Water temperature drives many in-stream processes and is an indicator of both local and broad-scale changes in climate, land use, and channel modifications. Though temperature is generally inexpensive and easy to monitor, few locations have records long enough to elucidate how stream temperature might change in the future. For example, across the U.S., only 4% of ~50k stream reaches have a year or more of daily temperature records. This data sparsity limits our ability to discern what those sites might look like in the future, e.g., which of these sites are tightly coupled to atmospheric processes and will therefore change with air temperature. Two recent advances in machine learning approaches allow us to leverage sparse, broad-scale observations to make accurate predictions of stream temperature. First, we used a recurrent graph network to inform the model of stream connections and share information across the stream network. Second, we pre-trained the machine learning model on a process-based stream temperature model to produce physically consistent predictions prior to exposing the model to observations. These methods were applied to a subset of the Delaware River Basin, US, where we accurately hindcasted (RMSE = 1.40 deg C) and forecasted (RMSE = 2.03 deg C for 1-day ahead) water temperature. This presentation highlights extension of these methods to 50k stream segments across the continental US and an assessment of how well these models reproduce long-term water temperature dynamics.

Primary Presenter: Samantha Oliver, U.S. Geological Survey (oliver.samanthak@gmail.com)

Authors:

Samantha Oliver, U.S. Geological Survey  (soliver@usgs.gov)

Jeremy Diaz, U.S. Geological Survey  (jdiaz@usgs.gov)

Lauren Koenig, U.S. Geological Survey  (lkoenigsnyder@usgs.gov)

Alison Appling, U.S. Geological Survey  (aappling@usgs.gov)

Simon Topp, U.S. Geological Survey  (stopp@usgs.gov)