Colored dissolved organic matter estimation using Sentinel-2 imagery in small-scale reservoir: Classification and...

Imbalanced datasets represent a significant challenge in accurately estimating Colored Dissolved Organic Matter (CDOM) concentrations in aquatic environments. This study integrates machine learning models with synthetic data generated via classification- and regression-based techniques. Five input variables-Band4/Band5, Band5/Band1, Band5/Band3, Band2/Band5, and Band1/Band2-were inferred from the reflectance of Sentinel-2 imagery using the HSIC-Lasso model, and these were combined with observed CDOM data from Namyang Reservoir. Based on the CDOM distribution, a 5 m^-1 threshold was used to classify data, maintaining a 3:1 class ratio in all datasets. Ensemble models such as Random Forest (RFR), eXtreme Gradient Boosintg (XGB) showed stable performance across oversampling techniques. In contrast, Support Vector Machine (SVR) and Deep Neural Network (DNN) exhibited marked improvements in prediction accuracy, particularly when combined with SMOTE-SVM and SmoteR techniques. Notably, DNN with SMOTE-SVM and DNN with SmoteR achieved the highest prediction accuracies among the classification-based and regression-based models, with accuracies of 0.88 and 0.89, respectively, significantly reducing errors in high-concentration regions. Spatial distribution analysis further emphasized the effectiveness of classification-based oversampling techniques in addressing data imbalances, with DNN with SMOTE-SVM and DNN with SMOTE-Tomek providing the most accurate predictions at high-concentration boundaries. This study highlights the potential of synthetic data and machine learning in water quality monitoring. The proposed framework provides a scalable and robust approach for estimating CDOM concentrations, offering valuable insights for environmental management and policymaking.