Predicting Stream Temperature Across Spatial Scales With Low Complexity ML

Stream water temperature is an important water quality parameter because it is critical for maintaining healthy ecosystems and for devising sound watershed management strategies. Water temperature regulates chemical and biological processes in streams and rivers, and is projected to be negatively impacted by climate change and extreme events. Accurate stream temperature predictions are essential for making optimal watershed management decisions, but are limited by data availability and the ability to integrate new data from different sources into models. Here we use low-complexity machine learning models (Multiple Linear Regression, Support Vector Regression, and Random Forest Regression/XGBoost) to predict monthly stream water temperature across spatial scales (point to watershed to regional) using data from 75 stations in the Mid-Atlantic and Pacific Northwest hydrologic regions. The models utilize a data integration tool, BASIN-3D to get access to the latest data for a range of input meteorological forcings, streamflow, water temperature and station metadata in a standardized format. Models are trained at the point scale, and at larger spatial scales using a generalized version of the models with additional station attributes. Performance for both the point scale (RMSE typically <1 degree C) and generalized model (RMSE typically <1.5 degree C) is good for stations without significant human impacts. To further improve performance of the generalized models, we additionally build separate models for stations in dammed and undammed basins, but find no significant improvement using this classification. Lastly, we validate these models using stations with low data availability and demonstrate how stations with well monitored historical records can be used to predict at undersampled or ungauged locations. Results support that low complexity models are able to generalize dynamics impacting stream temperature across spatial scales using solely meteorology, discharge, and simplistic station attributes. These simplistic models allow for stream temperature predictions at stations with minimal data and selection requirements: thereby expanding stream temperature prediction capabilities which are essential for proper water management.