Comparative Evaluation of Machine Learning Models for Predicting Cimbex Quadrimaculata Populatio

quadrimaculata observations in Diyarbakır (Eğil) and Elazığ (Keban), Türkiye, spanning 2020–2022.

• Objective framed as a comparative evaluation of AI-driven machine learning approaches for predicting population density of Cimbex quadrimaculata using ecological data with environmental predictors.
• Data source comprises meteorological and vegetation variables collected for C.
• A three-pronged modeling design was employed on the same dataset: binary classification, multiclass classification, and regression, enabling cross-formulation comparison of performance, generalizability, and explainability.

• The study defines distinct target formulations corresponding to the three modeling tasks, enabling assessment under stratified cross-validation and nested cross-validation schemes.
• For classification tasks, model performance was assessed with accuracy, F1 score, and AUC under a stratified 10-fold cross-validation approach.
• For regression tasks, evaluation utilized a nested cross-validation framework with metrics including R-squared, RMSE, and MAE.
• Ensemble-based boosting algorithms were central to the assessment, including Gradient Boosting, XGBoost, and LightGBM, chosen for their capacity to model nonlinearities and interactions in complex ecological data.

• In binary classification, Gradient Boosting achieved high accuracy and strong AUC, supporting its effectiveness in this formulation.
• XGBoost demonstrated strong overall predictive capability across the study, particularly in broader analyses.
• In regression contexts, LightGBM and Random Forest showed notable performance with cross-validated R-squared around 0.73, indicating substantial explained variance.
• SHAP analyses were employed to enhance model interpretability, consistently identifying temperature- and humidity-related variables as influential predictors across model types.

• The results underscore the capacity of ensemble-based learning methods to capture multidimensional, nonlinear, and interacting effects present in ecological datasets.
• The study emphasizes generalizability and interpretability as critical considerations in model selection when handling environmental datasets with nonlinear relationships.

Where such specifics are not reported, the findings are presented as stated without extrapolation.

• The provided content does not include explicit numeric confidence intervals beyond the listed metrics, and source details do not elaborate on potential biases or data limitations.
• It is not reported whether external validation beyond the internal cross-validation schemes was conducted.