Satellite and AI-Driven Rainfall Nowcasting Framework for Climate-Smart Agriculture in the Sahel: The Case of Burkina Faso
Bado Xavier*, Belko Abdoul Aziz Diallo, Souleymane Zio and Borli Michel Jonas Some
ABSTRACT
Accurate precipitation forecasting is vital for Sahelian countries like Burkina Faso, where rainfed agriculture drives the economy and erratic rainfall complicates water management and disaster preparedness. This study develops and evaluates three machine learning models—CatBoost, CNN, and a hybrid CNN-LSTM—for precipitation nowcasting using multi-source satellite data. Leveraging Google Earth Engine, we integrated GPM-IMERG (V07)(calibrated precipitation), GOES-16 (cloud and moisture indices), elevation, and CHIRPS(calibrated precipitation) data. GPM-IMERG (V07) was selected over CHIRPS based on higher correlation with ground-based observations from nine weather stations over 2010–2020. Model training used data from July 10, 2017, to December 31, 2021, with testing and validation from January 1, 2022, to June 21, 2024. GPM-IMERG (V07) outperformed CHIRPS in Probability of Detection (POD) and Critical Success Index (CSI). CatBoost achieved an RMSE of 1.23, MAE of 0.42, and POD of 84%, while CNN recorded an RMSE of 1.29, MAE of 0.32, and POD of 57% (threshold 0.2). The CNN-LSTM hybrid effectively captured spatial and temporal precipitation patterns. This research provides a reproducible framework that enhances forecasting tools for West Africa, with significant implications for supporting disaster preparedness, and agricultural planning.
