Research On The Key Technologies Of Object-oriented Typical Target Extraction For PolSAR Imagery

PolSAR (Polarimetric Synthetic Aperture Radar) can obtain scatteringcharacteristics of targets through recording electromagnetic wave scatteringinformation of several polarization states. The PolSAR data are represented byscattering matrix which takes the advantage of combining the polarizationcharacteristics, power and phase of targets into scattering matrix to provide abundantfeatures for the interpretation of targets. Therefore it is being theoretical andapplicable signification to develop PolSAR remote sensing technology, to minepolarization characteristics in the scattering matrix and to carry out efficientclassification and recognition of targets.With the increasing of PolSAR imagery resolution, the traditional processingmethods based on single pixel hasn’t been met the need not only in computerefficiency but also in the results. Whereas the application of object-oriented analysismethods in PolSAR imagery are poor and the technology is not advanced. Thedeficiencies limit the efficient application of PolSAR imagery deeply.Aiming at the central problems about object-oriented information extraction forPolSAR imagery currently, taking care of the complexity of the existent theories andproblems, with the latest theories of polarimetry, computer vision, image processingand graph theory, this thesis made an deep study on the correlative fields with thestrategy from speckle reduction and image segmentation to object-oriented targetextraction, and some meaningful conclusions were obtained.The main contents and innovations in this thesis can be summarized as follows.1. Beginning with the vector properties of polarization electromagnetic wavetransmitted by PolSAR, the partially polarised wave, completely polarised wave anddiversified representation of scattering characteristics of targets are summarized andanalyzed. And polarimetric compositions, polarimetric decompositions wereexpounded. An anatomy was made on the polarimetric interferometric SAR theory atlast. Above-analysis offers the essential theoretical supports for the study ofsubsequent algorithms.2. A novel speckle reduction algorithm for PolSAR imagery based oncurvature-preserving PDE was proposed. On the basis of summarizing the bornmechanism, characteristics of speckle and status of speckle reduction for PolSARimagery, aiming at the current problems about speckle reduction for PolSAR imagery, the application of tensor-driving PDE filters in speckle reduction were studied deeply.Aiming at the shortage of existent algorithms which can’t preserve the curvaturestructures, with the characteristics of PolSAR data, a novel speckle reductionalgorithm for PolSAR imagery based on curvature-preserving PDE was proposedwhich has the ability of not only reducing the speckle and preserving edge details butalso preserving the curvature structures such as corners.3. A novel segmentation algorithm for PolSAR imagery based on improvedspectral clustering was proposed. With the motivation of finding antinoisesegmentation algorithms with for PolSAR imagery, with the characteristics of PolSARdata, a novel segmentation algorithm for PolSAR imagery based on improved spectralclustering was proposed through introducing reverse Wishart distance as distancemeasurement. The antinoise property of our algorithm which could be applied to thePolSAR data was proved by experiments.4. A novel hierarchy aggregation multiscale segmentation algorithm for PolSARimagery based on spectral graph portioning was proposed through introducing theidea of multiscale segmentation algorithm by hierarchy aggregation. Initializedsegmentation was operated on the PolSAR data firstly. And then multiscale graphswere built by selecting feeds. The similarity was computed by joining cues of intensity,shape, size, and edge strengthen etc. The segmentation processing was completedafter deciding the relationship of objects by their saliency measures. The antinoiseability and multiscale analysis capability of our algorithm for PolSAR data wasproved by experiments, so that information of objects with different scales could beobtained.5. A novel multiscale segmentation algorithm for PolSAR imagery based onKummerU distribute was proposed. At the background of high resolution PolSAR,aiming at characteristics of high resolution PolSAR data with rich texture,complicated clutter model, heterogeneous clutter model based on KummerU distributewas analyzed systematically and the parameter estimation methods were summarized.The similarity based on KummerU distribute was designed by introducing maximumlikelihood function. A novel multiscale segmentation algorithm for PolSAR imagerybased on KummerU distribute was proposed with the idea of multiscale segmentationalgorithm by hierarchy aggregation.6. Taking forest, water and residential area as example, polarimetriccharacteristics and object-oriented extraction methods of typical targets were studied. Forest and water extraction gene with better applicability were proposed separately,and the polarimetric characteristics, namely the phase information of correlationcoefficient in the circle polarization basis, which is sensitive to residential areaextraction was studied. The extraction routines for forest, water and residential areawere designed separately.