Segmentation of multispectral remote-sensing images based on Markov random fields

An unsupervised approach for texture segmentation of multispectral remote-sensing images based on Gaussian Markov random fields (GMRFs) is proposed. At first, the authors treat the false-color information of SPOT satellite images as RGB attributes and then transform them to HSI attributes. Secondly, a scale-space filter is used to threshold the hue histogram to quantize the color set which represents the principal color components in the original image. Thirdly, the global GMRF parameters are estimated from the original image for global segmentation. Fourthly, they label each pixel of the image based on the quantized color set and the GMRF parameters to maximize a posterior color distribution probability to achieve the global segmentation. Fifthly, a criterion is used to judge whether every pixel in the global-segmented image is within a local textured region or not. Finally, the pixels in a local textured region are further estimated the local GMRF parameters and clustered based on the parameters. Seven SPOT images were segmented to demonstrate the ability of the proposed approach. Moreover, the scale-space filter, the MRF-based global segmentation, and the pure local (texture) parameter classification are sequentially evaluated their performance.