Climate-Sensitive Modeling of Rice Stock in Odisha Using Hybrid Machine Learning and Explainable AI

Climate variability has increased uncertainty in agricultural production and post-harvest management, yet most existing studies focus on average crop yields at aggregated spatial scales and rely either on linear statistical models with limited flexibility or black-box machine learning models with poor interpretability. As a result, district-level forecasting of climate-sensitive rice stock dynamics remains insufficiently addressed, particularly in highly vulnerable regions such as Odisha, India. This study addresses this gap by developing a district-level, climate-aware forecasting framework that jointly improves accuracy, robustness, and interpretability. A stacked hybrid model integrating an Artificial Neural Network and XGBoost through a linear meta-regressor is proposed to capture nonlinear climate effects and structured residual patterns in rice stock variability. Monthly agro-climatic variables including temperature, rainfall, humidity, wind speed, and evapotranspiration are combined with district-level rice stock data for representative agro-climatic zones of Odisha. The model is evaluated against statistical baselines, time-series benchmarks, and standalone machine learning models using a time-aware validation strategy. Results show that the proposed hybrid model outperforms all comparators, achieving anR2of 0.9727 and an MSE of 1.73 × 10⁹ . SHAP-based explainability identifies evapotranspiration and maximum temperature as dominant drivers, while high humidity and cumulative rainfall negatively affect stock stability, especially under extreme conditions. These findings demonstrate that climate-aware hybrid modelling provides reliable and interpretable forecasts, supporting practical applications such as adaptive sowing decisions, irrigation planning, storage management, and evidence-based rice stock regulation in climate-vulnerable regions.