Matthews et al. (2025) Error-correction across gauged and ungauged locations: A data assimilation-inspired approach to post-processing river discharge forecasts

Identification

• Journal: Hydrology and earth system sciences
• Year: 2025
• Date: 2025-11-11
• Authors: Gwyneth Matthews, Hannah Cloke, Sarah L. Dance, Christel Prudhomme
• DOI: 10.5194/hess-29-6157-2025

Research Groups

• Department of Meteorology, University of Reading, United Kingdom
• European Centre for Medium-range Weather Forecasts (ECMWF), United Kingdom
• Department of Geography and Environmental Science, University of Reading, United Kingdom
• Department of Mathematics and Statistics, University of Reading, United Kingdom
• National Centre for Earth Observation (NCEO), United Kingdom

Short Summary

This study presents a novel data-assimilation-inspired post-processing method to error-correct river discharge ensemble forecasts across both gauged and ungauged locations. The method successfully improves the skill of the ensemble mean and maintains spatial and temporal consistency in the corrected forecasts.

Objective

• To present and evaluate a novel technique for spreading observation information from gauged to ungauged locations in a computationally efficient and temporally varying manner within a post-processing environment.
• To determine if data assimilation techniques can be adapted for post-processing to propagate observational information to ungauged locations.
• To assess if the resulting error-corrected ensemble predictions of river discharge are more skillful than the raw ensemble.

Study Configuration

• Spatial Scale: Entire Rhine-Meuse catchment, with a drainage area of 195 300 km², a channel length of approximately 38 370 km, and consisting of 7812 grid boxes at a 5 km × 5 km spatial resolution.
• Temporal Scale: Daily timesteps for hindcasts from 1 January 2021 to 31 December 2021, with a maximum lead-time of 15 days. Daily river discharge observations from 21 December 2020 to 15 January 2022 were used.

Methodology and Data

• Models used:
  • Data Assimilation: Local Ensemble Transform Kalman Filter (LETKF) with state augmentation.
  • Hydrological Model: LISFLOOD (used in the European Flood Awareness System - EFAS).
• Data sources:
  • River discharge hindcasts: Copernicus Emergency Management Service’s European Flood Awareness System (EFAS) version 4 (51-member ensemble, 5 km × 5 km resolution).
  • Meteorological forcings: 51-member medium-range ensemble from the European Centre for Medium-range Weather Forecasts (ECMWF).
  • Observations: Daily in-situ river discharge observations from 89 stations in the Rhine-Meuse catchment (for assimilation and validation in leave-one-out experiments). Additional observations from 505 stations across Europe were used for initial error ensemble generation.
  • Ancillary data: LISFLOOD static and parameter maps (local drainage direction, channel length).

Main Results

• The proposed method successfully propagates error information along the river network, enabling error correction at every grid box while maintaining spatial and temporal consistency.
• The skill of the ensemble mean is improved at almost all locations, including stations both upstream and downstream of assimilated observations.
• The error-corrected ensemble means show a stronger correlation with observations, with an average increase from 0.82 to 0.92.
• Over half of the stations (47 out of 89) show improvement in mean bias, and the method generally improves the underestimation of flow variability seen in raw hindcasts.
• The Normalised Root Mean Square Error (N-RMSE) is reduced for most stations, though 14 stations, typically in upstream reaches, show reduced skill.
• The raw hindcast ensemble loses skill more quickly than the error-corrected ensemble, particularly for lead-times longer than 5 days.
• At short lead-times (1 day), the raw hindcast ensemble is underdispersed; the error-corrected ensemble shows slight improvement but remains overconfident.
• At longer lead-times (beyond 7 days), the error-corrected ensemble tends to be overdispersed, leading to an under-confident forecast.
• The error-adjustment of a single forecast for the entire Rhine-Meuse catchment took an average of 8.5 minutes, suggesting potential for operationalization with parallelization.

Contributions

• Introduces a novel, computationally efficient, and temporally varying data-assimilation-inspired post-processing technique for error-correcting river discharge forecasts at both gauged and ungauged locations.
• Adapts data assimilation techniques (state augmentation, LETKF) for a post-processing environment, eliminating the need for computationally expensive additional hydrological model runs.
• Develops new techniques for localization, covariance inflation, and initial error ensemble generation that are transferable across different modeling systems and river catchments.
• Provides a method that successfully propagates observational error information along the river network, addressing a critical challenge in hydrological forecasting for ungauged areas.
• Offers potential for improved post-event analysis and future development into operational forecast post-processing systems, providing more accurate knowledge of future river states.

Funding

• National Centre for Earth Observation (grant nos. NE/X019063/1, NE/W004984/1, and NE/Y006216/1)
• Natural Environment Research Council (grant no. NE/S015590/1)
• Engineering and Physical Sciences Research Council (grant no. EP/R513301/1)
• European Centre for Medium-range Weather Forecasts

Citation

@article{Matthews2025Errorcorrection,
  author = {Matthews, Gwyneth and Cloke, Hannah and Dance, Sarah L. and Prudhomme, Christel},
  title = {Error-correction across gauged and ungauged locations: A data assimilation-inspired approach to post-processing river discharge forecasts},
  journal = {Hydrology and earth system sciences},
  year = {2025},
  doi = {10.5194/hess-29-6157-2025},
  url = {https://doi.org/10.5194/hess-29-6157-2025}
}