Segment based classification of Indian urban environment

Abstract

This paper presents results of segment based classification of an Indian urban environment. This approach to classification involved three stages. In the first stage, a region based multispectral segmentation of the image was carried out after determining suitable automatic threshold values considering textured nature of imagery. The second stage involved refinement of initially segmented image, iteratively by merging smaller segments with the most similar adjacent segments until they satisfied a homogeneity criterion. Finally, these segments were classified into 12 different classes using various spectral and textural properties of segments. Three different types of classifications were performed: the per-pixel Gaussian maximum likelihood classification (GMLC), per-segment GML classification, and the per-segment neural classification. Result showed that per-segment classification improves overall classification accuracy by more than 25% in comparison to per-pixel approach.