Polarimetric SAR Image Terrain Classification Based On Sparse Coding And SVM

Polarization SAR(Polarimetric Synhetic Aperture Radar),have measured each resolution cell for targets and recorded measured fully polarimetric scattering echo, from which people can analyze many polarization characteristics of the targets, is a kind of radar imaging system have coherence and use multi-channel measurement, called polarimetric synthetic aperture radar. Due to the special methods for polarized SAR imaging, it can get a wealth of objective information on characteristics and classification characteristics for polarimetric SAR data measured, to determin the pixel category has become the frontier issues about wether polarimetric SAR image can be achieved applied research and polarimetric SAR image terrain classification research has become the focus of future research technology development.In this thesis, on the basis of polarization scattering characteristics launched polarimetric SAR image terrain classification research on polarization scattering characteristics of different encoding methods, the characteristics of the nonlinear mapping to feature space where different from the original one,using SVM classification. Three new methods for polarimetric SAR image terrain classificationis based on this idea have been proposed, the main contents are as follows:1.The algorithm based on Gaussian pyramid pooling codeing and SVM for polarization SAR terrain classification has been proposed. Firstly, the feature information of polarimetric SAR image will be fully tapped, then introduce the Gaussian pyramid model into the space constituting by the scattering characteristic feature extracted from coherent matrix for the constructing space layers, then use the maximize pooling coding.Using this method,the space information between the field and the characteristic features structural information have been taken into account better than before, which researchers have just considered the characteristics of one single pixel leading to the accuracy result of polarization SAR images terrain classification is not high enought, and that the results of the classification for polarimetric SAR image improved a lot, classifying clearly the complexity graphics.2.The classification method based on ultra-sparse vector encoding and SVM been proposed for polarization SAR terrain classification. This method introduce super-vector coding method which once to encode the image feature set for polarimetric SAR image's scattering, using polarization SAR images- label matrix and other prior information to sparse the scattering by mapping the scattering to the non-linear and linear classifier linearly separable feature space, which is a new polarimetric SAR image feature extraction process, using this method allows the characterization of the original features improved, so that the accuracy for polarization SAR image classification has benn improved. The results of the three groups terrain classification of polarimetric SAR image data show that the method for polarimetric SAR image feature classification results better than others.3.The algorithm for polarimetric SAR image classification based on sparse Fisher vector encoding and SVM has been porposed. This method use Sparse Fisher Kernel framework as each pixel's feature extractor, combined with SVM classifier to classify the target, with Gaussian mixture model is modeled for samples with label.Then encod characteristics under the probability distribution model parameters.The new feature been obtained after encoding the original feature and with a stronger classification ability than original feature,just need to train a linear support vector machine classifier for classification, classification results can be very good. Three sets of results terrain classification of polarimetric SAR image shows that the method is effective.