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

Area of Science:

• Agricultural Science
• Climate Science
• Data Science

Background:

• Accurate crop yield prediction is vital for food security and climate change adaptation.
• Agricultural systems face increasing vulnerability due to climate change.
• Understanding factors influencing crop yield is crucial for sustainable agriculture.

Purpose of the Study:

• To develop a hybrid modeling framework for predicting irrigated wheat yield.
• To identify key predictive features for wheat yield in Razavi Khorasan Province, Iran.
• To provide a decision-support tool for site-specific crop management and climate adaptation.

Main Methods:

• Utilized a hybrid modeling framework combining Mutual Information (MI), Recursive Feature Elimination (RFE), and NSGA-III for feature selection.
• Employed a Stacking Regressor meta-learner with LightGBM (LGBM) and Deep Neural Network (DNN) for yield prediction.
• Analyzed 47 variables from climatic, edaphic, and nutritional datasets spanning 2004-2023.

Main Results:

• Identified an optimal subset of 10 features, including temperature, soil properties, and nutrient levels, for wheat yield prediction.
• Achieved a test-set R-squared of 0.44, indicating moderate explained variance in a complex system.
• SHAP analysis revealed significant influence of regional heterogeneity and interactions between climatic, soil, and nutritional factors.

Conclusions:

• The hybrid optimization-learning pipeline effectively characterizes a substantial portion of wheat yield variation.
• The developed model serves as a practical decision-support tool for site-specific management and climate adaptation planning.
• Further research can refine models to capture remaining variability for enhanced agricultural resilience.