Sawada et al. (2026) Technical note: Benefits of Bayesian estimation of model parameters in a large hydrological model ensemble

Identification

• Journal: Hydrology and earth system sciences
• Year: 2026
• Date: 2026-11-02
• Authors: Yohei Sawada, Shinichi Okugawa
• DOI: 10.5194/hess-30-6737-2026

Research Groups

• Department of Civil Engineering, Graduate School of Engineering, The University of Tokyo, Tokyo, Japan

Short Summary

This study investigates the benefits of Bayesian estimation for quantifying parametric uncertainty within a large hydrological model ensemble. It finds that Bayesian parameter estimation significantly improves individual model performance, especially for complex models, thereby influencing the interpretation of model structural comparisons. However, it offers only marginal improvements to multi-model ensemble hydrological predictions, suggesting that the high computational cost of Bayesian methods is often not justified for practical forecasting when structural diversity is present.

Objective

• To clarify how Bayesian quantification of parametric uncertainty benefits large hydrological model ensemble-based assessment of structural uncertainty.
• To examine if Bayesian estimation of model parameters affects the interpretation of model comparisons (scientific benefit).
• To assess if Bayesian estimation of model parameters improves the accuracy of rainfall-runoff simulations in a large hydrological model ensemble and if this improvement justifies the high computational cost (practical benefit).

Study Configuration

• Spatial Scale: 51 river basins in Japan.
• Temporal Scale: Study period: 1986–2015. Calibration periods: 5 years (1986-1990) and 1 year (1986). Validation periods: 25 years (1991–2015) or 24 years (1992–2015) depending on calibration period shifts. Spin-up period: 5 years.

Methodology and Data

• Models used:
  • MARRMoT v1.3 (Modular Assessment of Rainfall-Runoff Models Toolbox)
  • 17 lumped conceptual hydrological models (IDs 02, 03, 04, 06, 07, 10, 11, 12, 13, 17, 18, 21, 24, 27, 30, 31)
  • Parameter estimation: Nelder-Mead algorithm (deterministic optimization), DiffeRential Evolution Adaptive Metropolis (DREAM) algorithm (MCMC-based Bayesian optimization) using Kling-Gupta Efficiency (KGE) as an informal likelihood function.
  • Model averaging: Bayesian Model Averaging (BMA), Bayesian Parameter Averaging (BPA), Bayesian Model and Parameter Averaging (BMPA).
• Data sources:
  • Meteorological forcings: Multi-model Ensemble for Robust Verification of hydrological modeling in Japan (MERV-Jp) dataset (daily total precipitation, daily mean temperature, daily mean potential evapotranspiration).
  • Daily river discharge data from MERV-Jp.
  • Performance metrics: Kling-Gupta Efficiency (KGE), Nash-Sutcliffe Efficiency (NSE).

Main Results

• Bayesian estimation of model parameters (BPA) systematically outperforms deterministic optimization for individual hydrological models, with KGE improvements exceeding 0.2 in some cases, particularly under data-scarce conditions (1-year calibration).
• More complex models with a larger number of parameters tend to benefit more significantly from Bayesian estimation, suggesting that previous evaluations of model complexity based on deterministic optimization might have underestimated their potential performance.
• Bayesian parametric uncertainty quantification influences the interpretation of structural uncertainty assessment, highlighting that the choice of calibration method can affect conclusions about appropriate model complexity.
• The overall improvement in multi-model ensemble hydrological predictions from BMA to BMPA is marginal (KGE improvement less than 0.1, NSE less than 0.05), even when individual models show substantial gains from Bayesian estimation.
• The high computational cost of MCMC-based Bayesian parameter estimation (e.g., 85 hours for 17 models in a single basin) is not justified for improving multi-model ensemble predictions, as the diversity of model structures effectively mitigates the adverse effects of suboptimal deterministic calibration.

Contributions

• First verification of the systematic superiority of MCMC-based Bayesian parameter optimization over deterministic methods within a large hydrological model ensemble.
• Demonstrates that Bayesian parametric uncertainty quantification significantly impacts the assessment of model structural uncertainty, particularly for complex models, challenging previous interpretations based solely on deterministic optimization.
• Provides a clear trade-off analysis, showing that while Bayesian estimation improves individual model performance, its practical benefits for multi-model ensemble predictions are limited, suggesting that structural diversity can compensate for parametric uncertainty.

Funding

• Foundation of River and basin Integrated CommunicationS (FRICS) grant
• JST Moonshot R&D Program (grant no. JPMJMS2281)
• JSPS KAKENHI (grant nos. 25H00760, 23K20967)
• Joint Usage and Research of JHPCN (grant nos. jh230003, jh240064, jh250035, jh260026)
• Computation performed on the Wisteria/BDEC-01 supercomputer in the University of Tokyo.

Citation

@article{Sawada2026Technical,
  author = {Sawada, Yohei and Okugawa, Shinichi},
  title = {Technical note: Benefits of Bayesian estimation of model parameters in a large hydrological model ensemble},
  journal = {Hydrology and earth system sciences},
  year = {2026},
  doi = {10.5194/hess-30-6737-2026},
  url = {https://doi.org/10.5194/hess-30-6737-2026}
}