Heldmyer et al. (2026) Predicting streamflow drought in the conterminous United States using machine learning and a donor-gage approach, 1982–2020

Identification

• Journal: Hydrology and earth system sciences
• Year: 2026
• Date: 2026-09-23
• Authors: Aaron J. Heldmyer, Roy Sando, Caelan E. Simeone, Michael Wieczorek, Scott Hamshaw, Phillip Goodling, Ryan McShane, Jeremy Diaz, David Watkins, Bryce A. Pulver, Apoorva Shastry, Konrad C. Hafen, John Hammond
• DOI: 10.5194/hess-30-5925-2026

Research Groups

• U.S. Geological Survey (Wyoming-Montana Water Science Center, Oregon Water Science Center, Maryland-Delaware-D.C. Water Science Center, Water Mission Area, Utah Water Science Center, Idaho Water Science Center)
• Department of Civil and Environmental Engineering, University of Waterloo

Short Summary

This study uses machine learning and a novel donor-gage approach to predict streamflow drought across the conterminous United States (CONUS) from 1982 to 2020. It identifies key meteorological drivers and physiographic characteristics influencing drought propagation, demonstrating that the donor-gage method provides comparable prediction performance to at-site models, particularly in ungaged locations.

Objective

• What are the common hydrometeorologic drivers of streamflow drought in the CONUS?
• What are the defining physiographic characteristics of basins most sensitive to influential drivers of streamflow drought?
• Can we use dynamic regionalization (i.e., donor gages) to predict daily streamflow drought at ungaged locations?

Study Configuration

• Spatial Scale: Conterminous United States (CONUS), covering 3198 basins monitored by U.S. Geological Survey (USGS) streamgages.
• Temporal Scale: Daily streamflow drought prediction for the period 1982–2020, with models trained on 1985–2015 data and tested on 1982–1984 and 2016–2020 data.

Methodology and Data

• Models used:
  • Random Forest classification models (3198 individual models).
  • Principal Component Analysis (PCA) for dimension reduction of variable importance.
  • Linear Regression to model principal components using basin characteristics.
  • Novel dynamic regionalization (donor-gage) approach for prediction at pseudo-ungaged locations.
• Data sources:
  • Daily streamflow data from 3198 USGS GAGES-II streamgages (1981–2020 climate years).
  • Climatic variables: Atlantic Multidecadal Oscillation (AMO), Pacific Decadal Oscillation (PDO), El Niño–Southern Oscillation (ENSO), Pacific–North American pattern (PNA), temperature (minimum, maximum), potential evapotranspiration (PET), precipitation, snow-water equivalent (SWE), soil moisture (10–40 cm and 40–100 cm depths), Standardized Precipitation Evapotranspiration Index (SPEI), sunspot data, and decimal date. These were transformed to percentiles and smoothed using 30, 90, and 365-day rolling windows.
  • Physiographic basin characteristics: 58 variables from the GAGES-II dataset, including climate, hydrologic catchment, land cover, soil, and topographic characteristics.

Main Results

• At-site Random Forest models achieved a mean Cohen's Kappa score of 0.42 (standard deviation, σ = 0.22) across all 3198 sites, with higher performance in the Central, Southeast, Northeast, Pacific coast, and central Southwest regions.
• The most important meteorological drivers for streamflow drought across the CONUS were soil moisture (especially at 10–40 cm and 40–100 cm depths), precipitation, and SPEI.
• Regional differences in drivers were observed: teleconnections, temperature, evaporative demand (PET), and snow-water equivalent (SWE) were important in the West, Southwest, and Northern Rocky Mountains, while precipitation and soil moisture were primary drivers in the Northeast, Southeast, and Northwest.
• Principal Component Analysis (PCA) of variable importances revealed distinct regional groupings of drought drivers:
  • PC1 (explaining ~25% variance) was dominated by teleconnections, temperature, and PET, influencing the western CONUS and parts of the Southeast and Northeast.
  • PC2 (explaining ~20% variance) was characterized by moisture-related metrics (precipitation, SPEI, soil moisture), showing a clear east-west split.
  • PC3 (explaining ~15% variance) was largely associated with soil moisture and SWE, distinguishing snow-dominated from baseflow-driven systems.
• The novel donor-gage prediction method achieved a mean Kappa score of 0.40 (σ = 0.22) across all sites, demonstrating comparable performance to at-site models (mean difference of 0.02 lower). It generally outperformed low-Kappa at-site models and underperformed high-Kappa at-site models, suggesting a regularization effect from ensemble averaging.

Contributions

• Developed and validated a machine learning-based methodology, including Random Forest models and a dynamic regionalization approach using donor gages, for predicting streamflow drought across the CONUS.
• Provided new insights into the spatiotemporal distribution of influential streamflow drought drivers by analyzing Random Forest variable importances and their clustering through PCA.
• Identified distinct, regionally coherent groupings of meteorological drivers (e.g., teleconnections/temperature/PET in the West vs. precipitation/soil moisture in the East) and their connection to basin physiography.
• Demonstrated the applicability and comparable performance of the donor-gage method for streamflow drought prediction at pseudo-ungaged locations, offering a valuable tool for water resource management in data-sparse areas.
• Enhanced the interpretability of machine learning models by allowing for the analysis and manipulation of specific donor gages, providing crucial information on the relative importance of model parameters.

Funding

• U.S. Geological Survey (USGS) Water Availability and Use Science Program
• Water Resources Mission Area Data-Driven Drought Prediction Project

Citation

@article{Heldmyer2026Predicting,
  author = {Heldmyer, Aaron J. and Sando, Roy and Simeone, Caelan E. and Wieczorek, Michael and Hamshaw, Scott and Goodling, Phillip and McShane, Ryan and Diaz, Jeremy and Watkins, David and Pulver, Bryce A. and Shastry, Apoorva and Hafen, Konrad C. and Hammond, John},
  title = {Predicting streamflow drought in the conterminous United States using machine learning and a donor-gage approach, 1982–2020},
  journal = {Hydrology and earth system sciences},
  year = {2026},
  doi = {10.5194/hess-30-5925-2026},
  url = {https://doi.org/10.5194/hess-30-5925-2026}
}