A novel hybrid NSGA-III and machine learning framework for modeling wheat yield variability using climatic, edaphic, and nutritional drivers

Abstract

Abstract Accurate prediction of crop yield remains a critical research priority due to the increasing vulnerability of agricultural systems to climate change and the growing need for food security. In this study, we developed a hybrid modeling framework to predict irrigated wheat yield in Razavi Khorasan Province, Iran, using long-term (2004–2023) climatic, edaphic, and nutritional datasets comprising 47 variables collected across 17 counties. After preprocessing, feature selection was performed using an integrated approach combining Mutual Information (MI), Recursive Feature Elimination (RFE), and the advanced NSGA-III multi-objective optimization algorithm. Final yield prediction was conducted with a Stacking Regressor meta-learner incorporating LightGBM (LGBM) and a Deep Neural Network (DNN). The optimal subset of 10 features—Tmin, TS, K, Silt, EC, HCO₃, Mg, Prec_OC, AI_Clay, and a regional indicator variable (county_te) —achieved a test-set R 2 of 0.44, reflecting a moderate yet meaningful level of explained variance given the multidimensional, nonlinear, and environmentally heterogeneous nature of the wheat production system. SHAP (SHapley Additive Explanations) analysis further highlighted the dominant influence of regional heterogeneity alongside complex interactions among climatic, soil, and nutritional factors. While the model does not capture all sources of variability, the results demonstrate that this hybrid optimization–learning pipeline reliably characterizes a substantial portion of wheat yield variation and offers a practical decision-support tool for site-specific management and climate adaptation planning.