Probabilistic precipitation and flood risk nowcasting on pan-European scale by using deep learning

We present probabilistic warning tools for flood hazard and risk based on deep learning techniques and weather radar observations. This work has been done in the EU-funded INLINE project. 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 (1-2 km and 5-15 minutes). Pan-European precipitation nowcasts are produced from OPERA rain rate composites using a convolutional neural network based on the SimVP architecture. Training the network with alternative loss functions together with three post-processing techniques are applied to improve the utility of the nowcasts. Underestimation of localized heavy precipitation is reduced by using a cumulative distribution transformation. Stochastic post-processing is then applied to produce ensemble members that reproduce the lost spatial variability. The ensembles are generated by Fourier-filtering and transforming white noise fields to reproduce the spatiotemporal correlation structure and the distribution of forecast errors. Finally, exceedance probabilities estimated from the ensembles are calibrated by using logistic regression. We present a verification study to determine the maximum time ranges of useful skill across different spatial scales and accumulation periods and to show that the proposed methodology outperforms simple extrapolation-based techniques in most cases. We also give examples of verification metrics where this is not the case. Precipitation rates from deterministic nowcasts are translated into color-coded hazard levels by using user-specified thresholds. The hazard levels are further translated into flood risk levels by using exposure information (i.e. population or critical infrastructure). In addition, we present different methods to translate ensemble nowcasts into probability-aware predictions of hazard and risk and give discussion of their practical utility. Practical use of the proposed methodology is demonstrated by using major flood events during the years 2024 and 2025 that affected multiple European countries.