Understanding water quality in coastal regions like the Bay of Bengal is becoming increasingly important, especially due to rising pressures from pollution, river runoff, and climate-driven changes. Chromophoric Dissolved Organic Matter (CDOM) is a key indicator of such changes, but tracking it accurately is difficult in these waters because of high turbidity and the coarse resolution of many satellite products. This study presents a new approach for estimating CDOM concentrations at a 300-m resolution by integrating Sentinel-3 OLCI satellite data with machine learning models optimized through Bayesian techniques. The model is trained using overlapping datasets from MODIS (1 km) and Sentinel-3 (300 m) covering the period from 2017 to 2024. We extract a range of inputs, including spectral bands (2–9, 11), to capture both spatial and temporal variability in turbid coastal waters. Three ensemble algorithms (Decision Tree, Random Forest, XGBoost) are tested and optimized their performance. SHAP analysis is used to interpret each band’s importance to predict CDOM. Results demonstrate that a Bayesian-optimized Random Forest model achieved the highest predictive accuracy (Testing R2 = 0.85, RMSE = 0.025), with Band 6 (560 nm) identified as the most significant predictor. The resulting high-resolution maps successfully delineate fine-scale CDOM patterns, particularly the outflow of organic matter from the Sundarbans mangrove forest. It demonstrates the framework’s utility for advanced coastal water quality monitoring and management. In addition, the framework’s transferability could advance satellite-derived water quality monitoring in similar tropical coastal systems.