Online-Seminarserie: Künstliche Intelligenz in der Hydrometeorologie

Larissa Tarasova (Helmholtz-Zentrum für Umweltforschung)

Prediction and Sub-Seasonal Forecasting of River Floods Using Deep Learning

Abstract

Spatially co-occurring or widespread floods pose a great threat to the resilience and there covery potential of the communities. A timely forecasting of such events plays a crucial role for increasing the preparedness of public and private sectors and for limiting the associated losses. Deep learning might be useful for this task. But which model architectures are especially skillful? Does this skill depend on the region or generation process of the corresponding flood? Do we have enough data to train such models? We answer these questions in several casestudies in Germany and China comparing forecast skill of three distinct architectures across regions with contrasting levels of dynamic predictability. Moreover, using explainable AI we investigate how generation processes of past widespread floods might change in the future and what impact this can have on the frequency and severity of future floods.

Biography

Larisa Tarasova is leading the group on “Watershed Dynamics and Hydrological Extremes” at Department Catchment Hydrology, Helmholtz Centre for Environmental Research – UFZ, Halle(Saale), Germany. She received PhD in 2020 from Martin-Luther-University Halle-Wittenberg. In her PhD research she has developed a process-based classification of streamflow events and river floods. Her recent research focuses on understanding flood generation processes at large scales and using deep learning and explainable AI to predict water extremes.