USEM: A Unified Model for Simultaneous Estimation of Multiple Nutrient Concentrations in Coastal Waters using Landsat 5/7/8 and Sentinel-2 imagery.

Accurate estimation of nutrient concentrations is vital for monitoring coastal eutrophication but remains challenging due to the lack of direct spectral features of nutrients and the constraints of single-sensor observations. To address these issues, this study proposed a unified simultaneous estimation model (USEM). The USEM model estimates four key nutrients-orthophosphate phosphorus (PO4P), dissolved inorganic nitrogen (DIN), total nitrogen (TN), and total phosphorus (TP)-in coastal waters simultaneously. We integrated multi-sensor satellite imagery from Landsat 5/7/8 and Sentinel-2 on the Google Earth Engine platform. A unified feature set was constructed by extracting common spectral bands across sensors and incorporating a sensor-type identifier. The eXtreme Gradient Boosting algorithm, with hyperparameters optimized via Bayesian optimization, was employed to build the USEM. The performance of USEM was compared against five other machine learning models: K-Nearest Neighbors, back-propagation neural network, support vector regression, random forest, and light gradient boosting machine. The results indicate that the USEM model showed higher accuracy, with coefficients of determination (R2) of 0.70, 0.77, 0.69, and 0.67 for PO4P, DIN, TN, and TP, respectively, on the independent test set. Additionally, it had the lowest root mean square error. The model also showed strong robustness in 10-fold cross-validation. The USEM enables long-term monitoring of coastal-water nutrient concentrations and directly supports eutrophication assessment, pollution-source tracing, and evaluation of management effectiveness.