Graph-based machine learning approach for river water quality prediction under data limitations

Accurate prediction of river water quality is essential for effective watershed management, yet it is often hindered by practical monitoring constraints, including infrequent grab sampling (e.g., monthly observations) and the lack of reliable streamflow data. These limitations restrict the applicability of conventional process-based water-quality models and necessitate alternative analytical tools. In this study, we propose a graph-based machine learning framework that integrates prediction and diagnostic analyses of river water quality, with chromaticity prediction in the Hantan River Basin, Republic of Korea, as a case study. Graph-based models outperformed purely temporal baselines, with the Graph Sample-and-Aggregate (GraphSAGE) model achieving a test R² of 0.82. Its sampling-based spatial aggregation integrates localized and distributed upstream information across the river network, allowing the model to capture nonlinear relationships mediated by implicit flow connectivity. Graph explanation analyses using PGExplainer identify the SC sub-watershed as the dominant pollution source and primary intervention area. In addition, feature attribution analyses distinguish persistent long-term drivers (e.g., TOC associated with major wastewater treatment plant discharges) from short-term episodic influences linked to facility-specific effluent spikes. Overall, these results demonstrate that graph-based machine learning can serve as a useful framework for both prediction and diagnostic interpretation of key water-quality drivers in data-limited river systems.