Mean spectra of training samples and mean of the pixels classified by the MLH classifier. The mean spectra of the training samples (solid lines) for each class, and the mean of the pixels classified by the MLH classifier into the respective classes (dashed lines), using the 30-band subsampled image cube. The vertical bars show the 1 standard deviation of the training samples. The mean of the predicted classes departs considerably from the mean of the training samples for a number of classes on the left, indicating a poor match between the known training representatives and the predicted members of the classes. In contrast, the match is very good for the classes on the right: the training and class means are virtually indistinguishable (and therefore the dashed line of the class mean may not be easy to see). This, however, does not mean excellent classification for all classes here, because a number of them have barely more than the training pixels classified into them.