Computer-assisted classification of suburban areas in satellite imagery through the use of an artificial neural network

Abstract

This thesis investigates the use of backpropagation artificial neural networks for improving the computer-assisted classification of suburban areas in satellite images. The accuracy of these artificial neural network classifiers in detecting suburban areas is compred to the accuracy of a standard multispectral classifier based on Euclidean minimum distance. The satellite image used for this study is a portion of a Landsat Thematic Mapper (TM) scene covering southern Marin County, California.

An artificial neural network operating on a single-pixel basis was found to improve significantly upon the accuracy of the Euclidean minimum distance classifier. Additional significant improvements in the suburban class accuracy of the neural networks resulted from the inclusion of contextual information in the form of data from pixels in 3x3 and 5x5 windows centered on the pixel to be classified.