Explainable AI for Hyper-Local Crop Yield Prediction in India

Accurate crop yield prediction is crucial for food security and economic stability in agricultural-dependent countries like India. However, while machine learning approaches have shown significant capabilities for yield prediction, a major limiting factor remains: most machine learning models are inherently complex and act mostly like a black box, making it difficult for farmers and policymakers to understand the rationale behind such predictions. This lack of transparency erodes trust and practical utility. Therefore, this paper proposed a crop yield prediction system that is not only highly accurate but also transparently explained. To develop the model, we utilized a comprehensive dataset of Indian agricultural data from 1997 to 2020. We applied several machine learning models and performed their hyperparameter tuning; among these models, XGBoost Regressor yielded the best performance, with an R -squared score of 0.93 on the test set. However, this is merely high predictive accuracy, which is not sufficient for any real-world application. In an attempt to address this limitation, an advanced explainable AI technique called SHapley Additive exPlanations was applied to interpret the model's predictions. The interpretability results show that the dominant feature of yield prediction is crop type, where coconut has the most significant impact, followed by controllable inputs like fertilizer and pesticide usages and climatic factors such as annual rainfall. This is realized by combining global and instance-specific explanations to enhance transparency and facilitate practical data-driven decisionmaking by agricultural stakeholders.