Application of Gaussian Markov random field model to unsupervised classification in polarimetric SAR image

The aim of this paper is to demonstrate that the Gaussian Markov random field (GMRF) model can be successfully applied to the classification pf multi-frequency polarimetric SAR data. As a special case of MRF, the GMRF has been shown to be an accurate compact representation from a single-band textured images or multi-band textured images. To apply the method to the classification of inter-channel correlated polarimetric SAR data, we first transformed the data into combination of uncorrelated principal component images. Both intensities (hh, hv, and vv) and phase difference (/spl phi//sub hh - vv/) images of L- and P-band data are considered for classification in the study area in Jeju Island, South Korea. The properties of the transformed data reveal that the images tend to be Gaussian and they are mutually uncorrelated. The GMRF model therefore can be applied to the classification of the transformed polarimetric SAR data. As the GMRF model is a type of classifier based on segment merging, the classification process begins from the initial guess consisting of large amounts of segments. Spatially and statistically similar regions are combined to update the segmented map for each iteration. The final classification map based on polarimetric characteristics shows improvements in the accuracy and efficiency of the classification frame for the tested polarimetric SAR data.