Discovering Localized River Parameters via Physics-Guided Machine Learning and the Muskingum-Cunge Method
Abstract
Deep learning-based rainfall-runoff models have consistently outperformed physics-based models and shown the ability to generate subbasin runoff with high accuracy. However, subbasin-level long short-term memory (LSTM) models do not scale properly to large rivers. The heterogeneity of inputs (meteorological forcings and geophysical attributes) is large enough that the standard assumption of uniform inputs along the entire river reach is no longer reasonable. The physics-based Muskingum-Cunge method uses a river network, a connected sequence of edges and nodes containing localized river attributes which we call a river graph, to model in-channel routing. Combining this process-driven method with our recently-developed deep learning framework for parameter learning enables routing of observed surface runoff and subsurface flow through the river network while also discovering individualized river segment parameters. Such parameters enable us to better understand the physical processes occurring in and around these rivers, and work to improve predictions including scalable flood routing.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H35S1255B