Comparative assessment of quantitative precipitation estimates, numerical weather prediction -and nowcasting methods for flood prediction in West Africa

Accurate prediction of extreme rainfall is critical for effective flood warning and mitigation, particularly in data-scarce regions. This study investigates the application of satellite-based precipitation, Numerical Weather Prediction (NWP) models, and nowcasting methods integrated with a hydrological model, the Ensemble Framework for Flash Flood Forecasting (EF5), for flood detection across 10 different events that occurred in different cities of West Africa. On the first part, five precipitation inputs—Integrated Multi-satellite Retrievals for GPM (IMERG) Early Run, Global Forecast System (GFS), Weather Research and Forecasting model (WRF), Global Satellite Mapping of Precipitation (GSMaP), and Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks – Dynamic Infrared Rainfall Nowcasting (PERSIANN-PDIRNow)—were analyzed, using IMERG-Early as the reference dataset for comparison. The study explores the ability of these precipitation products to provide flood predictions, demonstrating the potential of satellite precipitation in replicating rainfall evolution and flood-response. Statistical evaluation using ,Kling–Gupta Efficiency (KGE), and Relative Bias (RB) indicates that the forecasts such as GFS was not able to capture the flooding events when simulated one day prior to the event whereas WRF was mostly linked with high FAR. GSMaP was the only dataset that was able to capture the flood events in most of the cases, while PDIR-Now was able to capture some but generally tended to show flooding signals prior to the event. Additionally, in the second part, analysis was done using nowcasting algorithms such as the Lagrangian persistence method, the Local INtensity Downscaling Approach (LINDA), and the Convolutional Long Short-Term Memory network (ConvLSTM), in which the Lagrangian method was able to detect 2 events among the 10 events, while LINDA and ConvLSTM were not able to detect the floods. This study underscores the assessment of the potential of integrating nowcasting algorithms with satellite precipitation for real-time flood forecasting in developing regions.