Exploring the Potential of Sentinel-2 and SVR for Predicting Non-OAC Nutrients (TP, NO_x) in the Gulf of Izmit

Monitoring non-optically active constituents (non-OACs), such as total phosphorus (TP) and oxidized nitrogen (NO_x), is critical for effective assessment of coastal eutrophication. However, these constituents present persistent challenges for remote sensing approaches due to their weak or indirect optical signatures. The Gulf of Izmit, a semi-enclosed and ecologically stressed coastal system in the Sea of Marmara, experiences ongoing eutrophication, yet systematic satellite-based monitoring of TP and NO_x remains largely unexplored in this region. This study investigates the feasibility of integrating Sentinel-2 imagery with in situ water quality measurements, applying a Support Vector Regression (SVR) algorithm to predict and map TP and NO_x concentrations across multiple seasons and years (April 2019, October 2022, December 2023). All satellite images were atmospherically corrected with the C2RCC processor, and SVR models were trained on the merged 2022–2023 data and validated independently on the 2019 dataset. The SVR approach yielded limited-to-moderate predictive performance for TP (validation R = 0.16, RMSE = 8.5 µg/L, MAE = 6.3 µg/L) and limited predictive performance for NO_x. The relatively low performance for NO_x and, to a lesser extent for TP, may be partly attributed to the temporal and seasonal differences between the model development and validation datasets, together with the limited sample size available for model development and evaluation. The independent 2019 validation dataset represented spring conditions that were not included in the autumn/winter model development data. Although the overall error rates were relatively high, the broad spatial patterns predicted by the models were generally consistent with in situ observations, particularly in identifying zones of nutrient enrichment across the sampling periods. Analysis of model feature importance identified chlorophyll-a and total suspended matter as influential predictor variables for both TP and NO_x prediction. As one of the first studies in Turkish coastal waters, this work highlights both the limitations and the potential of satellite-based prediction of non-OAC nutrients, and provides guidance for improving future methods and expanding applications in coastal water quality monitoring.