SWOT-based Simulation of River Discharge with Temporal Graph Neural Networks
Keywords: Graph Machine Learning, Graph Neural Networks, Streamflow, Rivers, SWOT
TL;DR: Using graph machine learning to simulate river flow with the latest data from the SWOT mission and other data sources
Abstract: The Surface Water and Ocean Topography (SWOT) satellite launched in December 2022 and has been providing high-accuracy monitoring of rivers' water surface elevation. Remote measurements do not, however, include discharge, an essential element for improved modelling of river dynamics. Furthermore, orbital characteristics of SWOT satellite overpasses result in spatial and temporal discontinuity of observations. The process of inferring hydrological information from one river portion to another is called regionalization and may help alleviate this issue. While most regionalization methods based on machine learning do not leverage geometrical river properties, recent works point to graph machine learning as a promising solution for simulation of river flow. This work applies state-of-the-art temporal graph machine learning techniques for regionalization and discharge estimation. More specifically, we leverage available SWOT data, in-situ gauge discharge and a temporal graph neural network to offer basin-scale simulation of river discharge in US rivers where the SWOT mission operates. We run experiments in different basins and against several in-situ gauges on which the proposed method offers convincing performance. The approach is compared to the drainage-area ratio method, a non-temporal graph neural network and a long short-term memory model, which it outperforms.
Submission Number: 6
Loading