Czakay et al. (2026) Composition, frequency, and severity of future rain-on-snow floods in Germany

Identification

• Journal: Natural hazards and earth system sciences
• Year: 2026
• Date: 2026-11-05
• Authors: Christian Czakay, Larisa Tarasova, Bodo Ahrens
• DOI: 10.5194/nhess-26-5609-2026

Research Groups

• Institute for Atmospheric and Environmental Sciences, Goethe University, Frankfurt am Main, Germany
• Department Catchment Hydrology, Helmholtz Centre for Environmental Research – UFZ, Halle (Saale), Germany

Short Summary

This study projects future changes in rain-on-snow (ROS) flood characteristics across Germany using an ensemble of deep learning models driven by downscaled climate projections. It finds a general decrease in ROS flood frequency but an increase in the severity of widespread ROS floods, accompanied by a shift towards more rainfall-driven events.

Objective

• To investigate the future changes in the composition, frequency, and severity of rain-on-snow (ROS) driven winter floods in Germany under a high-emission climate scenario (RCP8.5).
• To assess these changes at both individual streamflow gauge stations and at a widespread, multi-catchment scale across Germany's four major river basins (Rhine, Danube, Elbe, Weser).

Study Configuration

• Spatial Scale: Germany, covering 96 river gauges within the Rhine, Danube, Elbe, and Weser River basins. Catchment areas range from 127 to 23,719 square kilometers.
• Temporal Scale:
  • Historical/Reference Period: 1951–2010
  • Future Periods (RCP8.5 scenario): 1981–2040, 2011–2070, 2041–2100
  • Daily temporal resolution for streamflow simulation and input data.

Methodology and Data

• Models used:
  • Long Short-Term Memory (LSTM) deep learning models (ensemble of 10 per gauge for streamflow simulation).
  • Explainable Artificial Intelligence (XAI) method: Integrated Gradients (IG) for attributing flood-generating processes.
  • 31 combinations of CORDEX CMIP5 Global Climate Models (GCMs) and Regional Climate Models (RCMs) (e.g., CLMcom-ETH-COSMO-crCLIM-v1-1, CNRM-ALADIN63, DMI-HIRHAM5, GERICS-REMO2015, KNMI-RACMO22E, MOHC-HadREM3-GA7-05).
  • Bias correction: Quantile Delta Mapping.
• Data sources:
  • Observed streamflow: Global Runoff Data Centre (GRDC).
  • Reanalysis data: ERA5 (0.25° × 0.25° spatial resolution) for LSTM training.
  • Climate projections: Downscaled CORDEX CMIP5 historical and RCP8.5 scenarios.
  • Elevation data: ETOPO Global Relief Model.

Main Results

• Frequency: The ensemble median projects a decrease in ROS flood frequency at most individual gauges (up to 60%), and for widespread floods in the Rhine, Elbe, and Weser River basins. Conversely, an increase in ROS flood frequency is projected for gauges in the Danube River basin and for widespread floods in this region.
• Composition: ROS floods are projected to become more rainfall-driven, with a significant decrease in the importance of snowmelt for flood generation across all regions. The IGROS (snowmelt importance ratio) decreases from a median of 2.75–3 in the reference period to 1.5–2 in the 2041–2100 period.
• Severity: The mean and maximum severity (based on normalized rank) of widespread ROS floods are projected to increase compared to the historical period across all major river basins in Germany. For individual gauges, the mean normalized rank increases for most, but the maximum rank shows mixed trends.
• Timing: Average ROS flood timing shows small shifts, with earlier occurrences in southern and eastern Germany and later occurrences in the northwest (generally within ±10 days for the ensemble median).
• Atmospheric Variables: A clear decrease in snow amount, snowmelt, and snowfall is projected, along with a significant reduction in the annual number of snow days. Liquid precipitation, however, is projected to increase in winter, compensating for some of the snowmelt loss.
• Model Variability: Significant variability exists among individual climate models regarding ROS flood trends, attributed to complex model chains, differences in precipitation and snowmelt variability, and varying melting season timings.

Contributions

• Provides the first comprehensive ensemble-based assessment of future changes in ROS flood characteristics (composition, frequency, severity) across Germany, covering both individual gauges and widespread events.
• Introduces a novel application of explainable artificial intelligence (Integrated Gradients) to attribute flood-generating processes in deep learning models, enhancing interpretability.
• Utilizes a computationally efficient deep learning approach (LSTM) to analyze a large ensemble of climate simulations, addressing limitations of previous studies that used fewer models or focused on single catchments.
• Highlights the critical need for ensemble-based assessments due to the substantial inter-model variability in projected ROS flood trends, even under consistent bias correction.

Funding

• German Research Foundation (Deutsche Forschungsgemeinschaft, DFG) – Research group FOR 2416 Space-Time Dynamics of Extreme Floods (SPATE) (grant no. FOR 2416)
• Goethe University Frankfurt (for open-access publication)

Citation

@article{Czakay2026Composition,
  author = {Czakay, Christian and Tarasova, Larisa and Ahrens, Bodo},
  title = {Composition, frequency, and severity of future rain-on-snow floods in Germany},
  journal = {Natural hazards and earth system sciences},
  year = {2026},
  doi = {10.5194/nhess-26-5609-2026},
  url = {https://doi.org/10.5194/nhess-26-5609-2026}
}