Watch: Operational Streamflow Forecasting With Theory-Guided Deep Learning

By Laura Read on January 21, 2021

Watch Alden Keefe Sampson, our Technical Director, present virtually at the 2021 American Meteorological Society's Annual Hydrology Meeting on behalf of his coauthors and colleagues, David Lambl and Eliza Hale.

The full citation information is available on the AMS website.


In recent years, deep learning models have had notable success in hydrological prediction [1, 2]. As we begin using these models in operational decision-making contexts, what techniques need to be applied and what new capabilities do these models provide? We share lessons from our company’s operational forecasting with deep learning models. Deep learning models can leverage near real time satellite data on land surface conditions, making what was formerly a challenging data assimilation problem into a simple change to a deep learning model. We share how incorporating hydrologically relevant satellite data provides a measurable improvement in accuracy across hundreds of basins in the continental US. We also show how to inspect the learned input-output relationships in a deep learning model to aid forecast interpretation, validate physical realism and build users’ confidence in model predictions. Finally, we discuss methods for model training and gauge data assimilation which lead to reliable, accurate forecasting. Throughout these topics we will highlight the common thread of how physical hydrologic theory can guide the application of deep learning approaches to streamflow forecasting.