EGU24-13019, updated on 09 Mar 2024
EGU General Assembly 2024
© Author(s) 2024. This work is distributed under
the Creative Commons Attribution 4.0 License.

Analyzing drought legacy effects on streamflow with machine learning

Anne Hoek van Dijke1,2, Sungmin Oh3, Xin Yu1,4, and Rene Orth1,5
Anne Hoek van Dijke et al.
  • 1Department of Biogeochemical Integration, Max Planck Institute for Biogeochemistry, Jena, Germany (
  • 2Louis Bolk Institute, Bunnik, The Netherlands
  • 3Department of Climate and Energy Systems Engineering, Ewha Womans University, Seoul, Republic of Korea
  • 4Department of Ecology, University of Innsbruck, Innsbruck, Austria
  • 5Faculty of Environment and Natural Resources, University of Freiburg, Freiburg, Germany

Prolonged periods of below-average precipitation decrease streamflow, deplete soil moisture and groundwater reservoirs, and affect vegetation health. These effects can last for several years even after precipitation returns to normal. This way, droughts can decrease or increase streamflow for post-drought years. These drought legacy effects were found in a few local studies, but they have not yet been studied at global scale. 
Here, we study drought legacy effects on streamflow in > 1100 catchments distributed across the globe using Long-Short Term Memory (LSTM) models. This type of data-driven model is very suitable for time-series predictions with long-term dependencies, and LSTMs are therefore frequently used to model streamflow. We train our LSTM model for each catchment to predict streamflow based on meteorological forcing data. For training, we include all available data between 1980 – 2019, but we exclude the drought legacy years (the two years after each drought year). We assume that our models do therefore not know about the drought legacy effects. After training we use the LSTM models to predict streamflow for drought legacy years. We then define the legacy effects as the difference between model errors (the difference between the predicted and measured streamflow) for drought legacy years, in comparison to the model errors for normal years.
Using this methodology, we find catchments that show no, positive, or negative drought legacy effects. In the next step we will study if these legacy effects vary along climate or land cover gradients. And we additionally include satellite data of vegetation greenness, evaporation, and terrestrial water storage in the LSTM training to study two hypotheses: 1) we find negative drought legacy effects due to a depletion of groundwater, and 2) we find positive drought legacy effects, because vegetation mortality leads to decreased evaporation after the drought.
Our study offers a new perspective on understanding drought legacy effects on streamflow using observational data and demonstrates the usefulness of machine learning in uncovering complex drought impacts. 

How to cite: Hoek van Dijke, A., Oh, S., Yu, X., and Orth, R.: Analyzing drought legacy effects on streamflow with machine learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-13019,, 2024.