Contextual classification with IRS LISS-II imagery

Abstract

A non-linear probabilistic relaxation contextual algorithm has been used to improve the classification results of the Gaussian maximum likelihood (GML) classifier. The research has been conducted at two study sites. The first is the Singrauli mines area in the Sidhi district and the second is the Bhopal city area, both in the state of Madhya Pradesh (India). IRS LISS-II data, with a spatial resolution of 36.25 m and radiometric resolution of 7 bits per pixel, have been used for the research. The GML classification has been performed with bands 1 (0.45–0.52 μm), 3 (0.62–0.68 μm), and 4 (0.77–0.86 μm). The overall classification accuracy for the 6-class GML classification for the Singrauli site was 90.70%. For the Bhopal site, the 5-class GML classification accuracy was 91.46%. When the non-linear probabilistic relaxation contextual algorithm with two different forms of compatibility coefficients was applied to this initial classification from the GML classifier, the overall classification accuracies after 20 iterations improved to 94.36 and 94.58% for compatibility coefficients based on correlation and mutual information, respectively, for the Singrauli site. For the Bhopal site the improvements were 94.98 and 92.39%, respectively. The overall classification accuracies with contextual information with both forms of compatibility coefficient were found to be statistically significantly different from and better than the GML classification for the Singrauli site. However, for the Bhopal site this was true only when the correlation based compatibility coefficient was used.