Reis et al. (2026) Improving flood forecasts: the combined impact of data assimilation and machine learning post-processing

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-09-16
• Authors: Gustavo Gabbardo dos Reis, Paul C. Astagneau, François Bourgin, Vazken Andréassian, Charles Perrin
• DOI: 10.1016/j.jhydrol.2026.136402

Research Groups

• Université Paris-Saclay, INRAE, HYCAR Research Unit, Antony, France
• WSL Institute for Snow and Avalanche Research SLF, Davos Dorf, Switzerland
• Institute for Atmospheric and Climate Science, ETH Zurich, Zurich, Switzerland
• Climate Change, Extremes and Natural Hazards in Alpine Regions Research Center CERC, Davos Dorf, Switzerland

Short Summary

This study assesses the combined impact of machine-learning-based post-processing, data assimilation, and calibration strategies on hourly streamflow forecasts for 687 catchments in metropolitan France. It finds that post-processing consistently improves forecast skill at short lead times, particularly for slow-response catchments, and can partially compensate for the absence of state updating, though it does not fully replace data assimilation.

Objective

• To quantify the respective and combined contributions of model calibration, state updating through data assimilation, and machine learning post-processing to hourly streamflow forecast skill for lead times up to 24 hours.
• Can machine learning post-processing compensate for the absence of state updating during forecasting?
• Can lead-time-dependent calibration combined with data assimilation be further improved by machine learning post-processing?
• How does the combined use of calibration, state updating, and machine learning post-processing affect forecast performance across catchment response times and prediction horizons?
• Which machine learning algorithm performs better in post-processing, Random Forest or Multilayer Perceptron?

Study Configuration

• Spatial Scale: 687 catchments across metropolitan France, ranging in area from 7 km² to 7918 km².
• Temporal Scale: Hourly streamflow forecasts at lead times of 3, 6, 12, and 24 hours. The analysis period for hydrometric data spans from January 1, 2007, to December 31, 2021.

Methodology and Data

• Models used:
  • Hydrological model: GR5H-RI (a modified version of the GR5H lumped rainfall–runoff model).
  • Data assimilation: Deterministic state-updating procedure based on direct insertion, adjusting routing store levels to match observed streamflow.
  • Machine learning post-processing models: Random Forest (RF) and Multilayer Perceptron (MLP).
• Data sources:
  • Hourly streamflow records: HydroPortail platform (French national hydrometric services).
  • Hourly precipitation data: COMEPHORE reanalysis product (Météo-France), 1 km² spatial resolution.
  • Potential evapotranspiration (PET): Estimated using the Oudin et al. (2005) formulation, based on air temperature and extraterrestrial radiation.
  • Mean daily air temperature: Météo-France SAFRAN gridded product.
  • Elevation data: Shuttle Radar Topography Mission (SRTM) digital terrain model (100 m resolution) for continental France, and BD ALTI v1.0 dataset (25 m resolution) by IGN for Corsica.

Main Results

• Machine-learning-based post-processing consistently improved forecast skill at short lead times, particularly for slower-response catchments.
• For the WsRf (lead-time-dependent parameters with data assimilation) approach, median bounded persistence scores (CPb) during flood events increased from 0.33 to 0.54 at a 3-hour lead time and from 0.44 to 0.56 at a 6-hour lead time with Multilayer Perceptron (MLP) post-processing.
• Performance gains (CPb > 0.05) were observed in 89.7% of catchments at a 3-hour lead time and 71.6% at a 6-hour lead time.
• Open-loop (OL) forecasts (without data assimilation) showed the largest relative improvements, with median CPb increasing from -0.56 to 0.16 at a 3-hour lead time and from -0.08 to 0.37 at a 6-hour lead time with MLP post-processing.
• Post-processing can partially compensate for the lack of state updating but cannot fully replace it; forecasts incorporating data assimilation generally maintained superior overall skill.
• Multilayer Perceptron (MLP) slightly outperformed Random Forest (RF) in continuous skill metrics (CPb), while differences were less pronounced for event-based evaluation (Critical Success Index, CSI).
• For WsRf at a 3-hour lead time, median CSI increased from 87.6% for raw forecasts to 91.1% with MLP and 91.4% with Random Forest.

Contributions

• This study extensively investigates the combined effects and interactions between machine-learning-based post-processing, data assimilation, and different hydrological model calibration strategies within a flood forecasting framework, an area previously insufficiently explored.
• It quantifies the complementary roles of data assimilation and machine learning post-processing across various lead times and catchment response times.
• It demonstrates the potential of such integrated frameworks for operational flood forecasting, highlighting a computationally efficient post-processing strategy applicable to different calibration and assimilation approaches using a minimal set of predictors.

Funding

• Service Central Vigicrues
• DRHYM project
• HPC resources from GENCI–IDRIS (Grants 2024-AD010116031 and 2025-AD010116031R)

Citation

@article{Reis2026Improving,
  author = {Reis, Gustavo Gabbardo dos and Astagneau, Paul C. and Bourgin, François and Andréassian, Vazken and Perrin, Charles},
  title = {Improving flood forecasts: the combined impact of data assimilation and machine learning post-processing},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136402},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136402}
}