Dynamic and scalable real-time prediction of heavy rainfall and resulting flooding using deep learning applications

• Applicants: Professor Dr.-Ing. Anselm Haselhoff; Professor Dr.-Ing. Markus Quirmbach
• Subject Area: Hydrogeology, Hydrology, Limnology, Urban Water Management, Water Chemistry, Integrated Water Resources Management
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 537919295

This research proposal addresses the improvement of heavy rainfall and flooding forecasts, particularly in urban areas. In contrast to purely physically based model approaches, data-driven model approaches from the field of artificial intelligence (AI), especially machine learning (ML), are to be further developed and applied in this work. The objectives of this research proposal result from the deficits of current ML models in the application context of nowcasting that were identified in the KIWaSuS project. Therefore, basic scientific investigations are to be carried out in the fields of precipitation and flood forecasting, on the basis of which a pre-competitive prototype can be developed that includes both components. The focus of the ML models for precipitation forecasting is on improving the quality (generalization capability) and reliability of the forecasts. To this end, previously used ML model structures are to be expanded in order to better take specific water management requirements (fusion of multimodal sensor data) into account and to enable uncertainty quantification for the high-dimensional temporal data sequences of spatially and temporally distributed time series (e.g. radar data). In the case of flood forecasting models, on the other hand, methods for scaling and generalizing the models to larger areas are primarily being investigated. Among other things, it will be examined whether the method of active learning leads to an improvement. After initially developing both model types separately, a hydro-meteorological model chain will ultimately be developed in which both forecasting models are coupled.

• DFG Programme: Research Grants (Transfer Project)
• Application Partner: fuseki GmbH