International Journal of Environment and Climate Change

Seasonal forecasts at coarse-resolution limit local decision-making and disaster (drought, flood…) preparedness in African small watersheds. To address these limitations, this study designs a Deep Learning-based downscaling framework using the U-Net Convolutional Neural Network (CNN) architecture to transform coarse 1° (~100 km) forecasts into high-resolution 0.05° (~5 km) gridded data. For this, it assesses raw seasonal forecasts of precipitation and temperature from three global models (ECMWF, Météo-France, and CMCC) and their ensemble mean (Ensmean) over Burkina Faso. Forecast skill was assessed against high-resolution reference datasets (CHIRPS for precipitation and CHIRTS for temperature) using standard statistical metrics (r, RMSE and MAE). The raw forecasts showed weak performance in representing spatial variability with biases above 20% for precipitation and more than 5 °C. However, after the downscaling process, the generated high-resolution outputs showed substantial improvements in skill compared to the raw forecasts, with gains of up to sixfold for precipitation and twenty-fold for temperature. A modified Taylor diagram incorporating these metrics was employed to identify ECMWF as the best-performing model for precipitation and METEO-FRANCE for temperature. By producing high-resolution seasonal precipitation and temperature datasets, this study demonstrates the added value of U-Net Deep Learning-based downscaling for Burkina Faso. While the initial application targeted the Nouhao sub-basin for drought modeling, the framework was extended to cover the entire country to make high-resolution data available for broader applications. These results provide new insights into integrating Deep Learning approaches into operational drought and flood prediction frameworks of Burkina Faso’s National Agency of Meteorology (ANAM) and the General Directorate of Water Resources (DGRE). They also contribute to improved understanding of complex seasonal climate hazards across West Africa and thereby enhancing related areas such as hydrological modeling.