Comparative evaluation of machine learning models for predicting Cimbex quadrimaculata population density across...

Area of Science:

• Ecology and Environmental Science
• Computational Biology
• Agricultural Science

Background:

• Classical statistical models struggle with the high variability and nonlinear relationships inherent in ecological datasets.
• Environmental variables like temperature, humidity, and altitude present challenges for accurate ecological predictions.
• Analyzing complex ecological data structures requires advanced analytical approaches.

Purpose of the Study:

• To compare and investigate the performance of AI-based machine learning methods for analyzing complex ecological data.
• To evaluate different modeling approaches (binary classification, multiclass classification, regression) on the same ecological dataset.
• To assess model performance, generalizability, and explainability using various target variable definitions.

Main Methods:

• Utilized an agricultural dataset of meteorological and vegetation variables, including population observations of Cimbex quadrimaculata.
• Applied three modeling approaches: binary classification, multiclass classification, and regression.
• Employed ensemble-based boosting AI algorithms (Gradient Boosting, XGBoost, LightGBM) and SHAP analysis for interpretability.

Main Results:

• Ensemble-based boosting algorithms demonstrated high accuracy and generalizability in capturing nonlinear relationships and interactions.
• Gradient Boosting achieved 94.3% accuracy and 0.983 AUC in binary classification; XGBoost showed 84.6% accuracy overall.
• LightGBM and Random Forest achieved R² ≈ 0.73 in regression analyses, indicating robust performance.

Conclusions:

• AI-based machine learning, particularly ensemble boosting methods, excels at analyzing complex ecological data structures.
• These advanced models provide accurate and robust predictions by capturing multidimensional relationships and interactions.
• SHAP analysis enhances model interpretability, identifying key environmental predictors like temperature and humidity.