A multispectral classification system based on an alternative spectral representation is described and its performance over a full landsat thematic mapper (TM) scene evaluated. Spectral classes are represented by their spectral shape -- a vector of binary features that describes the relative values between spectral bands. An algorithm for segmenting or clustering TM data based on this representation is described. After classes have been assigned to a subset of spectral shapes within training areas, the remaining spectral shapes are classified according to their Hamming distance to those that have already been classified. The performance of the spectral shape classifier is compared to a maximum likelihood classifier over five sites that are fairly representative of the full landsat scene considered. Although the performance of the two classifiers is not significantly different within a site, the performance of the spectral shape classifier is significantly better than the maximum likelihood classifier across sites. A full-scene spectral shape classifier is then described which combines spectral signature files that associate classes with spectral shapes derived over the five sites into a single file that is used to classify the full scene. The classification accuracy of the full-scene spectral shape classifier is shown to be superior to that of a stratified maximum-likelihood classifier. The spectral shape classifier is implemented in C and is able to process an entire landsat TM scene in about one hour on a single processor SUN SPARC 10 workstation with 128 megabytes of RAM.

Date Published: 17 June 1996
PDF: 12 pages
Proc. SPIE 2758, Algorithms for Multispectral and Hyperspectral Imagery II, (17 June 1996); doi: 10.1117/12.243224

Published in SPIE Proceedings Vol. 2758:
Algorithms for Multispectral and Hyperspectral Imagery II
A. Evan Iverson, Editor(s)