Author(s): Rodrigo Bezerra; Bruno Brentan; Ruben Imhoff; Andre Rodrigues

Linked Author(s):

Keywords: Flood forecasting; Machine learning; QPF nowcasts; Radar rainfall; Hydrological modelling

Abstract: The increasing frequency and intensity of extreme rainfall events underscore the urgent need for reliable shortterm flood forecasting systems. Machine learning (ML) models have demonstrated strong potential for datadriven hydrological prediction; however, their effective forecast lead time remains constrained by the catchment's hydrological response, as rainfall information must propagate through the basin before discharge signals emerge. Integrating radar-based Quantitative Precipitation Forecasts (QPF) or nowcasting, offers a promising yet underexplored opportunity to extend ML based flood forecast horizons. This study investigates , the potential and limitations of using radar nowcasts as input for ML flood forecasting in the Geul River catchment—a steep 344 km² basin in the southern Netherlands that suffered catastrophic flooding in July 2021, causing over €200 million in local damages. The event exposed critical weaknesses in the operational forecast chain, particularly regarding the use of high-resolution rainfall predictions. Two baseline configurations are compared: (i) a forecast system using only observed rainfall up to time t, and (ii) a system combining observations up to t with “perfect” rainfall forecasts for t + T. This idealized framework quantifies the theoretical skill gain achievable through QPF integration, providing a benchmark for future hybrid flood forecasting systems.

DOI: https://doi.org/10.64697/iahr.proc.hic2026.258

Year: 2026