A segmentation approach to classification of remote sensing imagery

In this paper we propose a new approach far land cover classification of remote sensing imagery. It is a two-stage technique, where in the first stage the global feature-based technique of histogram thresholding generalized to multidimensional cases developed by Khotanzad and Bouarfa (1990) is used, and in the second stage a local feature-based region growing technique is generalized to grow multiple non-contiguous regions in parallel. The Khotanzad and Bouarfa technique has the advantage of being a non-iterative unsupervised classification technique, but suffers from a failure to detect regions of small spatial extent which may have high local contrast but low weightage in the global feature space. Our proposed technique divides the image into blocks of suitable size so that regions of small spatial extent are detected in the block's histogram, and they are grown across neighboring blocks. The proposed technique is illustrated with actual remote sensing imagery. A number of choices of feature space for the first stage, and different measures of similarity for the second stage were investigated on remote sensing data, both visually as well as quantitatively in terms of classification accuracy. It was found that the xyz colour space (Ohta et al. 1980) for the first stage, and the J-M distance for the second stage similarity measure, gave the best results in terms of classification accuracy. Though the approach is unsupervised and non-iterative in nature, it has given a classification accuracy of better than 91 per cent for a five-class landcover classification.