EGU24-8899, updated on 08 Mar 2024
EGU General Assembly 2024
© Author(s) 2024. This work is distributed under
the Creative Commons Attribution 4.0 License.

Deep Learning for Spatially Distributed Rainfall–Runoff Modeling

Martin Gauch1, Frederik Kratzert1, Vusumuzi Dube1, Oren Gilon1, Daniel Klotz2, Asher Metzger1, Grey Nearing1, Florence Ofori1, Guy Shalev1, Shlomo Shenzis1, Tadele Tekalign1, Dana Weitzner1, Oleg Zlydenko1, and Deborah Cohen1
Martin Gauch et al.
  • 1Google Research
  • 2Helmholtz Centre for Environmental Research (UFZ), Leipzig, Germany

Deep learning approaches have emerged as the state of the art for rainfall–runoff modeling. Yet—until now—the best-performing models have typically been used with inputs that are averaged across possibly large catchment areas, modeling each (sub-)basin independently. This lumped modeling approach is in contrast to reality, where rivers form networks of connected subbasins. This discrepancy limits our ability to accurately and interpretably predict certain types of rivers. Here, we present recent work to build graph-based deep learning models that explicitly account for this network structure. These models promise to unlock improvements in both the quality and interpretability of predictions:

Catchment size: Lumped models lack information about spatial heterogeneity, i.e., they do not know where inside a basin events (such as precipitation) occur. This makes it hard to model large basins, especially at high temporal resolution. Spatially distributed models can explicitly learn to account for travel times between subbasins inside a larger catchment, which also allows to analyze runoff generation separately from routing behavior.

Data assimilation: In lumped models, it is hard to include (real-time) measurements from upstream river sections that could improve predictions at downstream locations. In a graph-based model of connected subbasins, any downstream prediction can benefit from upstream information that is propagated along the river network.

Human intervention: It is unclear how to explicitly represent human water extraction, reservoirs, or dams in lumped models. Graph-based models provide more flexibility, as we can account for the spatial location of human interventions explicitly in the graph structure and learn to represent their influence on runoff.

In this work, we compare different types of spatially distributed deep learning models with lumped deep learning models and traditional physics-based hydrologic modeling and routing approaches on >600 gauges of the LamaH dataset [1].

[1] Klingler, C., Schulz, K., and Herrnegger, M.: LamaH-CE: LArge-SaMple DAta for Hydrology and Environmental Sciences for Central Europe, Earth Syst. Sci. Data, 13, 2021.

How to cite: Gauch, M., Kratzert, F., Dube, V., Gilon, O., Klotz, D., Metzger, A., Nearing, G., Ofori, F., Shalev, G., Shenzis, S., Tekalign, T., Weitzner, D., Zlydenko, O., and Cohen, D.: Deep Learning for Spatially Distributed Rainfall–Runoff Modeling, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-8899,, 2024.