Improving flood forecasts: the combined impact of data assimilation and machine learning post-processing
Abstract
Accurate streamflow forecasts are essential for flood risk management and impact mitigation. Coupling hydrological models with machine learning techniques has gained increasing attention as a means to improve forecast skill, with post-processing emerging as a promising approach to correcting systematic model residuals. However, the interaction between machine-learning post-processing, data assimilation and calibration strategies remains insufficiently explored. In this study, we assess the contribution of machine-learning-based post-processing to hourly streamflow forecasts for 687 catchments in metropolitan France, generated using a GR-family hydrological model (GR5H-RI) under three forecasting approaches differing in calibration strategy and use of data assimilation. Random Forest and Multilayer Perceptron models are used to post-process forecasts at lead times of 3, 6, 12 and 24 h. Forecast performance is evaluated using both continuous skill metrics relative to persistence and threshold-based metrics for flood event detection. Results show that (1) post-processing consistently improves forecast skill at short lead times, especially for slower-response catchments, with the median persistence score increasing from 0.33 to 0.54 at a 3-h lead time in the highest-skill forecasting setup and performance improving in 89.7% of catchments; (2) the largest gains occur for forecasts without data assimilation, suggesting that post-processing can partially compensate for the lack of state updating but cannot fully replace it; and (3) neural network-based post-processing slightly outperforms tree-based models in continuous metrics, while differences are more limited for event detection. Overall, results highlight the complementary roles of data assimilation and machine-learning post-processing and demonstrate the potential of such frameworks for operational flood forecasting.
Related
- Improving global precipitation forecasts with an AI weather model trained on satellite observations
- A cost-effective data assimilation system for improving ROMS simulations in the Southwest Atlantic
- Improving large-scale river routing models with ESA long-term CCI discharge data assimilation: advancing accuracy for climate studies
- The interface of data assimilation and machine learning
- Explore the ideology of deep learning in ENSO forecasts
BibTeX
@article{reis2026improving,
title = {{Improving flood forecasts: the combined impact of data assimilation and machine learning post-processing}},
author = {Gustavo Gabbardo dos Reis and Paul Astagneau and François Bourgin and Vazken Andréassian and Charles Perrin},
journal = {Journal of Hydrology},
year = {2026},
month = sep,
doi = {10.1016/j.jhydrol.2026.136402},
url = {https://hal.inrae.fr/hal-05766794v1},
note = {CC BY},
}
RIS
TY - JOUR
TI - Improving flood forecasts: the combined impact of data assimilation and machine learning post-processing
AU - Gustavo Gabbardo dos Reis
AU - Paul Astagneau
AU - François Bourgin
AU - Vazken Andréassian
AU - Charles Perrin
PY - 2026
DA - 2026/09/28
JO - Journal of Hydrology
DO - 10.1016/j.jhydrol.2026.136402
UR - https://hal.inrae.fr/hal-05766794v1
AB - Accurate streamflow forecasts are essential for flood risk management and impact mitigation. Coupling hydrological models with machine learning techniques has gained increasing attention as a means to improve forecast skill, with post-processing emerging as a promising approach to correcting systematic model residuals. However, the interaction between machine-learning post-processing, data assimilation and calibration strategies remains insufficiently explored. In this study, we assess the contribution of machine-learning-based post-processing to hourly streamflow forecasts for 687 catchments in metropolitan France, generated using a GR-family hydrological model (GR5H-RI) under three forecasting approaches differing in calibration strategy and use of data assimilation. Random Forest and Multilayer Perceptron models are used to post-process forecasts at lead times of 3, 6, 12 and 24 h. Forecast performance is evaluated using both continuous skill metrics relative to persistence and threshold-based metrics for flood event detection. Results show that (1) post-processing consistently improves forecast skill at short lead times, especially for slower-response catchments, with the median persistence score increasing from 0.33 to 0.54 at a 3-h lead time in the highest-skill forecasting setup and performance improving in 89.7% of catchments; (2) the largest gains occur for forecasts without data assimilation, suggesting that post-processing can partially compensate for the lack of state updating but cannot fully replace it; and (3) neural network-based post-processing slightly outperforms tree-based models in continuous metrics, while differences are more limited for event detection. Overall, results highlight the complementary roles of data assimilation and machine-learning post-processing and demonstrate the potential of such frameworks for operational flood forecasting.
SN - 0022-1694
SN - 1879-2707
ER -