In our newly published study, we developed a novel hybrid framework that integrates the NSGA-III multi-objective optimization algorithm with advanced machine learning techniques to model wheat yield variability across diverse agroecological conditions. The framework combines feature engineering, evolutionary optimization, LightGBM, deep neural networks, and explainable artificial intelligence (XAI) to identify the most influential climatic, edaphic, and nutritional drivers of wheat productivity.

Using a unique long-term dataset (2004–2023) covering 17 counties and 47 environmental variables in northeastern Iran, we employed a three-stage feature selection strategy involving Mutual Information (MI), Recursive Feature Elimination (RFE), and NSGA-III optimization. The algorithm identified an optimal subset of only 10 variables capable of explaining a substantial portion of yield variability while avoiding unnecessary model complexity.

The final stacking model, combining LightGBM and Deep Neural Networks, successfully captured complex nonlinear interactions among climate, soil, and nutrient variables. Results revealed that minimum temperature, temperature seasonality, soil salinity (EC), bicarbonate concentration, potassium, magnesium, silt content, and interactions between precipitation and soil organic matter were among the most influential factors shaping wheat yield patterns.

To move beyond black-box predictions, we applied SHAP (SHapley Additive Explanations) analysis to interpret model behavior. The results uncovered hidden relationships between climatic stressors, soil constraints, and regional environmental characteristics, highlighting how nighttime warming, salinity stress, and soil-climate interactions jointly influence yield formation. The analysis also demonstrated that regional heterogeneity remains a dominant driver of wheat productivity across the study area.

Beyond predictive performance, this work illustrates how multi-objective evolutionary algorithms and explainable machine learning can be integrated into agricultural decision-support systems. The proposed framework offers a practical pathway toward climate-smart agriculture, improved input management, and location-specific adaptation strategies for sustainable wheat production.

The methodology is transferable to other crops and agroecosystems where complex interactions between climate, soil, and management practices limit the effectiveness of traditional modeling approaches.

📄 Full article:
https://doi.org/10.1038/s41598-026-48918-0