Thematic Classification of Multispectral Imagery

Remote sensing systems designed to monitor the surface environment employ a multispectral design such as parallel sensor arrays thereby detecting radiation in a small number of broad wavelength bands in order to provide increased spectral discrimination. These broad-band multispectral systems allow discrimination of different types of vegetation, rocks and soils, clear and turbid water, and some man-made materials. In a multispectral image the brightness values in the different wavelength bands encode the spectral information for each image cell, and can be regarded as a spectral pattern. Spectral classification methods seek to categorize the image cells on the basis of these spectral patterns. In supervised classification the analyst designates a set of “training areas” in the image, each of which is a known surface material that represents a desired spectral class. The classification algorithm computes the average spectral pattern for each training class, then assigns the remaining image cells to the most similar class. In unsupervised classification the algorithm derives its own set of spectral classes from an arbitrary sample of the image cells before making class assignments. This paper describes the supervised classification of multispectral data obtained from LANDSAT 5 TM using Matlab Simulation using different band combinations. The specific band combination that yields better classification is the end result.