Mehrvand et al. (2026) A novel approach to uncertainty-conscious long-term streamflow reconstruction in ungauged basin combining machine learning and dendro-hydrological indicators

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-11-04
• Authors: Shahriar Mehrvand, Alexandre Florent Nolin, Marie‐Amélie Boucher, Kurt C. Kornelsen, Jacques Clément Tardif
• DOI: 10.1016/j.ejrh.2026.104092

Research Groups

• Département de génie civil et de génie du bâtiment, Université de Sherbrooke, Quebec, Canada
• Laboratory of Tree-Ring Research, University of Arizona, AZ, USA
• Department of Environment and Climate Change, Ontario Power Generation, Ontario, Canada
• Centre for Forest Interdisciplinary Research, University of Winnipeg, Manitoba, Canada

Short Summary

This study develops an uncertainty-aware framework combining hydrological modeling, machine learning, and dendrohydrological validation to reconstruct 74 years of natural streamflow in the regulated Abitibi River watershed. The reconstructed ensemble reveals statistically significant increases in annual and spring streamflow, with estimated trends of 1.91 m³/s/year and 2.67 m³/s/year, respectively.

Objective

• Reconstruct long-term natural streamflow in a regulated watershed using machine-learning-based parameter estimation.
• Incorporate parameter and structural uncertainty using a multi-model ensemble framework.
• Validate reconstructed streamflow using independent dendrohydrological indicators and regulated observations.
• Quantify long-term trends in natural streamflow using hierarchical Bayesian modeling.

Study Configuration

• Spatial Scale: Abitibi River watershed, northeastern Ontario and northwestern Quebec, Canada. Drainage area: 27,300 km². Elevation range: approximately 500 m.
• Temporal Scale: Daily streamflow simulations for 1950–2023 (74 years). Validation period: 1950–2017 for flood-rings, 1960–2023 for regulated streamflow.

Methodology and Data

• Models used:
  • Machine Learning: Artificial Neural Networks (ANN), Random Forest (RF), Light Gradient Boosting Machine (LGBM) for hydrological model parameter estimation.
  • Hydrological Models (within Raven platform): GR4J-CemaNeige, HBV-EC, HMETS, MOHYSE, SAC-SMA.
  • Statistical Model: Hierarchical Bayesian linear model for trend analysis.
• Data sources:
  • Meteorological Reanalysis: Environment and Climate Change Canada’s Regional Deterministic Reanalysis System (RDRS) v2.1 (2000–2018, 10 km resolution), ERA5-Land reanalysis (1950–2023, 0.1° × 0.1° resolution).
  • Dendrohydrological Indicators: Flood-ring chronologies (mean earlywood vessel area (MVA), number of earlywood vessels (N), relative frequencies F1, F2, F12) from Harricana River (HAR), Lake Duparquet (DPL), and Little Abitibi River (LAR) (1950–2017).
  • Observed Streamflow: Water Survey of Canada stations: Abitibi River at Onakawana (04ME003), Abitibi River at Island Falls (04ME001), Driftwood River at Monteith (04MB002) (all regulated); Harricana River at Amos (04NA001) (natural, since 1914).
  • Physio-climatic attributes for ML training: Database of 675 unregulated watersheds across the United States and Canada (e.g., watershed area, mean annual precipitation, altitude, slope, land cover, soil types, latitude).

Main Results

• The reconstructed 15-member ensemble of natural streamflow for 1950–2023 shows strong agreement with independent flood-ring signatures from naturally flowing rivers, with significant correlations ranging from 0.35 to 0.72.
• The ensemble successfully tracked major historical flood years (e.g., 1960, 1979, 1989, 2013, 2014) identified by flood-ring chronologies.
• Flow duration curves showed that simulated natural flows generally exceed regulated flows at high flows and exhibit wider dispersion at low flows, consistent with the absence of regulation.
• The mean Continuous Ranked Probability Score (CRPS) for daily natural streamflow simulations against regulated observations during the spring discharge period was 171.75 m³/s.
• Lagged correlation analysis between simulated natural and observed regulated daily streamflow revealed maximum Pearson correlation coefficients up to 0.790 (GR4J_ANN), with optimal lags indicating both earlier and later natural flows relative to regulated flows.
• A hierarchical Bayesian linear model revealed statistically significant increasing trends in natural streamflow from 1950 to 2023:
  • Annual streamflow: 1.91 m³/s/year (95% HDI: [0.62, 3.06] m³/s/year), representing a 4.39% increase per decade relative to the baseline mean of 435.07 m³/s.
  • Spring streamflow: 2.67 m³/s/year (95% HDI: [0.81, 4.63] m³/s/year), representing a 3.27% increase per decade relative to the spring baseline mean of 817.42 m³/s.
• Uncertainty assessment using the drainage-area ratio (DAR) scaled Harricana River flow as a proxy showed an annual CRPS of 49.39 m³/s, a daily CRPS of 130.44 m³/s, and a spring discharge CRPS of 207.17 m³/s.
• The reliability diagram indicated that the ensemble captures the uncertainty structure with reasonable reliability, though with slight underdispersion at lower confidence levels.

Contributions

• Novel application of machine learning to estimate hydrological model parameters for streamflow naturalization in regulated, data-scarce basins.
• Integration of uncertainty representation through a multi-model hydrological ensemble for long-term streamflow reconstruction.
• Independent validation of reconstructed streamflow using dendrohydrological flood-ring indicators in a regulated basin context.
• Provides a robust strategy for reconstructing long-term natural streamflow in regulated watersheds, addressing the lack of natural streamflow records.

Funding

• Canada Research Chairs (NSERC-CRC)
• Natural Sciences and Engineering Research Council of Canada Collaborative research program (NSERC-CRD grant number: CRDPJ513709-17)
• Partners: Ouranos, Hydro-Québec, Ontario Power Generation (OPG), The University of Winnipeg.

Citation

@article{Mehrvand2026novel,
  author = {Mehrvand, Shahriar and Nolin, Alexandre Florent and Boucher, Marie‐Amélie and Kornelsen, Kurt C. and Tardif, Jacques Clément},
  title = {A novel approach to uncertainty-conscious long-term streamflow reconstruction in ungauged basin combining machine learning and dendro-hydrological indicators},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.104092},
  url = {https://doi.org/10.1016/j.ejrh.2026.104092}
}