Gaussian Mixture Model and Classification of Polarimetric Features for SAR Images

Aiming at the various statistical characteristics such as peak tailing presented in the high-resolution synthetic aperture radar (SAR) images, we model the polarimetric features according to the Gaussian mixture model (GMM) and come up with a constrained distance estimation algorithm for the parameters of multivariate Gaussian mixture distribution. Under the framework of greedy expectation maximum algorithm, a constraint distance function is designed and the number of mixed components and model parameters arc automatically estimated in this parameter estimation algorithm. Consequently the classification of polarimetric SAR images is realized under the Bayesian framework. The classification results of three groups of image data from Radarsat-2 in San Francisco and other places indicate that the proposed GMM classification algorithm possesses an overall accuracy higher by 7% -10% comparing with those by the classical classification algorithms. Moreover, its dependence on sample number is small. The more accurate classification results can be obtained in heterogeneous regions such as urban and farmland.