Segmentation of radar imagery using the Gaussian Markov random field model

Abstract

Segmentation of highly speckled radar imagery is achieved by the use of the Gaussian Markov random field model. The model considers two regions to be different if one or more than one of the following conditions is true: (1) their first-order statistics are different, (2) their second-order statistics are different and (3) their spatial textures are different. It proves that the optimal model parameters can be obtained via solving linear algebraic equations for single-channel images and linear iterations for multi-channel images. Techniques of wavelet transforms and watershed process are used to obtain initial segmentation. Various grey and colour examples, including synthetic radar images, air-borne and space-borne synthetic aperture radar images, are tested, showing the accuracy and efficiency of the method. Regions whose mean differences are as small as 0.5dB and ratios of the standard deviation to the mean as high as 0.35, are separated with an accuracy of more than 95%. It would require their mean differences to be as large as 7.5dB to separate such speckled regions if only pixel values were used in segmentation.