Large-scale streamflow regionalization in ungauged West African catchments: How do classical and deep learning approaches compare?

International audience In West Africa, limited access to hydrometric data remains a major challenge for advancing surface water research and improving water management. Since the early 1980s, many gauging stations have been decommissioned, leaving gaps in reliable streamflow records across numerous catchments. Parameter regionalization of hydrological models is commonly employed to enable runoff prediction in ungauged catchments. This study represents an assessment of rainfall-runoff model regionalization across West Africa. We used an unprecedented dataset of 189 near-natural catchments to compare two contrasting approaches: (i) a benchmark conceptual modeling framework using the GR4J model, regionalized with three parameter-transfer techniques (spatial proximity, physiographic similarity, and Random Forest), and (ii) a data-driven framework based on Long Short-Term Memory (LSTM) neural networks. Using a leave-one-out resampling approach, regionalization approaches were evaluated using different performance metrics: (i) the Kling-Gupta Efficiency (KGE), calculated between simulated and observed streamflows, (ii) the relative bias (rBias) on several hydrological signatures computed with observed or simulated discharge and (iii) the difference between observed and simulated flood quantiles. Results show that the conceptual modeling approach with traditional parameter-transfer techniques consistently underperforms compared to the LSTM, failing to reproduce key hydrological signatures. In contrast, the LSTM model showed better generalization performance, accurately simulating streamflow with a median KGE of 0.67 and reliably capturing hydrological signatures and flood quantiles across West Africa’s diverse climates and landscapes with lower biases. These findings highlight the potential of data-driven approaches to enhance hydrological prediction in data-scarce regions, supporting more effective flood risk management and water resource planning.