Integrating machine learning approaches to evaluate faba bean varietal performance under fall and spring plantings

This study applies machine learning (ML) to predict grain yield and classify grain yield in spring- and fall-sown faba bean ( Vicia faba L.) varieties. Five feature selection methods, Correlation-based Feature Subset Selection, Correlation-Based Attribute Evaluator, Principal Components Attribute Transformer, ReliefF Attribute Evaluator, and Stepwise Regression, were evaluated in combination with Linear Regression, Random Forest, and Multilayer Perceptron for yield prediction. For classification into Low, Medium, Good, and Exceptional categories, Logistic Regression and Random Forest were applied using season-specific features. For spring planting, Linear Regression consistently outperformed other models, achieving high correlation coefficients (0.95–0.98) and low error metrics (MAE: 1.01, RMSE: 1.62), with pod weight and grain yield per branch emerging as key predictors. For fall faba bean yield, Linear Regression and Multilayer Perceptron achieved high accuracy (CC 0.84–0.99) with key predictors including pod weight, shoot dry weight, seeds per plant, and yield per branch. Classification performance varied by season. Logistic Regression achieved moderate accuracy in spring (79.5 %) but struggled with high-yield classes. In contrast, Random Forest performed best in fall, reaching 91.9 % accuracy (Kappa 0.88) with strong class-specific F-measures (0.83–0.94). Overall, Linear Regression and Multilayer Perceptron were effective for yield prediction, while Random Forest was most suitable for fall classification. This study demonstrates that utilizing the appropriate ML methods can enhance prediction accuracy, mitigate class imbalances, and facilitate more informed decisions for sustainable faba bean production.