Machine learning framework for the reconstruction of missing daily rainfall time series from multisource satellite and in situ observation in Burkina Faso

Accurate climate data are essential for sustainable development and climate resilience, yet missing data remain a major challenge in data-sparse regions such as Burkina Faso. This study quantifies the performance gain of hybrid classification–regression machine learning (ML) models for reconstructing missing rainfall time series, relative to raw satellite data imputation. Unlike simple bias correction, the approach combines five satellite products with ML, learning from the diversity among datasets rather than from any single dataset’s bias; a two-step regression–classification scheme forms its final stage. Once trained, the model requires no in situ data, an approach termed the crossed ML approach. The reconstruction uses five satellite products (TAMSAT, RFE, ARC, CHIRP, CHIRPS), geographic coordinates, climatic zone, and month of year, together with in situ observations from Burkina Faso’s national rain gauge network (2000–2020). Regression models (linear regression, decision trees, random forests, boosted trees, multilayer perceptrons) are coupled with classification models for rainfall event detection, yielding 25 crossed ML models compared against 5 uncrossed ML models and 5 raw satellite datasets, evaluated using RMSE, MAE, ME, PBIAS, POD, and FAR. ML models outperform most satellite estimates, particularly in error reduction. The best models reduce the annual median RMSE by about 2.5 relative to most satellite datasets, improve the coefficient of determination by over 0.5, and lower the false alarm rate by more than 0.6. While uncrossed ML models yield higher detection probability, crossed ML models better balance detection and false alarms. These findings highlight ML’s potential to improve rainfall reconstruction and data availability.