The automated classification of remotely sensed data is an important research topic, as more and more such data is becoming available for scientific, military, and commercial purposes. This paper describes a new algorithm to automatically segment multiband satellite images into seven predefined classes.
The main idea of the proposed method is to split the problem into two parts. The first part breaks the image into a set of small patches, and segments each patch individually. The second part agglomerates segmented patches into a segmentation map of the whole image. The segmentation of a small patch is a simpler problem than the segmentation of the whole image, as it can be assumed that the number of classes coexisting within the patch is small, and each region is contiguous.
Unfortunately, it is difficult to evaluate the effectiveness of the technique, as the experiments are designed and described poorly. The authors, without any justification, use normal distributions to model each region in a synthetic test image. Natural images often have much more complex and spatially correlated patterns. Thus, a technique that works well on a simplistic model often does not work on natural images. A table showing the results is not well presented, and I could not interpret the meaning of some of the entries. The authors propose a variance estimation technique that only uses pixels around the boundary of a patch, and report more accurate results than the one that uses all of the pixels in the patch. The claim is suspicious, since estimating variance with a small number of samples is a dangerous practice, due to a high standard error for such an estimate.