Decadal basin-wide nitrogen recovery and localized estuarine deterioration in the Bohai Sea revealed by physics-guided spatially adaptive machine learning

Abstract

Excessive dissolved inorganic nitrogen (DIN) drives coastal eutrophication in semi-enclosed seas, yet retrieving this non-optically active nutrient in optically complex waters remains difficult, limiting decadal assessments of management progress. Here we reconstruct surface DIN across the Bohai Sea from 2013 to 2023 using a physics-guided, spatially adaptive machine-learning framework (GW-PGML-XGBoost) trained on 4,409 in-situ match-ups. The reconstruction reveals a decadal divergence: most of the basin recovered, while localized estuarine zones deteriorated. On the independent test set, the framework reproduced observations with R2 = 0.73 and RMSE = 0.094 mg L−1. Ablation attributed this skill to the coupled spatial and physical constraints, which raised R2 from 0.58 for standalone XGBoost. Prediction uncertainty, dominated by observational noise rather than model structure, was propagated into the trends, confirming that the divergence is robust to retrieval error. Sea surface salinity, sea surface height, and water depth were the dominant predictors, reflecting physical mixing and terrestrial loading. The climatology shows a persistent nearshore-high to offshore-low gradient and a seasonal cycle of summer minima and autumn–winter maxima. DIN declined significantly across the basin, tracking reductions in sewage, atmospheric, and agricultural nitrogen inputs. In Bohai Bay, the decline reached (−2.11 ± 1.17) × 10−3 mg L−1 yr−1 (p < 0.01), the steepest in the study area. Laizhou Bay, the most DIN-rich subregion, held steady at a high baseline (p = 0.39). In contrast, the localized estuarine increases coincided with intensified riverine inputs. Nutrient management in semi-enclosed seas should therefore shift from uniform annual-load reduction toward spatially precise, high-flow-season strategies.