Polarimetric SAR Image Classification Based On Scatter Matrix And Sparse Deep Learning

Polarmetric Synthetic Aperture Radar(Pol SAR)is a powerful tool for obtaining the information of ground to space,spectrum and incident(or observation)direction changes.People continue to improve the spatial resolution of the image,at the same time,they use of multi-channel polarization method to obtain a large number of feature information,thus providing more information support to improve image classification capabilities.The complete polarization data from the target obtained by polarmetric synthetic aperture radar can provide a large number of rich target scattering information,so that people can more effectively carry out the remote sensing images classification and recognition tasks.And the sparse representation is able to remove the redundancy in these features,making the classification faster and more efficient.In this thesis,we mainly study the method of polarimetric SAR image classification,and systematically study the scattering information and sparse coding of polarimetric SAR images.The main contents are as follows:A method based on sparse coding and dictionary pair learning for classification of polarimetric SAR image is proposed.This method adds an analysis dictionary to the traditional dictionary learning method.That is,we can learn the polarized scattering characteristics after sparse coding by the dictionary pair of the synthesis dictionary and the analysis dictionary,and then we can make a multiple classification of images according to the residuals.Several experiments performed on different datasets show that this method can show good classification accuracy and competitive running speed.This thesis provides a method based on polarization-texture feature sparse coding and dictionary pair learning for classification of polarimetric SAR image.This method combines the polarization scattering feature and the image texture feature in an original polarmetric SAR image.After the sparse coding of the obtained polarization-texture feature,the dictionary pair is used to learn these features,which makes the characteristics in the dictionary pair learning more abundant and complete.Compared with using polarization scattering characteristics alone,this method exhibits highly competitive classification accuracy without increasing the cost of running time.A method based on polarization-texture feature sparse coding and stacked sparse autoencoder for classification of polarimetric SAR image is proposed in this thesis.This method utilizes multiple sparse autoencoders to construct a stacked sparse autoencoder model.With this model,more useful information can be excavated to obtain a higher level characteristics than the post-coding polarization-texture feature.Semi-supervised learning is carried out according to some known labeled samples,and finally all samples will be classified by the Softmax classifer.Three datasets are validated,and this method achieves a good classification result.