Studies On Projection Pursuit Method Of Polarimetric SAR Image Classification

The POLarimetric Synthetic Aperture Radar (POLSAR) measures the scatteringecho of a target in different polarimetric modes and obtains the object'shigh-dimensional characteristics to describe the objection. A classification ofPOLSAR image is an important research of remote senseing application. Usually,POLSAR image is classified by physical emitting features or by statisticcharacteristics. Due to the complexity of the objection of the earth's surface and theeffect of coherent echo, there is certain classification precision problem, Therefore,based on the scattering characteristics of the objection, this thesis presents newstatistical classification methods of POLSAR image, named the Sequential ProjectionPursuit Cluster-Model (SPPCM), Projection Pursuit Wavelet Learning Network(PPWLN) and Mixture Wishart Model (MWM). These methods combine thescattering characteristics of the objections of the POLSAR. The details of this thesisare as followings:(1)A wavelet-kernel estimation of the Cook Family projection index is proposed.And its asymptotic unbiasedness and the convergence in mean squared are proved.Compare with the estimation of Cook Family projection index based on theorthogonal polynomial or the kernel function, wavelet-kernel estimation not onlyestimates precisely the projection index, but also bypasses the difficult problem of theselection of the order of orthogonal polynomial or the selection of the bandwidth ofkernel function. It provides a new tool for fulfilling projection pursuit method.(2)Bootstrap estimation of the Cook Family projection index is proposed andstrong consistent of this estimation is proved. Based on bootstrap estimation, rules ofsample selection and quasi-optimal projection direction are defined in order to providea theory foundation for fast fulfillment of projection pursuit.(3)The unsupervised classification method of POLSAR image, based on theSPPCM, is proposed. By using the scattering characteristics of the objection, thismethod projects data to the optimal direction to classify POLSAR image time aftertime. Unlike the current classification methods based on the scattering characteristic,this method can simultaneously makes use of multi-dimensional characteristics of theobjection to classify POLSAR image. Experimental results show that the quasi-optimal projection direction car be found with a few bootstrap samples,fulfilling the fast classification of POLSAR image. The classification results aregreatly improved.(4)The unsupervised restoring method of optical images is proposed by usingPPWLN. PPWLN trains the network by approximating degradation factors so as tobypass the difficult task of estimating point-spread function when little is knownabout degradation factors. The supervised classification method of POLSAR image isproposed by using the PPWLN. Compare with the Wishart classifier proposed byLee, U.S. Navy Lab, this method can avoid the procedure of filter POLSAR data. Theexperiment results show that this method has the better behavior of classifying targetswith complex characteristics comparing with the Wishart classifier under sameconditions.(5) The unsupervised classification method of POLSAR image, based on MWM,is proposed. This method directly classify POLSAR image without prior information.The experimental results show that this method is obviously superior to the Wishartclassifier on classifying targets with large homogeneous area.(6)The typical classification methods of the POLSAR image and our methods arecompared and analyzed. The strongpoint and shortcomings of our methods and theapplicability of our methods are analyzed.