J et al. (2025) Can discharge be used to inversely correct precipitation?

Identification

• Journal: Hydrology and earth system sciences
• Year: 2025
• Date: 2025-11-11
• Authors: Ashish Manoj J, Ralf Loritz, Erwin Zehe
• DOI: 10.5194/hess-29-6115-2025

Research Groups

• Chair of Hydrology, Institute of Water and Environment (IWU), Karlsruhe Institute of Technology, Karlsruhe, Germany
• Department of Hydrology and Atmospheric Sciences, The University of Arizona, Tucson, United States

Short Summary

This study investigates the feasibility of using observed streamflow to inversely correct catchment-average precipitation from reanalysis products using LSTM networks. It demonstrates that discharge significantly improves precipitation estimates, especially for high-magnitude events, and enhances subsequent hydrological forward modeling of streamflow and soil moisture.

Objective

• To explore the feasibility of inversely correcting catchment-average precipitation time series from reanalysis products (e.g., ERA5-Land) using observed streamflow data, particularly to improve estimates for high-impact events and enhance subsequent hydrological forward modeling of streamflow and soil moisture.

Study Configuration

• Spatial Scale: Continental scale (Europe), utilizing a training dataset of 1800 catchments and testing on unseen catchments, including smaller out-of-sample catchments ranging from 50 km² to 200 km².
• Temporal Scale: Daily resolution over a training period of 25 years (1 October 1980 to 30 September 2005) and a testing period of 15 years (2006 to 2020, or 2015 for CAMELS-GB catchments).

Methodology and Data

• Models used:
  • Long Short-Term Memory (LSTM) ensemble networks (regional model for inverse precipitation prediction).
  • HBV (Hydrologiska Byråns Vattenbalansavdelning) conceptual hydrological model (for forward streamflow modeling).
  • CATFLOW process-based hydrological model (for forward soil moisture modeling).
• Data sources:
  • Caravan dataset and its community extensions (GRDC-Caravan, CAMELS-GB, Caravan Spain, Caravan Switzerland, Caravan Germany) for hydrometeorological time series and catchment attributes.
  • ERA5-Land reanalysis product (0.1° × 0.1° spatial resolution, daily aggregated temporal resolution) for meteorological forcing.
  • E-OBS gridded observational precipitation product (0.25° × 0.25° spatial resolution, daily temporal resolution) as the training target for precipitation.
  • Observed discharge data from relevant state and national authorities.
  • HydroATLAS for catchment attributes.
  • MERRA-2 reanalysis product (0.625° × 0.5° spatial resolution, daily aggregated temporal resolution) for soil moisture comparison.
  • GLDAS-2.2 reanalysis product (0.25° × 0.25° spatial resolution, daily temporal resolution) for soil moisture comparison.

Main Results

• Including discharge information in the LSTM model led to an average improvement in precipitation prediction, with a significant increase in median Nash–Sutcliffe Efficiency (NSE) of approximately 29 % on days with precipitation amounts greater than 5 mm.
• Out-of-sample tests demonstrated that the inversely estimated precipitation better reproduced small-scale, high-impact events that were poorly represented in the original reanalysis products.
• Using the inversely generated precipitation time series for classical hydrological "forward" modeling significantly improved streamflow estimates (e.g., HBV model NSE improved from 0.57 to 0.70 for the Elsenz Schwarzbach catchment) and soil moisture dynamics (increased correlation with other reanalysis products).
• At the continental scale, the LSTM model incorporating discharge reduced underestimation errors for mean wet day precipitation and the 95th percentile limit of wet days compared to the model without discharge.
• The approach successfully transferred knowledge to out-of-sample catchments, providing hydrologically consistent storm estimates and improving runoff coefficient estimations for flood events.

Contributions

• Demonstrates the novel application of deep learning (LSTM networks) to inversely assimilate streamflow data for correcting catchment-average precipitation from reanalysis products at a continental scale.
• Extends the "hydrology backwards" concept from small, well-monitored research catchments to large samples across diverse hydro-climatic conditions in Europe.
• Quantifies the significant information gain from observed discharge, particularly for high-magnitude precipitation events, highlighting its underutilised potential in meteorological applications.
• Provides evidence that inversely estimated precipitation improves the accuracy of subsequent forward hydrological modeling for both streamflow and soil moisture dynamics.
• Offers a methodology with significant implications for generating coherent long-term statistical records for catchment forcings, improving flood event design, and addressing the "Predictions in Ungauged Basins" (PUB) problem.

Funding

• Deutsche Forschungsgemeinschaft (German Research Foundation - DFG) via the project "Implementation of an InfraStructure for dAta-BasEd Learning in environmental sciences (ISABEL)" (grant no. 496155047).
• Karlsruhe Institute of Technology (KIT) for article processing charges.

Citation

@article{J2025Can,
  author = {J, Ashish Manoj and Loritz, Ralf and Gupta, Hoshin and Zehe, Erwin},
  title = {Can discharge be used to inversely correct precipitation?},
  journal = {Hydrology and earth system sciences},
  year = {2025},
  doi = {10.5194/hess-29-6115-2025},
  url = {https://doi.org/10.5194/hess-29-6115-2025}
}