For additional information please see "Use of Landsat for Managing Coastal Ecosystems of the USVI" https://www.ncei.noaa.gov/data/oceans/coris/data/NOAA/nesdis/ora/LandChangeAnalysis/Documents/sapnabatishDissFINAL.pdf
The results were more conclusive for St. Croix than for St. John, which were smaller in size and had a greater amount of cloud cover than St. Croix.
The classification accuracy was improved by removing all areas greater than approximately 9 m in depth from the classification. This was due to the fact that, in clear water, Landsat TM and ETM+ sensors are unable to discriminate spectral signatures of objects greater than approximately 9 m in depth.
Using the visible shallow water extent as a guide, each of the georectified shallow water areas and atolls were clipped from the Landsat ETM+ imagery. In another set of scenes, all water areas were clipped, leaving only terrestrial areas. They were atmospherically corrected, and using unsupervised classification, land use and change maps were created for the benthic and terrestrial portions of each scene.
Because water and land have spectral characteristics very different from one another, the benthic and terrestrial portions of each satellite image were classified separately. For a complete description of the processing steps please see: https://www.ncei.noaa.gov/data/oceans/coris/data/NOAA/nesdis/ora/LandChangeAnalysis/Documents/sapnabatishDissFINAL.pdf