Probabilistic rainfall and flash flood nowcasting on pan-European scale

We present probabilistic warning tools for flood hazards and risks based on deep learning (DL) techniques. The scope is on nowcasting heavy rainfall and associated flash floods in short time ranges (0-3 hours) and at high spatial and temporal resolutions (2 km and 15 minutes). Rainfall nowcasts are produced from pan-European OPERA radar composites by using a convolutional neural network based on the SimVP architecture. Two post-processing techniques are applied to enhance the utility of the DL-based nowcasts. First, underestimation of heavy rainfall is reduced by applying quantile mapping. Second, ensembles that provide realistic estimates of forecast uncertainty are generated by utilizing a stochastic technique. With these enhancements, the DL-based nowcast is shown to outperform the traditional extrapolation-based nowcasting techniques. The rainfall nowcasts are translated into color-coded hazard levels by using user-specified thresholds and statistically optimized probability thresholds that maximize hits and minimize false alarms. These are further translated into flood risk levels by using exposure information. Real-time feed of the warning products is displayed in a web platform developed in the EU-funded INLINE project. Demonstrations of the proposed methodology are given using major flood events during the years 2024 and 2025 that affected multiple European countries.