Sar Target Recognition And PolSAR Classification Based On Deep Learning

With the progress of science and technology,satellite remote sensing technology has developed rapidly.The scale of remote sensing images is increasing,and the image resolution is also improving.The interpretation of remote sensing data such as synthetic aperture radar(SAR)and polarimetric synthetic aperture radar(Pol SAR)has become a hot topic.The traditional image-processing algorithm requires researchers to design complex feature-extraction procedures,and it is difficult to achieve good results when faced with complex and diverse remote sensing data.With the development of deep learning and neural network,deep neural network has been applied to data processing of SAR and Pol SAR.However,due to the differences between SAR/Pol SAR images with natural images,it is difficult for ordinary deep learning algorithms to fully exploit their advantages.In this thesis,for the problem of SAR image target recognition and Pol SAR image terrain classification,three new deep neural network models are proposed,and their effectiveness and advancement are demonstrated by some experiments.The main research contents of this thesis are summarized as follows:1)A deep memory convolutional neural network(M-Net)is proposed.On the basis of convolutional neural network(CNN),M-Net adds an information recording module to record the spatial features of the samples,and then uses the spatial similarity information of the sample features to predict the labels of unknown samples.Since M-Net adds the part of the information recording module,if traditional training method of CNN is used,it is likely that the network will not converge or converge slowly.In order to solve this problem,this thesis proposes a transfer parameter technique to train M-Net in two steps.The first step is to train an ordinary CNN with the same structure as M-Net,and the training parameters are saved.The second step is to train the whole M-Net based on the saved parameter.This way of two step training not only avoids the non-convergence issue of M-Net to some extent,but also speeds up the training process.Compared with the normal CNN,M-Net can better recognize the SAR image targets and alleviate the overfitting problem of CNN on the SAR dataset.2)A depthwise separable convolutional neural network with dense connections(SDNet)is proposed.According to characteristics of Pol SAR data,SDNet uses depthwise separable convolution to replace standard convolution,to independently extract features over each channel in Pol SAR images.It can avoid extracting redundant features from data,and reduces training parameters.SDNet also introduces dense connections to directly connect non-adjacent layers of the networks.With the dense connections,SDNet can not only preserve the information of the front layer,improve the reuse of features,enhance information transmission of the network,but also further reduce the parameters of the model.Compared with normal CNN,SDNet is more lightweight and its training parameters decrease to less than 1/9.3)A spatial feature-based convolutional neural network(SF-CNN)is proposed.The network adopts a dual-branch CNN structure.The two branches have the same structure and share parameters.SF-CNN can receive more than one sample as input.The special structure expands the original training set by combining different samples and it can alleviate the problem of too little high-quality labeled training data in Pol SAR image classification tasks.SF-CNN can map high-dimensional Pol SAR image to low-dimensional feature space.In low-dimensional feature space,SF-CNN enhances the ability of network to extract distinguishable features by maximizing or minimizing the distance between feature centers of samples.In order to dig up the relationship between the samples,the test sample features are compared with every training sample feature when testing.Finally,the labels of test samples are determined by the distance comparison result.The result of SF-CNN in Pol SAR image classification task is better than that of ordinary CNN.