Integrating anthropogenic hydrodynamics and GOCI satellite observations for water quality retrieval in highly engineered coastal systems: A case study of Saemangeum

Retrieving water quality parameters in highly engineered coastal systems remains a challenge for standard satellite algorithms due to the complex decoupling of optical properties from natural hydrodynamics. This study develops a physics-aware, data-driven framework to retrieve chlorophyll-a (Chl-a) and total phosphorus (T-P) in the Saemangeum Reservoir, a representative dike-enclosed waterbody, using decadal Geostationary Ocean Color Imager (GOCI) data (2011−2020). Unlike conventional approaches relying solely on spectral data, we explicitly integrate anthropogenic structural drivers—specifically, seawater-exchange fluxes controlled by sluice-gate operations—into the modeling architecture. We systematically evaluated 255 spectral-hydrodynamic input combinations across four architectures: Random Forest, XGBoost, Transformer, and MLP-Mixer. To address the optical complexity of T-P, which lacks direct spectral absorption features, we incorporated Total Nitrogen (T-N) as an operational proxy for nutrient loading, enabling the models to capture the non-linear covariation between nutrients and optically active constituents. Our results demonstrate that incorporating anthropogenic hydrodynamic variables significantly improves retrieval accuracy, with the MLP-Mixer achieving the highest performance (Chl-a R 2 = 0.81; T-P R 2 = 0.83) by effectively learning long-term temporal dependencies. However, spatial cross-validation revealed that ensemble tree-based models offer superior generalization in spatially heterogeneous zones, suggesting a trade-off between temporal precision and spatial robustness. Explainable AI analysis further confirmed that thermal forcing, nutrient coupling, and anthropogenic hydrodynamic variables are key contributors to water-quality variability. Ultimately, this framework demonstrates a transferable modeling concept for monitoring human-modified aquatic environments, showing that explicitly accounting for anthropogenic forcing can improve satellite-based water-quality retrieval, particularly in managed systems where operational auxiliary measurements (e.g., near-real-time nutrient proxies) are available.