Polarimetric Synthetic Aperture Radar Imagery Classification Based On Region Similarity And Characteristic Dimension Reduction

Classification of Polarimetric Synthetic Aperture Radar (PolSAR) is a significant research in SAR imagery process and Analysis. During last decades, many Algorithms have been proposed and great progress has been made in classification based on pixels. However, these methods couldn’t deal with the effect of noise in SAR imagery on classification map and have difficult in taking full advantage of spatial information. Therefore, these methods hardly provide classification map with high precise. Meanwhile, the classification method based on regions can deal with these shortages, and it is researched in this thesis.Region classification includes imagery segmentation and classification of regions, the progress of these two aspects in PolSARdata is firstly summarized in this paper, and the difficulty of segmentation is how to get an accurate image segmentation results taking into account the computational efficiency and reducing speckle noise; lacking of effective initial clustering method restricts the development of regional non-supervised classification; the trend of supervised classification research is to make valid application of polarization scattering characteristic. A series of regional segmentation and region classification methods are proposed to deal with these difficulties by summarizing the existing research and introducing or drawing on the graph partitioning, clustering research and data dimensionality reduction algorithms. Overall, imagery is segmented by region segmentation method and the initial segmentation and region merging is improved to get a better result; some initial clustering methods based on region similarity are proposed in unsupervised classification, and the polarimetric scattering characteristic is processed by dimensionality reduction in region supervised classification. Specifically, the main researches of this paper are summarized as follows:(1) A method of PolSAR imagery segmentation based on regional statistical is proposed. The watershed is used to get initial segmentation. The CFAR ratio of average is introduced to instead of traditional difference gradient which is not CFAR, and the geodesy corrosion reconstruction is introduced to deal with the excessive over-segmentation of watershed, therefore a better initial segmentation can be obtained. An object function giving a better description of region similarity is presented by combination with the statistical distribution of SAR data and hypothesis testing, which can provide a better segmentation.(2) The normal cut is applied to region initial clustering.The similarity matrix whose element is the similarity of region with other is constructed based on revised wishart distance, and then the similarity matrix is process by Ncut to get the initial clustering centers.(3) An initial clustering method is developed combination with learning the process of similarity matrix from clustering trend and K-means. The reorder method of similarity matrix in clustering trend is introduced to provide a preliminary clustering, and stable clustering centers is obtained then by k-means iteration of distance between preliminary clustering centers and regions until the condition.(4) Affinity propagation is introduced to the research of region initial clustering. The initial clustering centers are obtained by adjusting the parameters of affinity propagation whose input is similarity matrix.(5) Laplacian egenmaps and supervised laplacian egenmaps are introduced to dimensionality reduction of PolSAR data. The extracted region polarimetric scattering characteristic is performed dimensionality reduction in order to remove the redundant information and make effective use of polarimetric scattering information, which eusure to obtain high precise supervised classification result.