Comparison of deep learning and conceptual models for the prediction of flood statistics in ungauged catchments in France

Predicting flows in ungauged basins is a key challenge for integrated water resource management and hydrological risk prevention. To fill these gaps, regionalization approaches have generally relied on conceptual or physically-based hydrological models whose parameters are calibrated on gauged rivers and then transferred to ungauged rivers. However, these methods often have significant limitations in terms of robustness and accuracy, particularly when applied in heterogeneous hydrological contexts [1].

With the growing adoption of machine and deep learning by the hydrological community, new opportunities are emerging for operational hydrology. In particular, recurrent neural networks such as Long Short-Term Memory networks (LSTM) have proven to be effective in exploiting large databases of observed flows and climate forcings compared to traditional locally or regionally calibrated approaches [2]. Nevertheless, LSTM is still rarely used in France for prediction outside of a few research projects (e.g. [3,4]).

The objective of this study is to compare an LSTM model with the GR5J model [5,6] in a regionalization exercise with a special focus in flood prediction. The evaluation is carried out on the French catchments of the Explore 2 project, which gather more than 600 catchments [7]. The GR5J model, which stands for Génie Rural Journalier à 5 Paramètres (5-parameter daily rural engineering model), is a widely used reference model in catchment hydrology modeling due to its simplicity and flexibility. GR5J parameters are regionalized with different algorithms, including traditional spatial proximity and catchment similarity methods as well as a machine learning method based on random forest regression [8]. A direct assessment of flood statistics is also performed with random forest regression as a benchmark for flood prediction.

The models are evaluated according to global criteria as well as hydrological signatures representative of flood hazards. The hydrological characteristics of the catchments are analyzed to identify favorable and unfavorable conditions for regionalization.

This study discusses the prospects offered by deep learning for hydrological regionalization and its future integration into operational applications such as hydrological projection.

 

[1] Guo et al. (2020). https://doi.org/10.1002/wat2.1487

[2] Kratzert et al. (2019). https://doi.org/10.5194/hess-23-5089-2019

[3] Hashemi et al. (2022). https://doi.org/10.5194/hess-26-5793-2022

[4] Puche et al. (2026, in review). http://dx.doi.org/10.2139/ssrn.5286855

[5] Perrin et al. (2003). https://doi.org/10.1016/S0022-1694(03)00225-7

[6] Le Moine (2008). https://hal.inrae.fr/tel-02591478v1

[7] Sauquet et al. (2025). https://doi.org/10.5194/egusphere-2025-1788

[8] Saadi et al. (2019). https://doi.org/10.3390/w11081540