Decoupling physical transport from redox coupling in coastal nutrient enrichment using explainable machine learning

Anthropogenic nutrient pollution drives coastal eutrophication, but distinguishing physical transport from internal biogeochemical processing along the land-sea continuum challenges linear models. We compared two hydrologically contrasting bays in Fujian Province that share a monsoon climate and intensive aquaculture but differ fundamentally in freshwater input: river-dominated Sansha Bay and marine-dominated Zhao'an Bay. An explainable AI (XAI) framework integrating Random Forest, SHAP, and partial dependence plots (PDP) decoupled the drivers of dissolved inorganic nitrogen (DIN), improving cross-validated R² over linear regression from 0.35 to 0.64 (RF) in Sansha Bay and from 0.20 to 0.69 (RF) in Zhao'an Bay. Using an integrated diagnostic that combined conservative-mixing analysis, SHAP driver rankings, and PDP response shapes, we classified Sansha Bay as transport-controlled: salinity was the dominant driver (28.6%) and DIN followed a continuous mixing gradient with no stable threshold. In contrast, Zhao'an Bay was reaction-controlled: salinity was the weakest driver (12.9%), while dissolved oxygen (29.3%) and dissolved reactive phosphorus (27.3%) produced non-linear thresholds and central retention hotspots. XAI thus distinguishes transport- from reaction-controlled systems without extensive sediment monitoring, and the identified thresholds indicate whether management should target external nutrient loads or internal benthic feedbacks.