Satellite image classification using genetically guided fuzzy clustering with spatial information

Land-cover classification of satellite images is an important task in analysis of remote sensing imagery. Segmentation is one of the widely used techniques in this regard. One of the important approaches for segmentation of an image is by clustering the pixels in the spectral domain, where pixels that share some common spectral property are put in the same group, or cluster. However, such spectral clustering completely ignores the spatial information contained in the pixels, which is often an important consideration for good segmentation of images. Moreover, the clustering algorithms often provide locally optimal solutions. In this paper, we propose to perform image segmentation by a genetically guided unsupervised fuzzy clustering technique where some spatial information of the pixels is incorporated. Two ways of incorporating spatial information are suggested. The characteristic of this technique is that it is able to determine automatically the appropriate number of clusters without making any assumptions regarding the dataset, while attempting to provide globally near-optimal solutions. In order to evolve the appropriate number of clusters, the chromosome encoding scheme is enhanced to incorporate the don't care symbol (#). Real-coded genetic algorithm with appropriately defined operators is used. A cluster validity index is used as a measure of the fitness value of the chromosomes. Results, both quantitative and qualitative, are demonstrated for several images, including a satellite image of a part of the city of Mumbai.