Unsupervised texture segmentation for multispectral remote-sensing images

An unsupervised texture segmentation approach for multispectral remote-sensing images is proposed. Firstly, a scale-space filter (SSF) based histogram thresholding is used to threshold each spectrum space of a multispectral remote-sensing image to detect the major clusters of the multispectral data to generate the principal multispectrum set. Secondly, a GMRF (Gaussian Markov random field) is used to model the multispectral texture image, then the global GMRF parameters in a posteriori distribution probability are estimated. We label each pixel of the image based on the principal multispectrum set and the global GMRF parameters to maximize a posteriori distribution probability (MAP). Thirdly, a uniformity criterion is presented to each pixel in the global segmented image to determine whether it should be estimated the local MRF parameters or not. A max-min distance clustering method is then used to cluster the estimated local MRF parameters to further segment the image. Several remote-sensing images were processed by the proposed approach to demonstrate the segmentation ability