Learning representations of salt marsh hydrodynamics
Abstract
High-frequency measurements of flow and transport are increasingly used to understand the health and functioning of tidal wetlands at ecosystem scales. A persistent challenge of this approach lies in extending inferences made in instrumented wetland channels to ungauged channels. Tradeoffs need to be made between instrumenting more channels, which permits one to account for spatial variability in wetland form and function, and instrumenting for longer periods, which captures both seasonal and interannual variability as well as rare events such as storms. Hydrodynamic models can fill gaps in observational records, but they also need to be parameterized and calibrated, which presents its own data requirements. Recently developed, data-driven stage-discharge models can successfully estimate discharge from water level measurements, but these models need training data from the channel of interest, preventing their use in ungauged channels. Deep neural networks can overcome these challenges by learning not the time-varying nonlinear mapping from stage to discharge but a representation of the underlying physics of flow and transport through tidal wetlands. Here, I construct neural networks to estimate flow from stage time series in ungauged salt marsh channels. I demonstrate how, by carefully choosing the architecture and objective function of recurrent neural networks, we can interpret the neural network as an effective model of salt marsh hydrodynamics. I train these models on flow and water level data collected in salt marshes of the Plum Island estuary, MA, USA, and find that they can successfully estimate the stage-discharge relationship in ungauged channels. Because these models construct an implicit hydrodynamic model from the provided water level data, they can serve as a data-driven basis for ecosystem-scale flux studies in ungauged wetlands.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2018
- Bibcode:
- 2018AGUFMEP51E1876K
- Keywords:
-
- 1824 Geomorphology: general;
- HYDROLOGYDE: 1942 Machine learning;
- INFORMATICSDE: 4217 Coastal processes;
- OCEANOGRAPHY: GENERALDE: 4558 Sediment transport;
- OCEANOGRAPHY: PHYSICAL