This manuscript presents the use of hyperspectral in situ reflectance measurements evaluated above the water surface to train a supervised classification of a simultaneously acquired Landsat 5 TM image. The optical signature of a submerged reef substrate is both attenuated and augmented by the presence of the atmosphere and intervening water column, thereby complicating the link between field and space-borne measurement. The motivation of this manuscript is to quantify the advantage (defined as increase in classification accuracy) conferred by each of five levels of increasingly complex image processing methods to correct for atmospheric and submergence effects. It is found that the maximum overall classification accuracy attained in areas of the image where water depth is unknown was 53%, but with the addition of depth information, accuracy increases to 76%. The results demonstrate the ability of a classifier trained solely by in situ optical measurements to accurately resolve the broad substrate types of a typical Red Sea fringing reef top. Although slightly different accuracy assessment protocols were used, the results also suggest that using in situ spectra for training provided better classification results than using from-image statistics. © Springer-Verlag 2003.