Rosenberg Institute Seminar Series - Laurel Larsen

Laurel Larsen, Associate Professor/Delta Lead Scientist, UC Berkeley/Delta Stewardship Council

Can a machine become a hydrologist? Integrating physical understanding with deep learning to develop practical and trustworthy hydrologic forecasts

Abstract

With watersheds experiencing unprecedented shifts in climate and hydrology, developing more accurate forecasts of streamflow has become paramount. Legacy forecasting systems used by public agencies do not harness the power of deep learning models because they are perceived as black boxes. However, more recently, hydrologists and computational scientists have been developing hybrid approaches to integrating physical understanding with deep learning to harness the latter’s predictive power while infusing it with physical grounding. In this talk I dive into recent work by the Environmental Systems Dynamics Laboratory to develop deep-learning and hybrid forecasts for streamflow and nitrate across a range of watersheds in the US. Results yield insight into watershed characteristics that result in better or worse predictivity under different forecasting strategies. We find that generally, combining physical understanding with deep learning produces more generalizable, physically consistent models with performance similar to that of pure deep-learning models.