perspectives on both the ecological and evolutionary aspects of the field. The goal of feature selection (FS) is to identify key interrelated features or remove irrelevant and redundant ones, thus minimizing model complexity, lowering storage requirements, improving the interpretability of the final model, and sometimes boosting performance. This study seeks to identify the most efficient feature selection technique for predicting the distribution of three bird species by comparing six different wrapper methods, evaluated across four classifiers: Random Forest (RF), LightGBM, Decision Tree (DT), and Support Vector Machine (SVM). The empirical analysis incorporates various methods, including 5-fold cross-validation, the Scott-Knott statistical test, and the Borda Count voting method. Additionally, three performance metrics were utilized: F1-score, Kappa, and accuracy. Experiments revealed that when comparing the wrappers, Sequential Feature Selector (SFS) gave satisfactory results when using RF and DT except for LGBM and SVM where Permutation Importance (PI) and Shapley Additive Explanations (Shap) were the best, respectively. Furthermore, Precipitation of wettest quarter, Temperature seasonality, Isothermally, and Precipitation of coldest quarter were the most relevant features for predicting species distribution.