Satellite-based spatial-statistical modeling of daily stream water temperatures at the CONUS scale

Stream temperature (ST) is a key driver of water quality and ecosystem health, and the analysis and forecasting thereof benefit from the availability of high spatiotemporal resolution ST datasets. However, such datasets are limited spatially and temporally across the CONUS, particularly for small, remote streams. Available models are limited in domain (regional), spatial resolution ( ≥ 10 km), temporal resolution (monthly), or a combination thereof. We address these limitations by developing a satellite remote sensing-based spatial-statistical model, TempEst 2 (“stream TEMPerature ESTimation, version 2”), to estimate daily mean and maximum temperatures at 1 km resolution for ungaged streams of any size across the CONUS. This contrasts with an earlier version, TempEst 1, which used a random forest algorithm at monthly resolution. TempEst 2 also improves over TempEst 1 in interpretability and computational efficiency. In TempEst 2, the streams used for model training and testing cover a wide range of urban and rural land uses, climates, and geographies. TempEst 2 requires minimal input data and tolerates sparse training gage networks, supporting generalization outside the CONUS. For CONUS applications, TempEst 2 is trained on 1,316 USGS ST gages. Quantitative performance impacts from locally-sparse gage networks are reported. Model performance is evaluated using cross-validation, walk-forward validation, and extrapolation validations over regions and elevation. TempEst 2 estimates daily mean temperatures with a median cross-validation gage RMSE of 2.0 °C, NSE of 0.91, and bias of 0.10%. TempEst 2 supports the development and analysis of high-resolution, large-domain ST datasets across the CONUS. • A satellite-driven statistical model was developed for daily stream temperature. • Spatial statistics are used to estimate stream seasonality and weather response. • Model performance is sustained across stream sizes and ecoregions. • Model performance is consistent for urban and less-altered sites. • Model components are interpretable to support further analysis.