This thesis investigates the concept of a spectral class and its relevance to remotely sensed imagery. Two study areas near Sydney were selected, and cluster analysis was performed on these areas using four clustering programs: the IBM ERMAN system, the CSIRO ORSER system, the DIPIX Image Analysis System and a newly written program CLUST on the University of N.S.W. Cyber computer. The clustering results of these four systems were compared to assess their performance and determine the extent to

more »

... hich spectral classes exist. It was found that the significance of spectral classes depends on the nature of the individual subimage. Water pixels are usually concentrated closely around a class distinctly separate from land pixels. However, the land areas of an image usually occupy a very broad region of spectral space, with very little natural separation into distinct clusters. This is probably because most pixels contain a mixture of landcover types. The iterative clustering method using Euclidean distance was found to be the best of those studied. Single pass methods were inadequate because an insufficient sample of the minor classes existed to generate signatures in one pass. Histogram programs had the disadvantage that they could not distinguish the major peaks of the histogram from subsidiary or small peaks. The description of the ORSER clustering program in this thesis is based upon a summary contained in reference 26. A more detailed treatment, which should be consulted by an intending user is given in reference 40.