Supervised classification utilized training texels of 30 x 30 to 90 x 90 pixels cut from GeoTiff orthotiles centered on the coordinates of the grab sample stations.  Each texel was assigned to a cluster training set based on that sample’s classification in the original (latent) cluster analysis calculated on similarity of sediment characteristics. However, none of the potential 2470 combinations of backscatter signal characters and their treatments were able to discriminate significantly among these 5 classes, meaning that variation among samples of at least 2 classes overlapped considerably. Recombination into 4 classes (combining Classes 3 and 4) yielded significant discrimination. Mapping of the results showed that one of these classes was likely to be legitimate when applied to the bayside, but additionally was duplicated as an artifact of edge between orthotiles on the oceanside because of fading at the swath margins. This means that backscatter was characteristic of the larger habitat distinctions shown in the latent dendrogram with confidence, and of lesser branches with less confidence. Therefore, the entire oceanside was characterized as one habitat, and classification of the bayside was attempted again in isolation. Recombination into 3 classes (“mud”, “sand”, “gravelly sand”) was able to resolve 3 classes significantly (score  = 0.33548) using input factors Contrast, Gray Mean, and Directionality with 30 x 30 pixel (15 x 15 m) texels. Despite good separation in the training texels, with some slight overlap at the 5% confidence ellipsoid for mud and gravel, most areas known to be muddy were classified as being gravelly sand in the resulting classification map. This is likely a function of reflective shell hash in acoustically dark mud having similar contrast to reflective gravel with acoustically dark shadows created by high relief. A test of natural separation (Davies-Bouldin Index) indicated four modes using these characters, so the same factors were used in an unsupervised classification allowing four latent classes. The four latent classes mapped very similar to the previous supervised classification but broke up the latent analog to the “gravelly sand” class. Class error was low at 0.1110. The newly resolved class was clearly mud with shell, based on video ground truthing. This class was combined with the mud class in compiling the final habitat classification map.