Improvised wheat yield forecasting via SHAP-based feature Selection and Bayesian optimized weighted ensemble model

Abstract Prediction of wheat yield is complex due to the nonlinear and multivariable nature of climatic influences, necessitating advanced analytical techniques. Punjab is a key wheat producer in India, which primarily grows during the rabi (winter) season. This study proposes a Hybrid Feature Selection (HFS) method that incorporate multiple feature selection strategies and validates the chosen features using Shapley Additive Explanations (SHAP) to ensure transparency and robustness. Using a dataset of 1014 data instances collected over 13 years from 2011–2023 in three districts of Punjab state specifically Amritsar, Ludhiana, and Patiala, the Bayesian Optimized Weighted Ensemble Model (BOWEM) is suggested to increase the accuracy of wheat yield prediction using key climatic parameters from November to April. The model effectively captures nonlinear yield-climate relationships by combining Random Forest, XGBoost, and CatBoost with district-specific optimal weights. Standard error metrics like Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE) are used to assess the model’s accuracy. The BOWEM tuned the weight allocations and exhibit high accuracy, especially in Amritsar. The model efficiency of 91.11% with low error metrics (MAE: 35.78, RMSE: 83.35) is achieved. It outperformed baseline and previously published studies. Statistical validation confirmed the importance of these improvements, demonstrating the strength of the proposed approach. The results highlight the effectiveness of combining multiple models to enhance yield prediction accuracy, providing a valuable tool for informed agricultural decision-making. The main contribution is integrating hybrid feature selection, SHAP-based interpretability, and Bayesian optimization in an ensemble framework for region-specific yield forecasting. The proposed framework enhances the interpretability and reliability of crop yield forecasting, contributing meaningfully to data-driven agricultural planning and climate-resilient farming strategies.