Applying Graph Neural Networks to Improve the Data Resolution of Stream Water Quality Monitoring Networks

Session Number

Project ID: CMPS 25

Advisor(s)

Kaize Ding, Northwestern University Department of Statistics

Discipline

Computer Science

Start Date

17-4-2024 8:35 AM

End Date

17-4-2024 8:50 AM

Abstract

In most US watersheds, surface water quality observations are scarce, making it challenging to assess goals, advise management, and calibrate high-resolution models. Popular statistical techniques, such as USGS's LOADEST and WRTDS, estimate daily pollution load using regression methods and the link between flow and pollutant concentrations. However, they do not consider upstream-downstream relationships. We suggest using Graph Neural Network (GNN) to improve spatial and temporal resolution of water quality data, considering the river network connections and the spatial-temporal inputs simultaneously. Using a physical flow direction graph, the GNN builds up numerical relationship edges among river monitoring stations. It then wraps up the time steps using spatiotemporal encoders. After imputations with multiple solvers, the GNN decodes the spatial connections and results in continuous monitoring data at each station. We test this method in the Maumee River Basin because of its prolonged records on river water quantity and quality.

Share

COinS
 
Apr 17th, 8:35 AM Apr 17th, 8:50 AM

Applying Graph Neural Networks to Improve the Data Resolution of Stream Water Quality Monitoring Networks

In most US watersheds, surface water quality observations are scarce, making it challenging to assess goals, advise management, and calibrate high-resolution models. Popular statistical techniques, such as USGS's LOADEST and WRTDS, estimate daily pollution load using regression methods and the link between flow and pollutant concentrations. However, they do not consider upstream-downstream relationships. We suggest using Graph Neural Network (GNN) to improve spatial and temporal resolution of water quality data, considering the river network connections and the spatial-temporal inputs simultaneously. Using a physical flow direction graph, the GNN builds up numerical relationship edges among river monitoring stations. It then wraps up the time steps using spatiotemporal encoders. After imputations with multiple solvers, the GNN decodes the spatial connections and results in continuous monitoring data at each station. We test this method in the Maumee River Basin because of its prolonged records on river water quantity and quality.