Modern computational approaches for rice yield prediction: A systematic review of statistical and machine learning-based methods

Context: Rice is a critical food source for over half the global population, making yield forecasting important for food security and agricultural planning. Advances in statistical, machine learning, and deep learning approaches, coupled with data sources such as hyperspectral and UAV-derived imagery, have transformed the practice of rice yield forecasting. Objective: This review assesses the effectiveness of such data-driven approaches for rice yield prediction, focusing on model accuracy, scalability, and generalizability, while identifying existing gaps and future directions. Methodology: A systematic review of 156 studies was conducted using PRISMA guidelines, covering machine learning models such as Neural Networks, Random Forest, and Support Vector Machines. The input data comprised satellite-based indices (e.g., NDVI, MODIS), climate variables (precipitation, temperature), and ground observations. Performance metrics, such as R2, MAE, and MAPE, were analyzed to evaluate model performance across various regions and conditions. Results and conclusions: Neural Networks and Random Forest models were the most commonly used, with satellite and climate variables being integral to predictions. Key challenges identified include limited data availability, overfitting, and difficulty in generalizing models across regions and rice cultivars. Several gaps remain in standardizing evaluation metrics, incorporating ground-truth data, and adapting models for real-world agricultural decision-making. Future research should focus on integrating multiple data sources, enhancing model scalability, and developing frameworks for better uncertainty quantification. Significance: Addressing these challenges can enable AI-driven models to better support agricultural management and policy, improving resource allocation and food security in the face of climate variability.