Polarimetric SAR Image Classification Based On Sparse Coding And Deep Bandelet Network

Polarimetric Synthetic Aperture Radar(Pol SAR)has been widely used in various fields because of the advantages of all-time and all-weather.Polarimetric SAR image classification is a very important aspect of SAR image interpretation.So it is of great significance for the research of polarization SAR image classification and a good classification algorithm can lay a solid foundation for subsequent image recognition.This thesis mainly proposes various classification algorithms for polarimetric SAR image,which based on the theories of sparse coding and deep learning.Sparse coding can reduce the redundancy of the feature of image sample set,which is conductive to representative feature extraction.Besides deep learning can extract a more abstract feature for the original polarimetric SAR data and obtain good results on the polarimetric SAR image classification.By combining the two theories above,this thesis proposes three methods to realize the classification of polarimetric SAR images.The main work is as follows:A novel classification method for polarimetric SAR images based on PCA(Principal Component Analysis)and SVGDL(Support Vector Guide Dictionary Learning)is presented in this thesis.The main idea of the method is as follows.Firstly,we extract the polarimetric neighborhood features on polarimetric SAR data.Then we apply the PCA on the feature matrix of polarimetric SAR image to realize dimensionality reduction for reducing computational burden and redundancy of the model.Finally,we use SVGDL model to learn dictionary and get the final classification result.The method is applied on three polarimetric SAR datas to obtain good result.It can be seen that the application of this method on the polarimetric SAR image classification can not only improve the classification accuracy but also reduce the computational complexity.This thesis proposes a novel classification method of polarimetric SAR images,which based on the SSAE(Stacked Sparse Auto Encoder)and SVGDL(Support Vector Guide Dictionary Learning).As one of the deep learning model,SSAE could provide more abstract and effective features for polarimetric data.Additionally,SVGDL model can realize dictionary learning for polarimetric SAR data and get the encoding vectors of polarimetric SAR data.Firstly,the polarimetric SAR data is filtered by the Lee filter method and the training samples and test samples are selected according to the true geographical distribution of the polarimetric SAR image.Then we apply SSAE on the polearimetric SAR data to extract more efficient features.Finally,the feature matrix is put into the SVGDL model to encode and realize classification.The experimental results show that the application of this method on the polarimetric SAR image classification can improve accuracy effectively.A classification method for polarimetric SAR image classification based on deep Bandelet network is presented in this thesis.Bandelet can get the optimal representation of image.In addition,the Convolutional Neural Network can extract features for two-dimensional image and reserve the information of spatial structure of image.The main idea of this method is that combining the Convolutional Neural Network with the multi-scale Bandelet.The Bandelet transform is applied on the convolutional layer of the Convolutional Neural Network to construct the Deep Bandelet Network.The Deep Bandelet Network could not only extract the image features effectively but also get higher accuracy of polarimetric SAR image classification without increasing the running time.