Skip to main content


Exploring the potential of machine learning techniques for flash flood forecasting

Project promoter:

Geo.X Research Network for Geosciences in Berlin and Potsdam



2017 - 2020


Project description:


The project aims to enhance the earliness and predictability of hazardous flash flood events forecasting based on using state-of-art weather data sources, physically-based hydrological models, and machine learning techniques. Here we address two key issues which limit our ability to make secure runoff predictions with short lead times: (1) the lack of reliable rainfall information at the required spatiotemporal scales, and (2) our understanding of the catchments' response to extreme local rainfall. Regarding these challenges, here we develop a novel nowcasting technique for radar rainfall sweeps with lead times up to two hours, and models for rainfall-runoff predictions, both based on top-ranked machine learning techniques: recurrent and convolutional deep neural networks. The added skill of developed research techniques has been verified at various lead times, and in different geographical regions.