Satellite-derived bathymetry using machine learning and optimal Sentinel-2 imagery in South-West Florida coastal waters

This study examines the use of the Multi-Spectral Instrument (MSI) in Sentinel-2 satellite in combination with regression-based random forest models to estimate bathymetry along the extended southwestern Florida nearshore region. In this study, we focused on the development of a framework leading to a generalized Satellite-Derived Bathymetry (SDB) model applicable to an extensive and diversified coastal region (>200 km of coastline) utilizing multi-date images. The model calibration and validation were done using airborne lidar bathymetry (ALB). As ALB surveys are very expensive to conduct, the proposed model was trained with a limited and practically feasible ALB data sample to expand the model's practicality. Out of the three different sub-models introduced using varying combinations of historical satellite imagery, the combined-band model with the largest feature pool yielded the highest accuracy. The results showed root mean square error (RMSE) values of 8% and lower for the 0-13.5 m depth range (limit of the lidar surveys used) for all areas of interest, indicating the model efficiency and adaptability to varying coastal characteristics. The influence of training sample locations on model performance was evaluated using three distinct model configurations. The difference between these configurations was less than 5 cm, which highlights the robustness of the proposed SDB model. The quality of the satellite imagery is a significant factor that influences the accuracy of the bathymetry estimation. A preliminary methodology incorporating spectral data embedded in Sentinel-2 imagery to effectively select the most optimal satellite imagery was also proposed in this study.