Daily predictions of water temperature for streams across the contiguous United States (1979-2021)

This model application data release provides the data processing and model code used to generate predictions of daily stream water temperature across the contiguous United States from 1979-2021. We used a recurrent graph convolutional network (RGCN) algorithm to make daily stream temperature predictions. Stream water temperature observations, along with forcing data consisting of daily meteorological information, a stream distance matrix, and static stream characteristics were used to predict daily stream temperature summaries (minimum, mean, and maximum) for 57,810 stream segments across the contiguous United States. This model application data release is organized as follows:

data_processing_code.zip contains the instructions and code needed to assemble inputs to the model. This directory contains a README.txt file that describes all major processing steps and outputs of this code.
model_code.zip contains code to process the outputs from data_processing_code.zip into model-ready data structures and implements the modeling algorithm. This directory contains a README.txt file that describes all model-ready input files, major processing steps, and an overview of the modeling steps.
national_temperature_metadata.xml describes the top-level files contained in this model application data release (model outputs and supporting reach-level metadata).
The model outputs are contained in a Parquet database, where chunks of data are stored in regional and subregional (HUC2 and HUC4) nested folders titled huc2={HUC2 ID}.zip}. Each HUC2 can be downloaded separately.
data_access_pattern.R gives an example of how to extract and use the stream temperature predictions in this data release.
reach_metadata.csv contains reach-level metadata that describes how the reach was used in the model (training or testing) and how the reach was classified (groundwater, atmospheric, reservoir, thermoelectric) for evaluation purposes.


The methods and results from this modeling effort are described in: Diaz, J., Oliver, S.K., Gorski, G. 2025. Evaluation of daily stream temperature predictions across the contiguous United States using a spatiotemporal aware machine learning algorithm. Environmental Modelling & Software, https://doi.org/10.1016/j.envsoft.2025.106655.