Multilevel Distribution Coding Model-Based Dictionary Learning for PolSAR Image Classification

This paper presents a new unsupervised classification method of polarimetric synthetic aperture radar (PolSAR) data based on dictionary learning. First, a multilevel distribution coding model is proposed to encode the probability distribution of the rearranged matrix of each pixel in a PolSAR image; this model can generate a stable and adaptive representation of the images, which can be used to extract better feature vectors of the PolSAR data using a new dictionary learning method. The proposed model can increase the separability of terrains and effectively discriminate one class of pixels from another. Then, the k-means clustering is used to perform initial classification of the PolSAR image, and the initial classification map defines training sets for classification based on the complex Wishart classifier. Finally, in order to improve the performance of classification, we use the maximum-likelihood (ML) classification based on complex Wishart distribution to refine the clustering result. Five PolSAR datasets, including the RADARSAT-2 C-band data of western Xi'an, China, are used in the experiments. Compared with the other two state-of-the-art methods, H/α-Wishart and Lee category-preserving classification methods, the proposed one shows improvements in accuracy and efficiency, as well as high adaptability and better consistency.