Lutonsky et al. (2026) Deep learning with upstream-aware terrain representations and rainfall adapter for generalizable pluvial flood prediction

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-09-25
• Authors: Thomas Lutonsky, Viktor C. Birschitzky, Stefan Jeschke, Andreas Buttinger‐Kreuzhuber, Nils Wandel, Daniel Cornel, Viola Haase, Reinhard Klein, Jürgen Waser, Michael H. Wimmer
• DOI: 10.1016/j.jhydrol.2026.136500

Research Groups

• VRVis Zentrum für Virtual Reality und Visualisierung Forschungs-GmbH (Wien, Austria)
• TU Wien, Institute of Visual Computing and Human-Centered Technology (Wien, Austria)
• University of Bonn, Department of Computer Science (Bonn, Germany)
• Stadtentwässerungsbetriebe Köln (Köln, Germany)

Short Summary

This study investigates the design of deep learning-based flood surrogates that can achieve robust spatial generalization across unseen terrains and modified topographies without retraining. The authors propose a two-stage training strategy with an attention-enhanced U-Net model and a rainfall adapter module to enable efficient adaptation to previously unseen rainfall events.

Objective

• Investigate the ability of deep learning-based flood surrogates to generalize across heterogeneous terrains or modified topographies.
• Develop a robust spatial representation that can be applied to various terrain configurations without retraining.

Study Configuration

• Spatial Scale: The study region covers approximately 20,000 km2 in Rhineland-Palatinate (RLP), Germany, with multiple catchments combined into single simulation domains.
• Temporal Scale: The simulations are driven by spatially uniform rainfall events, including a 2 h, 100-year Euler Type II event and natural rainfall patterns.

Methodology and Data

• Models used: Attention-enhanced U-Net model with multi-head self-attention in the bottleneck and attention gates at skip connections.
• Data sources: Hydrodynamic simulations generated using scenarify, a computational framework that solves the shallow-water equations.

Main Results

• The proposed two-stage training strategy enables the model to achieve robust spatial generalization across unseen terrains and modified topographies.
• The model demonstrates accurate predictions of maximum water depths for various rainfall scenarios, with an RMSE of 0.05 m and CSI0.3 of 0.83 over all test domains.

Contributions

• This study contributes to the development of deep learning-based flood surrogates that can generalize across heterogeneous terrains or modified topographies without retraining.
• The proposed two-stage training strategy enables efficient adaptation to previously unseen rainfall events while maintaining robust spatial generalization.

Funding

• Not specified in the paper.

Citation

@article{Lutonsky2026Deep,
  author = {Lutonsky, Thomas and Birschitzky, Viktor C. and Jeschke, Stefan and Buttinger‐Kreuzhuber, Andreas and Wandel, Nils and Cornel, Daniel and Haase, Viola and Klein, Reinhard and Waser, Jürgen and Wimmer, Michael H.},
  title = {Deep learning with upstream-aware terrain representations and rainfall adapter for generalizable pluvial flood prediction},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136500},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136500}
}