Fertilizer-Focused Stacked Ensemble Model for Explainable Paddy Yield Prediction Using Context-Residualized Nutrient Sensitivity Learning

Abstract

Objectives: To build an explainable model that predicts paddy yields based on the fertilizer application, removing the influence of the climatic, irrigation and spatial factors. Method: The study suggests a new model called Fertilizer-Focused Stacked Ensemble Model (FFSEM). Context residualisation eliminates contextual effects. Mutual information, fertilizer-context synergy, redundancy suppression and stability penalization generate fertilizer sensitivity scores. Softmax weighting forms a Context-Integrated Feature (CIF) space. The Gradient Boosting meta-learning technique is used for stacking regressors like Light GBM, XGBoost and Multilayer Perceptron (MLP). Findings: FFSEM gave the best Root Mean Square Error (RMSE of 0.0406), Mean Absolute Error (MAE=0.0287), Mean Absolute Percentage Error (MAPE of 0.202) and R2 of 0.9512 for paddy yield of Tamil Nadu district level data. The model performed better than the existing ensemble models. Nitrogen was found to be the most important fertilizer component and Theni, Dharmapuri, Thiruvarur and Dindigul districts had been identified as high fertilizer-sensitive districts using SHapley Additive exPlanations (SHAP). Novelty: The proposed model provides context-residualized fertilizer sensitivity learning, dependency, synergy, redundancy and stability-aware nutrient weighting. Keywords: Paddy yield, Fertilizer sensitivity, Stacked ensemble, Context residualization, Explainable artificial intelligence