On the concept of spectral class

Abstract Classification of remote sensing image data requires the multispectral domain to be segmented into distinct regions called spectral classes. These are related to the geometric character of the classifier and consequently have multidimensional spectral shapes that are different for parallelepiped, minimum distance and maximum likelihood classification. This is particularly significant for regions of multispectral space that are substantially continua with few discernible density maxima, as is frequently the case in practice.