Diaz et al. (2025) Evaluation of daily stream temperature predictions (1979–2021) across the contiguous United States using a spatiotemporal aware machine learning algorithm

Identification

• Journal: Environmental Modelling & Software
• Year: 2025
• Date: 2025-08-19
• Authors: Jeremy Diaz, Samantha K. Oliver, Galen Gorski
• DOI: 10.1016/j.envsoft.2025.106655

Research Groups

• U.S. Geological Survey, Water Mission Area, Reston, VA, USA
• U.S. Geological Survey, Upper Midwest Science Center, Madison, WI, USA

Short Summary

This study developed and evaluated a recurrent graph convolution network to predict daily minimum, mean, and maximum stream temperatures across over 50,000 stream reaches in the contiguous United States (CONUS) for 42 years (1979–2021). The model achieved satisfactory performance with reach-level root mean square errors (RMSE) below 2 °C and robust uncertainty quantification, providing the most spatially complete stream temperature modeling to date for water availability assessments.

Objective

• To generate consistent, complete, and accurate daily predictions of minimum, mean, and maximum stream temperature across CONUS river networks (approximately 58,000 stream segments) using recurrent graph convolution networks (GNNs) that leverage both temporal and hydrologically relevant spatial information.
• To comprehensively evaluate model performance across all reaches and by stream type (e.g., reservoir, groundwater influence, atmospheric), and assess its ability to capture ecologically relevant metrics (e.g., magnitude and timing of peak temperatures, long-term changes).

Study Configuration

• Spatial Scale: Contiguous United States (CONUS), covering 57,810 river segments (reaches) from the National Hydrologic Geospatial Fabric version 1.1 (NHGFv1.1).
• Temporal Scale: 42 years, from 1 October 1979 to 31 December 2021, providing daily predictions.

Methodology and Data

• Models used:
  • Recurrent Graph Convolution Network (RGCN) combining Long Short-Term Memory (LSTM) for time series and Graph Neural Networks (GNNs) for spatial relationships.
  • Cluster Graph Convolutional Network (Cluster-GCN) for efficient training on large networks.
  • Simple linear model (baseline) for comparison.
• Data sources:
  • Stream Temperature Observations: Multi-agency database (Oliver et al., 2024) comprising approximately 13.2 million daily observations across 26,404 stream reaches.
  • River Network: National Hydrologic Geospatial Fabric version 1.1 (NHGFv1.1).
  • Meteorological Data: GridMET gridded daily dataset (Abatzoglou, 2013) for daily minimum and maximum air temperature, relative humidity, and precipitation.
  • Stream Characteristics: Static physical attributes from NHGFv1.1 including elevation, slope, and width.
  • Infrastructure Data: Thermoelectric-power water use data (Galanter et al., 2023) and metrics describing decadal reservoir storage intensity and degree of river regulation (Wieczorek et al., 2021) derived from the National Inventory of Dams (USACE, 2018).

Main Results

• Overall Performance: The model achieved median site-level RMSEs for the test period of 2.05 °C for daily minimum, 1.98 °C for daily mean, and 2.20 °C for daily maximum stream temperature. Nash-Sutcliffe Efficiency (NSE) values ranged from 0.90 to 0.92.
• Uncertainty Quantification: The 90% prediction intervals (PI90) captured 90.7% of observations overall, indicating a slight overestimation of uncertainty. PI90 width generally increased from 5 °C to over 8 °C as observed temperatures rose from 0 °C to over 30 °C.
• Generalizability: The model demonstrated good generalizability to unseen times and locations, with no statistically significant difference in RMSE between gaged (seen in training) and ungaged (not seen in training) reaches during the test period.
• Performance by Stream Type: Performance varied by stream type:
  • Best at atmospherically dominated sites (median RMSE ≤ 1.8 °C).
  • Worse at sites influenced by upstream reservoirs (median RMSE = 2.47 °C), deep groundwater (median RMSE = 2.18 °C), and thermoelectric plants (median RMSE = 2.50 °C).
• Groundwater Influence: The model struggled to accurately classify groundwater-influenced sites (11–13% recall) and did not adequately capture large phase lags between air and water temperature characteristic of shallow groundwater sites. It showed a tendency for positive bias in summer and negative bias in winter at groundwater sites.
• Ecological Metrics: The model accurately predicted average annual daily mean temperature (Tannual, RMSE < 1.7 °C, R² > 0.67). However, it struggled with the timing of the warmest 7-day period (DOYTmax7DMA, R² < 0.23) and the annual number of days exceeding 30 °C (R² < 0.53).
• Long-term Trends: The model consistently underestimated the observed rate of change in water temperature, although both predictions and observations indicated predominantly increasing temperature trends.

Contributions

• First CONUS-wide predictions of daily minimum, mean, and maximum stream temperatures for approximately 58,000 stream reaches, significantly enhancing the integration of water quality parameters into national water availability assessments.
• Provides the most spatially complete stream temperature modeling to date, serving as a baseline for future improvements.
• Comprehensive evaluation of model performance across diverse stream types and ecologically relevant metrics, including robust uncertainty quantification.
• Demonstrated the capability of recurrent graph convolution networks to generalize well in both space and time for large-scale hydrological predictions despite using a limited set of input features.

Funding

• U.S. Geological Survey (USGS) Water Mission Area’s Water Availability and Use Science Program
• Predictive Understanding of Multiscale Processes project
• Integrated Water Prediction Program

Citation

@article{Diaz2025Evaluation,
  author = {Diaz, Jeremy and Oliver, Samantha K. and Gorski, Galen},
  title = {Evaluation of daily stream temperature predictions (1979–2021) across the contiguous United States using a spatiotemporal aware machine learning algorithm},
  journal = {Environmental Modelling & Software},
  year = {2025},
  doi = {10.1016/j.envsoft.2025.106655},
  url = {https://doi.org/10.1016/j.envsoft.2025.106655}
}