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

Polarization Synthetic Aperture Radar(Pol SAR)has become one of the indispensable objects in the development of Synthetic Aperture radar at home and abroad,and polarization SAR image classification is a crucial research technology in the SAR image interpretation.It determines the category of the pixel according to the target polarization information obtained by resolution unit.Comparing with the shallow network model structure in the complex variable function of the representation,the complexity of the network computing and access to the useful information,the deep network model structure has a big advantage.The Sparse Automatic Encoder(SAE)is one of the deep network model structures.In this thesis,Polarized SAR images are classified based on wavelet-SAE,The network utilizes the self-learning ability of neural networks and has good approximation ability for all nonlinear functions,and for the same learning task,the time and space complexity of the wavelet network structure are often on the low side.The main contents include the following aspects:The method of classifying polarimetric SAR images by wavelet-SAE is described in detail.Gaussian wavelet function,Morlet wavelet function and Mexican hat wavelet function are used as the activation function of the network.The algorithm is better than the comparison algorithm on classification effect,and the wavelet-SAE uses far less total time than the other algorithms.Among them,Gaussian wavelet-SAE for polarized SAR classification has not only the highest classification accuracy but also the shortest time.This thesis systematically explains the polarimetric SAR image classification method based on Gaussian Pyramid Pooling Coding(GPPC)and wavelet-SAE.The image original data is coded and the spatial neighborhood information of the data is used.The results show that the classification accuracy of the algorithm has been significantly improved,especially for the German Oberpfaffenhofen area and San Francisco Bay area where classification effect is remarkable.In this thesis,the polarimetric SAR image classification method based on Super Vector Coding(SVC)and wavelet-SAE is expounded in detail.Combining the image spatial correlation and sparse coding the feature with the prior probability of the pixel,the experiments show that SVC has fully studied the original data,and has made a further improvement in the accuracy rate in the Flevoland region under the premise of improving the classification precision of high-resolution images.