Climate-Aware Crop Yield Forecasting Using Hybrid Statistical-Machine Learning Methods: A Case Study of Odisha, India

This is the authors' abstract. We don't add key points for this paper.

Odisha has advanced agricultural productivity by blending traditional and modern practices, recording its highest-ever fruit yield of 2.95 million metric tonnes in 2024–2025 and expanding vegetable cultivation to over 762,560 hectares. Innovations like the system of rice intensification (SRI) have improved efficiency, notably reducing water use by 34% and boosting rice yield by 54% in districts like Mayurbhanj. This study introduces a hybrid framework to evaluate the impact of climate variability on crop yields in Odisha, combining interpretable regression with powerful machine learning models. Climate indicators such as rainfall, temperature, humidity, GDD, SPI, and heatwaves are aligned with crop production data to assess their influence. Regression models estimate marginal effects, while machine learning models, Random forest (RF), XGBoost (XGB), and multi-layer perceptron (MLP), capture complex patterns. Their outputs are fused via a stacked ensemble for improved accuracy and interpretability. The hybrid model achieved top performance across all metrics (lowest RMSE and MAE, highest R-squared), with SHAP analysis confirming key climate drivers. This framework offers a robust decision-support tool for policymakers and farmers, enabling adaptive, climate-resilient agricultural planning in Odisha and similar regions.