Spatial Information Based PolSAR Imagery Segmentation And Classification

Polarimetric Synthetic Aperture Radar(PolSAR)is an advanced remote sensing system for earth observation.Compared with single-polarization type,PolSAR is capable of obtaining more abundant and complete scattering information of land covers and targets,therefore it has been increasingly used in national economic development and military field.PolSAR image segmentation and classification play key roles in PolSAR interpretation,and they are very important in improving the utilization level of PolSAR data,from the aspects of both theoretic sense and application value.Statistical modeling and polarimetric target decomposition are two main ways for PolSAR interpretation.However,along with the increase of image resolution,spatial information is getting more and more prominent,and it is helpful to improve the performance of PolSAR interpretation.This dissertation studies spatial-information-based PolSAR image segmentation and classification,in a systematic way,and it mainly focuses on the extraction of spatial information,as well as the application of spatial information in PolSAR segmentation and classification.The main content of the dissertation can be summarized as follows:(1)Integrating spatial information and H/ ? decomposition for PolSAR classification.The polarimetric decomposition method that is based on eigenvalue/eigenvector decomposition,as well as the classification methods based on the partition of H/ ? plane are introduced at first.In order to alleviate the effect of speckle noise and to improve the classification result,this dissertation proposes to integrate the spatial correlation information with the polarimetric information,besides applying the frameworks of fuzzy C-means clustering and Wishart iterative refinement.In the iterative procedure,the prior probability extracted from local MRF analysis is combined with the original fuzzy membership,generating a more reliable fuzzy membership function.In addition,the proposed method adopts adaptive neighborhood and utilizes adaptive smoothing factor,then it can execute smooting operation adaptively,according to the local spatial characteristics.Therefore,this method can prevent oversmoothing and preserve the local spatial details.The experimental results indicate that the spatial information can effectively restrain the pepper-and-salt phenomenon in the classification result,and improve the classification accuracy.(2)Polarimetric target decomposition of PolSAR data.Several classic decomposition methods are reviewed,including Yamaguchi-four-component decomposition,Sato-four-component decomposition,and multiple-component decomposition.However,in urban areas,these methods cause scattering mechanism ambiguity between rotated buildings and forests.In order to enhance the power of double-bounce scattering of rotated buildings and also suppress their volume scattering,this section proposes a modified method for multiple component polarimetric decomposition.This method applies a general volume scattering model to describe the HV scattering from vegetated areas,and a new volume scattering model for the HV scattering from rotated buildings.In addition,this new volume scattering model is closely related to the polarization orientation angle of the buildings,thus it can adaptively determine the volume scattering power.The decomposition experiments using real PolSAR data set and the subsequent classification experiments demonstrate the effectiveness of the proposed method.(3)PolSAR edge detection and superpixel segmentation.Since the distance measure based on PolSAR statistical distribution is the base of edge detection and superpixel generation,the Wishart distance and SIRV distance between regions are derived at first,and it is indicated that the SIRV distance can better discriminate the complex scattering characteristics in heterogeneous areas.In order to overcome the limitation of traditional fixed-shape windows in estimating the distribution parameters,the directional span-driven adaptive(DSDA)window is investigated,and then the modified edge detection method is proposed.This method can effectively detect the tiny and inconspicuous edges in heterogeneous areas.As to superpixel segmentation,this section combines the Wishart distance with SIRV distance,generating an integrated distance measure,and then introduces the entropy rate method into PolSAR superpixel segmentation.The proposed method can obtain homogeneity-adaptive segmentation results,with well preserved details in heterogeneous areas.(4)Superpixel-based PolSAR segmentation.Compared with the traditional Wishart distribution,KummerU distribution can describe the PolSAR data more accurately,especially for the data with rich textural information.In this section,the KummerU distribution is introduced,and the parameter estimation method based on the matrix log-cumulants is also deduced.Based on the segmentation method which applies the classical hierarchical region merging,a two-stage merging strategy is proposed,which includes the Wishart merging stage(WMS)and the KummerU merging stage(KUMS).WMS quickly merges the initial superpixels that are definitely from the same land covers.Then,in KUMS,the KummerU energy loss is combined with the edge penalty and the homogeneity penalty to guide the iterative merging procedure.The experiments using real PolSAR data sets indicate that the proposed method can significantly improve the calculation efficiency and segmentation accuracy.