Machine Learning Forecasts Rice Yields in Data-Scarce Regions Using Satellite Data

Researchers investigated the feasibility of using machine learning to predict rice yields in Sierra Leone, a country with limited agricultural data infrastructure. They trained various models, including XGBoost, Gradient Boosting, and Random Forest, using 25 years of national crop production data. Initially, models relying solely on crop statistics did not outperform a simple persistence benchmark. However, when the models were augmented with readily available satellite climate data, specifically CHIRPS rainfall and NASA POWER temperature, the forecasting accuracy improved substantially. An XGBoost model using only climate data reduced forecast error by one-third. The study highlighted early-season rainfall as a key predictor, suggesting that potential yield risks can be identified months before harvest. The findings offer practical policy recommendations for Sierra Leone's agricultural strategy, emphasizing the value of integrating climate data for better decision support, even in resource-limited environments. The entire methodology is provided as an open-source pipeline.