Boardman et al. (2025) Improving model calibrations in a changing world: controlling for nonstationarity after mega disturbance reduces hydrological uncertainty

Identification

• Journal: Hydrology and earth system sciences
• Year: 2025
• Date: 2025-11-17
• Authors: Elijah N. Boardman, Gabrielle Boisramé, Mark S. Wigmosta, Robert K. Shriver, A. A. Harpold
• DOI: 10.5194/hess-29-6333-2025

Research Groups

• Graduate Program of Hydrologic Sciences, University of Nevada, Reno, USA
• Mountain Hydrology LLC, Reno, Nevada, USA
• Hydrologic Sciences Division, Desert Research Institute, Reno, Nevada, USA
• Energy and Environment Directorate, Pacific Northwest National Laboratory, Richland, Washington, USA
• Department of Natural Resources and Environmental Science, University of Nevada, Reno, USA

Short Summary

This study demonstrates how equifinality in hydrological models leads to significant uncertainty in predicting post-disturbance streamflow changes after a megafire. It introduces a novel metamodel framework that controls for non-stationary model error (bias shift) to drastically reduce this uncertainty, providing more robust estimates of hydrological responses to environmental disturbances.

Objective

• To investigate how calibration equifinality impacts process-based simulations of the hydrological response to a megafire.
• To determine if equifinality and predictive uncertainty can be reduced by evaluating the model's representation of hydrological change through quantifying stationarity across pre- and post-disturbance periods.

Study Configuration

• Spatial Scale: Upper San Joaquin River Basin, California, USA, covering 4244 km² with an elevation range of 100 m to 4200 m. Model simulations were performed at a 90 m resolution. The Creek Fire burned 1540 km² within this area.
• Temporal Scale: Water years 2005–2024 (20 years total), with calibration focused on 2015–2024 (6 years pre-fire, 4 years post-fire) and validation on 2005–2014. Meteorological data were disaggregated to a 3-hour timestep. Post-fire streamflow changes were analyzed over a 4-year period.

Methodology and Data

• Models used:
  • Distributed Hydrology Soil Vegetation Model (DHSVM) for process-based hydrological simulations.
  • Statistical metamodel (Bayesian linear regression) trained on DHSVM outputs to correct for bias shift and reduce uncertainty.
  • Empirical annual water balance model (Bayesian multiple linear regression) for comparison.
• Data sources:
  • Streamflow: Reconstructed unimpaired daily outflows at Millerton Lake (California DWR, 2024).
  • Burn Severity: Landsat-based data from Monitoring Trends in Burn Severity (MTBS Project, 2022).
  • Vegetation Properties: RCMAP data (Rigge et al., 2021a, b) for fractional cover, Landfire data (2022) for vegetation classification, NLCD (Dewitz and U.S. Geological Survey, 2019) for abiotic land surface classes. Leaf Area Index (LAI) estimated empirically and refined by calibration.
  • Subsurface Properties: Regional soil survey databases (Gupta et al., 2022; Soil Survey Staff, 2022) disaggregated using Random Forest models.
  • Channel Geometry: National Hydrography Dataset (U.S. Geological Survey, 2019).
  • Meteorological Data: gridMET (Abatzoglou, 2013) disaggregated using MetSim (Bennett et al., 2020).
  • Snow Water Equivalent (SWE): Airborne Snow Observatory (ASO) data (Painter et al., 2016) for spatial distribution and total volume.
  • Calibration: Multi-objective Bayesian optimization using 14 sensitive parameters (meteorology, vegetation, subsurface, snowpack dynamics) and 7 objective functions (daily streamflow NSE, log-scaled NSE, >95th-percentile RMSE, yearly MAPE, April–July MAPE, pixel-wise SWE RMSE, total SWE volume MAPE).

Main Results

• Equifinality in water balance partitioning leads to substantial uncertainty in post-fire streamflow predictions, with different model calibrations yielding up to a six-fold variation in the 4-year post-fire streamflow change (e.g., 13 to 97 mm/year, or a 1400% range in dry years).
• Controlling for non-stationary model error (bias shift) significantly (p < 0.01) reduces both equifinality and predictive uncertainty. A strong correlation (r = 0.96–0.99) exists between the mean streamflow bias shift and the annual streamflow change attributable to fire.
• The statistical metamodel framework, correcting for bias shift, reduces uncertainty in the post-fire streamflow change by 80% compared to pure statistical model ensembles and 82% compared to pure process-based model ensembles.
• The estimated streamflow increase in the first four years after the Creek Fire is 11% ± 1% (90% credible interval). During drought conditions (2021), the increase was 18% ± 4%.
• In absolute terms, the additional streamflow attributable to the fire was approximately 0.11 ± 0.03 km³ in 2021 and 0.38 ± 0.04 km³ in the wet year of 2023.
• Selecting a "stationary sub-ensemble" (parameter sets with near-stationary bias) significantly reduces uncertainty in Leaf Area Index (LAI) by 74% and log-transformed transmissivity by 50%.
• Blindly trusting highest-performing parameter sets based on traditional goodness-of-fit metrics (e.g., NSE) can be misleading; the highest-NSE parameter set underestimated the post-fire streamflow change by 79% relative to the metamodel mean.

Contributions

• Demonstrates that equifinality in hydrological models, particularly concerning water balance partitioning, is a critical and often unconstrained source of uncertainty in simulating environmental disturbances like megafires.
• Introduces a novel and robust framework for reducing model equifinality and predictive uncertainty by leveraging the concept of non-stationary model error (bias shift) after disturbance.
• Proposes a statistical metamodel approach that fuses process-based model outputs with observational constraints on bias stationarity, enabling more confident and statistically justified uncertainty quantification for hydrological changes.
• Highlights the "reward hacking" pitfall in high-dimensional model calibration, cautioning against over-reliance on single goodness-of-fit metrics and advocating for evaluation against independent metrics like bias stationarity.
• Provides a generally applicable conceptual framework for improving hydrological model simulations of various types of environmental change by ensuring model stationarity with respect to the investigated variability.

Funding

• National Science Foundation Graduate Research Fellowship Program (grant no. 1937966)
• National Science Foundation Division of Earth Sciences (grant nos. EAR 2011346 and EAR 2012310)
• National Science Foundation Office of Integrative Activities (grant no. OIA 2148788)

Citation

@article{Boardman2025Improving,
  author = {Boardman, Elijah N. and Boisramé, Gabrielle and Wigmosta, Mark S. and Shriver, Robert K. and Harpold, A. A.},
  title = {Improving model calibrations in a changing world: controlling for nonstationarity after mega disturbance reduces hydrological uncertainty},
  journal = {Hydrology and earth system sciences},
  year = {2025},
  doi = {10.5194/hess-29-6333-2025},
  url = {https://doi.org/10.5194/hess-29-6333-2025}
}