Physics-guided graph-based multimodal prediction of chlorophyll-a in river networks

Short-term prediction of chlorophyll-a (Chl-a) is essential for eutrophication management and early warning of algal blooms. Existing Chl-a prediction studies commonly rely on remotely sensed observations, process-based models, or data-driven models. However, remotely sensed observations are often discontinuous, data-driven models alone may struggle with nonlinear and network-dependent dynamics, and process-based models are limited in capturing observation-consistent, short-term Chl-a variability. To address these limitations, we develop a multimodal prediction framework based on a graph neural network (GNN) that explicitly represents the river network as a directed graph. The framework integrates (i) remotely sensed Chl-a observations and meteorological data with (ii) process-based hydrological and water-quality states that provide continuous, physically consistent information on streamflow and constituent transport along the river network. These process-based variables are generated using the Soil and Water Assessment Tool (SWAT), and provide physics-guided information that complements observation gaps and supports learning of upstream–downstream dynamics along the river network. The proposed framework is applied to Geumho River Watershed (2,092 km2), and predictive performance is evaluated using the coefficient of determination and the root mean square error. Comparative analyses are conducted with an RNN model and conventional machine learning models, including Random Forest and XGBoost, to assess the validity of the GNN-based approach in learning structural connectivity within river networks. This study demonstrates the applicability of a graph-based multimodal prediction framework integrating satellite observations and physics-guided hydrological information and provides a foundation for the development of intelligent early warning systems.