Flash floods are one of the most dangerous categories of floods and a major natural hazard. They are characterized by a very fast reaction time, with the flood wave occurring within minutes after the onset of precipitations. They pose a threat to both human lives and infrastructures in small alpine catchments. Forecasting systems delivering alerts in real-time can reduce the economic and human cost associated with flash flooding. Radar-based meteorological forecasts, thanks to technological progress in the last decades, are now able to produce spatially detailed forecasts with rapid update frequency. However, despite the continuous improvement of weather forecasting systems, predicting intense precipitation events with high spatial accuracy and sufficient lead time remains a challenge.
The goal of this study is to use ensemble weather forecasts that provide several scenarios for the short-term evolution of precipitation patterns. These products explicitly represent of the uncertainty inherent to any weather forecast. Numerous studies in the literature highlight the potential of using probabilistic methods to increase the accuracy and lead time in the context of extreme event prediction. However, the use of hydrological ensembles in flood forecasting is still in its infancy, with few studies on flash flood forecasting on small watersheds (< 100 km2).
In this study, two different ensemble products are used for flash flood forecasting. The first is the output of the NWP model COSMO1E. The second is generated by perturbing a deterministic nowcasting weather forecast, by means of a simple spatial transposition. The RS3.0 rainfall-runoff modelling software converts these meteorological ensemble into discharge ensembles. The forecast performance are then evaluated for two small watersheds in Switzerland with risks of flash flooding: the Altbach in Kloten and the Flon in Lausanne. Ensemble-based alarm systems are compared to deterministic alarm for historical events, offering insight into the added value of this new method.
The scope of this study is limited, with a short study period and only two watersheds being investigated. Nevertheless, results show that both ensembles have higher skill than deterministic forecasts for flash flood forecasting. In Kloten, ensembles allow a better estimation of the flood peak intensity that tends to be underestimated by deterministic systems. And in Lausanne, where the performance of a deterministic systems are not sufficient for flash flood forecasting, the ensemble approach drastically increase the proportion of detected events. In both cases, these increased detection performances come with a trade-off: the ensemble alarm systems tend to generate higher rates of false alarms, even more so when using operational-like continuous simulations.