Unsupervised PolSAR Image Classification and Segmentation Using Dirichlet Process Mixture Model and Markov Random Fields With Similarity Measure

Abstract

The Markov random fields (MRF) is skillful in incorporating the spatial-contextual information of images and has been widely applied to remote-sensing image classification and segmentation. However, the traditional MRF-based method is unable to determine the precise number of clusters automatically. It is known that the Dirichlet process mixture model (DPMM) takes the number of clusters as a model parameter and estimates it in image classification. Therefore, the DPMM is a powerful and potential method for classification tasks. Then, in this paper, by fusing the DPMM model and a similarity measure scheme into the MRF framework, we propose a novel unsupervised classification and segmentation method for polarimetric synthetic aperture radar (PolSAR) images, abbreviated as DPMM-SMMRF. First, the DPMM built by the multidimensional Gaussian distribution is introduced into the MRF framework, which enables the proposed DPMM-SMMRF model to identify the underlying number of clusters automatically. Second, to utilize the polarization information adequately and modulate the spatial correlation, the similarity measure between the neighboring polarimetric covariance matrices is utilized to construct the interaction term; thus, providing strong noise immunity and enhancing the ability of the classification of the sample pixels. Then, for updating the class labels and the parameters in the proposed DPMM-SMMRF model, we propose a detailed sampling procedure based on the Gibbs sampling. Experiments on real PolSAR images demonstrate that the proposed DPMM-SMMRF model can automatically recognize the number of clusters and simultaneously obtain higher classification accuracy, more accurate edge location, and smoother homogeneous areas compared to several recent MRF models.