USGS: Hot spots and cold snaps: Daily stream temperature data across the US from 1979-2021 ~ MAVEN'S NOTEBOOK

Snapshot of modeled stream temperature data from across the lower 48 United States on May 25, 2020. Data from https://doi.org/10.1016/j.envsoft.2025.106655

USGS scientists recently published four decades of modeled daily stream temperature values for nearly 60,000 river reaches across the lower 48 United States

From the USGS:

Temperature is a key water quality parameter that affects a range of stream functions, including fish habitat, cooling capacity for power plant operations, and risk of harmful algal bloom outbreaks. Despite its importance, temperature observations are historically sparse in both space and time.

Scientists at the USGS have drastically expanded the availability of stream temperature estimates across the U.S. by combining USGS and other agency data across the country. These data were used to create a first-of-its-kind dataset with daily stream temperature values modeled across 57,810 stream reaches from 1979-2021.

Machine Learning in Action

Producing this large set of stream temperature predictions was made possible by adapting methods for modeling social media and online shopping networks to U.S. stream reaches. These methods allowed USGS scientists to leverage otherwise patchy stream temperature readings to generate modeled data that are more complete in space and time. The models used observational data compiled from regional scales and expand the coverage of modeled stream temperature data to cover the lower 48 United States. The observed and modeled temperature data are publicly available on ScienceBase.

Stream temperature is driven by air temperature, but it is also affected by a variety of natural and human factors. For example, groundwater inputs like springs or seeps generally cool streams in the summer and warm streams in the winter. Built structures like dams or reservoirs also affect water temperature and are further complicated by periodic water releases to downstream areas. Other factors, like canopy cover and disturbances like wildfires and urbanization, can alter the relationship between air and stream temperatures, further complicating modeling efforts.

180 degree panorama photograph, showing water, rock formations, a partly cloudy sky. The sun is setting on the right hand side. — Human-managed water systems such as the Flaming Gorge Reservoir, Wyoming’s largest manmade reservoir, affect stream temperature downstream.

The machine learning methods used to create this new record of modeled temperature data provide a flexible approach where the models can adaptively learn relationships between stream temperature and natural or human drivers. The authors tested the accuracy of the model results by comparing the modeled data against observed stream temperatures in areas below dams, at sites with high groundwater inputs, and downstream of thermoelectric power plants.

Building on History

photo of a fixed-mount thermal infrared camera installed on the USGS Platte Kill at Dunraven NY streamgage. — Photo of a fixed-mount thermal infrared camera installed on the USGS Platte Kill at Dunraven NY streamgage in the Neversink River watershed. The camera is positioned to track focused groundwater discharge along a bedrock contact on the opposite side of the river. This technology is part of the USGS Next Generation Water Observing System.

The USGS has a record of producing cutting-edge modeled water data for the nation. This current stream temperature modeling work started in the Delaware River Basin, where temperature monitoring expanded by the Next Generation Water Observing System (NGWOS) program enabled temperature forecasts for reservoir operators who actively manage cold water habitat for a prized trout fishery.

The national stream temperature modeling effort built on previous university collaborations and regional model in the Delaware River basin, which explored an approach called process-guided deep learning (PGDL), which the USGS developed in collaboration with university partners. PGDL can improve the accuracy, applicability, and trustworthiness of deep learning models by reducing data requirements and ensuring predictions are physically consistent with known processes. PGDL has been used by the USGS to advance predictions of temperature in groundwater-dominated systems, dissolved oxygen in streams, lake temperature across the Upper Midwest, and nitrate in groundwater.

Another USGS effort, EcoSHEDS brought together temperature records from multiple local and federal agencies into a multi-decade dataset of temperature observations across the country. The observational data compiled by EcoSHEDS drastically improved the ability to use data-intensive models and evaluate model performance from local to national scales.

These ongoing investments keep USGS science at the cutting edge, highlight the value of our data, and help manage economically important natural resources.

Read the paper about this new dataset and learn more about modeling stream temperature through our interactive data visualization website.