Preprints
https://doi.org/10.5194/egusphere-2026-6443
https://doi.org/10.5194/egusphere-2026-6443
22 Nov 2026
 | 22 Nov 2026
Status: this preprint is open for discussion and under review for Hydrology and Earth System Sciences (HESS).

Can we rely on deep-learning hydrological models to assess the quality of precipitation products?

Mohamed Saadi, Pierre Brigode, Marc Dumont, Ludovic Oudin, and Hélène Roux

Abstract. Process-based models are commonly used to evaluate the accuracy of precipitation products because they are highly sensitive to errors in precipitation due to mass-balance constraints. In this respect, it is unclear whether to rely on deep-learning models to discriminate between precipitation products, since their larger flexibility may allow them to compensate for the errors in input precipitation. Here, we address this question by comparing a 1-km, radar-based reanalysis (COMEPHORE) against (1) a gauge-based topography-aware product (SAFRAN, 8 km) and (2) a global, model-based reanalysis (ERA5-Land, ~9 km). Over a set of 445 French catchments, we assessed the quality of each precipitation product using three hydrological models: a process-based conceptual model (GR4J), a locally trained, and a regionally trained long short-term memory (LSTM) models. To assess the reliability of the evaluation by LSTM models, we retrained and tested the three hydrological models on shifted and randomly corrupted precipitation inputs to analyze the extent to which LSTM models mobilized their flexibility relatively to GR4J to compensate for the introduced errors. Our key results are: (1) The best model performances were obtained with the radar-based reanalysis COMEPHORE, with comparable performances for SAFRAN and a large deterioration with ERA5-Land, highlighting the need for gauge-based correction. (2) All models ranked the precipitation products similarly. In particular, the flexibility of LSTM models was not sufficient to compensate for the poorer quality of ERA5-Land. (3) All models responded similarly to unrealistic precipitation inputs. However, LSTM models mobilized their flexibility (larger parameter space, absence of mass-balance constraints) to limit the drop in their performances under highly corrupted precipitation inputs. These counterfactual experiments shed new light on the mechanisms by which LSTM-based models may outperform traditional process-based models, and call for efforts to improve precipitation estimates for hydrological applications.

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.
Share
Mohamed Saadi, Pierre Brigode, Marc Dumont, Ludovic Oudin, and Hélène Roux

Status: open (until 03 Jan 2027)

Comment types: AC – author | RC – referee | CC – community | EC – editor | CEC – chief editor | : Report abuse
Mohamed Saadi, Pierre Brigode, Marc Dumont, Ludovic Oudin, and Hélène Roux
Mohamed Saadi, Pierre Brigode, Marc Dumont, Ludovic Oudin, and Hélène Roux
Metrics will be available soon.
Latest update: 22 Nov 2026
Download
Short summary
Deep-learning rainfall-runoff models were rarely (or never) used for the specific task of evaluating precipitation products, perhaps due to their larger ability in compensating for the errors in precipitation inputs, making them less selective when it comes to precipitation quality. Here we show that the flexibility of deep-learning models is not sufficient to compensate for the poor quality of precipitation inputs, highlighting the dominance of the choice of precipitation product.
Share