A graph-based machine learning framework for river water quality management under data limitations.

Area of Science:

• Environmental Science
• Water Resource Management
• Machine Learning Applications

Background:

• Accurate riverine water quality prediction is challenged by sparse data and limited streamflow information, common in resource-constrained watershed monitoring.
• Existing methods often struggle to integrate diverse hydro-environmental variables effectively for robust forecasting.

Purpose of the Study:

• To develop and validate a novel three-module machine learning framework for riverine water quality prediction, interpretation, and management.
• To apply this framework to chromaticity prediction in the Hantan River Basin, addressing data limitations.

Main Methods:

• A three-module framework combining graph neural networks (GNNs) or recurrent networks for prediction, explainable AI for interpretation, and counterfactual analysis for management.
• Utilized a dataset of 1667 monthly observations from 59 sites covering 37 hydro-environmental variables.
• Employed independent training, validation, and testing sets for rigorous performance assessment.

Main Results:

• Graph-based models, particularly the enhanced Graph Sample-and-Aggregate, outperformed recurrent baselines, achieving a test R² of 0.82.
• Interpretability analyses identified the SC sub-watershed as a primary intervention region and distinguished long-term from short-term pollution drivers.
• Counterfactual analysis demonstrated feasible downstream chromaticity targets (14-15 CU) with success rates of 26-40%.

Conclusions:

• The proposed machine learning framework significantly enhances riverine water quality prediction accuracy and interpretability.
• It provides a cost-effective, decision-support tool for watershed management, especially under data-limited conditions.
• The study highlights the effectiveness of GNNs in capturing pollution source characteristics and transport pathways.