Zheng Duan

Dissolved Organic Carbon (DOC) is a key component of the inland water carbon cycle, playing a critical role in water quality, aquatic ecosystem dynamics, and carbon flux assessments. In this study, we combine Sentinel-2 Multi-Spectral Instrument (MSI) data with in-situ water quality, meteorological, and temporal variables to predict DOC concentrations in Lake Erken, Sweden. Three Machine Learning (ML) models, eXtreme Gradient Boosting (XGBoost), Random Forest Regression (RFR), and Gaussian Process Regression (GPR), are evaluated. XGBoost achieved the best performance with a MAPE of 3.46% and RMSE of 0.45 mg C/L. To enhance interpretability, we applied SHapley Additive exPlanations (SHAP) analysis, which identified that the year (Y), total phosphorus (TP), total nitrogen (TN), and Sentinel-2 Band 2 (B2) were the most influential predictors of DOC. These results demonstrate the potential of integrating remote sensing and ML techniques for accurate DOC prediction in inland waters, thereby improving our understanding of the carbon cycle and providing insights to support sustainable water management.