Polarimetric Information Reconstruction Based On Deep Learning

Quad-polarization Synthetic Aperture Radar(SAR)can obtain comprehensive polarization information,but it has the defects of narrow swath width and low spatial resolution.The compact polarization SAR is a dual-polarization system,and its swath width is large,but the polarization information is less than quad-polarization SAR,and it is difficult to replace quad-polarization SAR.For the problem that the quad-polarization swath width is small,the information reconstruction of the compact polarization data can be performed to supplement the polarization information missing in the polarization dimension.For the problem of low spatial resolution of quad-polarization SAR,the low-resolution quad-polarization image can be reconstructed in the spatial dimension to complement the lack of polarization information in the spatial dimension and improve its spatial resolution.The main difficulty of the above problems is that the reconstructed quad-polarization SAR data needs to be highly consistent with the polarization information of the real data,and the reconstruction accuracy of the traditional method is low.The powerful learning ability and nonlinear fitting ability of deep learning method can help to solve the consistency problem of quad polarization SAR image reconstruction.Therefore,this paper focuses on the reconstruction of the polarimetric information of the compact polarization SAR data based on deep learning and the spatial dimensional information reconstruction of the quad-polarimetric SAR image.The main work and contributions of this paper are as follows:1.In polarimetric SAR images,there are ground objects of different scales(such as buildings and farmland),and feature extraction needs to be performed corresponding to receptive fields of different sizes.Aiming at this problem,this paper proposes a dual-branch convolution kernel self-selecting complex convolutional neural network(DB-SK-CVCNN)to perform quad-polarization reconstruction on compact polarization data.The two branches of the network use convolution kernels of different sizes to extract features from polarization images.In this thesis,the Selective Kernel(SK)convolution unit is introduced and improved to generate more fine-grained attention information,which can make the neural network adaptive adjust the receptive field.The experimental results show that DB-SK-CVCNN can guarantee the accuracy in the reconstruction of various types of ground objects,and also fully retain the polarization information.2.There are large empty areas and local dense areas in the quad-polarization SAR image.In dense areas,the texture details are richer,the scattering mechanism changes much,and the super-resolution reconstruction on this area is difficult.To solve this problem,this paper proposes an attention-based complex convolutional dual-branch adversarial generative network(CV-DB-GAN)to improve the resolution of quad-polarization SAR images.The introduction of Convolutional Block Attention Module(CBAM)makes the neural network pay more attention to the densely distributed areas,so that the network can reconstruct richer texture details.Experiments show that the quantitative analysis and polarization decomposition results of CV-DB-GAN are better than the two state-of-the-art super-resolution reconstruction methods for quad-polarization SAR,and the reconstructed quad-polarization SAR images are more accurate in phase.