Unsupervised segmentation of multitemporal interferometric SAR images

Abstract

This paper shows how to segment large data sets of multitemporal and interferometric SAR images using an unsupervised, fuzzy clustering method. An adaptive feature extraction (principal component transformation) is employed which may drastically reduce the number of images and improves the final results. This also speeds up the fuzzy clustering iteration part considerably. The method is applied to data over two areas in Sweden: one typical urban area with forest and farmland surroundings and a forested area. The best classification accuracy is obtained when classifying the data into two classes, agreeing with the predictions of the cluster validity parameters used in this study. The method always finds the dominating land-covers in the images first. These are then subdivided as more clusters (classes) are identified, indicating that the segmentation is moderately hierarchical. The final classification results, between 65% and 75%, are comparable to those obtained in other studies. Analyzing the final cluster signatures reveals that the current unsupervised method has several similarities with rule-based methods.