A GNN-RNN Framework for Rice Yield Prediction in South Asia

Climate change poses unprecedented risks to agriculture production, thus accurate and timely crop yield forecasting is pivotal for ensuring global food security and agricultural market stabilization, especially in South Asia, a key region for rice production and export. While a growing number of studies have explored the potential of machine learning-based models for rice yield prediction, these efforts are often limited to local scales and ignore the spatiotemporal nature of rice yield in the modeling process. In this study, we propose a graph-based recurrent neural network (GNN-RNN) framework for predicting district-level rice yields in South Asia using publicly available data. The model integrates multi-source datasets, including climate observations, satellite-derived phenological metrics, soil maps, and historical yield records. By aggregating these inputs at the district scale through rice distribution masks, we extract time-series features with a Convolutional Neural Network (CNN) and utilize a GNN-RNN model to process spatiotemporal embeddings. The GraphSAGE algorithm captures geographical relationships, while the RNN component enhances predictions by incorporating temporal dependencies. Validation against five baseline machine learning models (CNN, CNN-RNN, LSTM, gradient boosting, random forest) from 2000 to 2020 shows the GNN-RNN outperforms alternatives, achieving an average R² of 0.75 and RMSE of 288 kg/ha for monsoon-season rice yields. Further tests confirm its robustness in both normal and extreme weather years, with leave-one-year-out RMSEs ranging from 234 to 366 kg/ha (11-18% of the long-term mean yield). The framework also quantifies uncertainty, with over 80% of observed yields falling within the 95% confidence interval, and prediction reliability improving throughout the growing season. This study demonstrates the potential of graph-based AI models for high-resolution crop yield forecasting, offering critical insights for food security and climate resilience. Future research could explore the model's application to extreme weather and pest impacts, as well as the integration of advanced remote sensing datasets to further enhance its predictive power.