Object-oriented PolSAR Image Classification Based On Sparse Representation

PolSAR image classification has attracted more and more attention as one of the most important applications of radar image processing. Since the heavy noise existing in PolSAR data will seriously influence the analysis and process, the feature extraction method and image classification method to overcome the noise are the major problems of PolSAR image classification. To solve above problems, the object-oriented method and the sparse representation theory are introduced in this article. The object-oriented method which extracts objects for classification by segmentation could suppress the interference of noise. The sparse representation theory is robust to noise which will improve the classification accuracy. The combination of these two methods can obtain excellent classification results. The four main works of this article are summarized as followed.1. Segmentation operation is firstly conducted for PolSAR image. Texture feature extraction is introduced to statistical region merging method to obtain objects. Both Bhattacharyya distance merging criteria and statistical region merging criteria are used to obtain an accurate segmentation image result. The texture feature is extracted by the improved LBP operator: regional homogeneity local binary pattern operator(RHLBP operator). By setting a threshold value, RHLBP operator presents good performance in discriminating heterogeneous regions and describing homogenous regions.2. The color feature is applied to the object-oriented PolSAR image classification.Because the PolSAR data do not provide the true color information of the ground objects,the color feature is not used for image classification before. However, the composed false color image can provide visual information for the ground objects. By introducing color feature from the false color image, this article verifies the effectiveness of color feature for PolSAR image classification.3. A new original data based dictionary updating method is proposed. The dictionary used for sparse representation is randomly extracted from the data whose class labels are given. The original dictionary has a serious possibility that contains noise pixels or defective dictionary atoms. The defective dictionary atoms can not represent other pixels that belong to the same class. To obtain an effective dictionary for sparse representation,this article proposes a new dictionary updating method. This method selects good dictionary atoms from the original dictionary to compose a new dictionary. Good classification results are gotten in this way while keeping the same size of the dictionary.4. The improved joint sparse presentation method is presented. To reduce the calculation amount of sparse representation classification for every pixel, the joint sparse presentation method is introduced. The traditional joint sparse presentation method uses all the pixels in one segmented region to calculate the common pattern. However, the sizes of segmented regions obtained by the segmentation method in this article are different and most segmented regions contain a large number of pixels, it will lead to bad classification results. In this article, the pixels in one segmented region are divided into data sets whose sizes are fixed and the data sets are used for joint sparse representation classification. Then, the segmented regions are labeled by counting the classification results of the data sets. Thus, the classification result are obtained.