Spatial assessment of rice straw yield under future climate pathways using interpretable stacked machine learning models

Abstract

This study develops a hybrid stacked ensemble (SE) machine learning framework integrated with CMIP6 climate projections to generate high-resolution spatial predictions of rice straw yield in Eastern India. The SE approach combines Cubist, Random Forest (RF), Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGB), Multivariate Adaptive Regression Splines (MARS), and Support Vector Machine (SVM), and was validated using 1,780 farmer-reported field observations. Among the models, SE-MARS achieved the highest predictive accuracy (R² = 0.775). Current straw yields were estimated to range from 0.10 to 6.78 t/ha, while future projections under SSP2-4.5 and SSP5-8.5 scenarios indicate yields reaching 4.75–8.85 t/ha in Bankura and parts of Birbhum. Feature importance analysis identified precipitation (pr) as the dominant predictor (Boruta score = 34.58; Sobol first-order ≈ 0.40; total effect ≈ 0.45), whereas available water capacity showed comparatively lower influence (13.32). SHAP results further confirmed soil moisture, precipitation, elevation, and soil temperature as key controlling factors. These findings demonstrate the robustness of the SE framework for climate-resilient straw yield assessment and sustainable residue management planning.