Polarimetric SAR Image Classification Based On Contourlet DCGAN

With the advancement of technology,the demand for remote sensing imaging technology is increasing.As the most popular technology in remote sensing imagery,polarimetric SAR has many advantages,such as all-weather,all-day,high resolution,side-view imaging,and so on.Polarimetric SAR images are rich in polarization information and have research value.The polarimetric SAR image classification is a fundamental and effective way to study polarimetric SAR data,and it is also an important part of the application of polarimetric SAR images.However,the traditional technology of the polarimetric SAR image classification can no longer meet people's needs.Deep learning is a research hotspot in recent years.It simulates the human cerebral cortex and has powerful feature representation capabilities.This paper studies the application of DCGAN in deep learning in the field of the polarimetric SAR image classification to improve its accuracy and reliability and promote its development.The main work in this paper is as follows:1.In order to solve the problem that labeling data manually is difficult,a polarimetric SAR image classification algorithm based on DCGAN is proposed.A large number of unlabeled samples is used to train DCGAN unsupervised.DCGAN can learn the distribution characteristics of data from a large number of unlabeled data.Therefore,this method can still achieve very high classification accuracy with few labeled data.A lot of experiments are conducted with five sets of data.Then this method is compared with five other methods.It was proved that with a large amount of unlabeled data information,only a small amount of labeled samples can achieve high classification accuracy,and a good classification result image can be obtained.2.Starting from multi-scale analysis,this paper proposes a polarimetric SAR image classification algorithm based on Contourlet-DCGAN.With improving the DCGAN,and combining with the Contourlet transform,the Contourlet-DCGAN model is proposed.It not only uses a lot of information of unlabeled data,but also adds Contourlet transform to sparsely represent the polarimetric SAR image and extract the edge and texture features of the images.Finally,experiments were performed with different resolutions and platform data,and other methods were compared.It is proved that the method can extract multi-scale multiresolution features and improve the classification accuracy.3.For the problem of many network parameters,this paper propose a polarimetric SAR image classification algorithm based on Contourlet Depthwise-DCGAN.Depthwise deconvolution is proposed according to depthwise convolution.Then the standard convolutions in Contourlet-DCGAN is replaced by depthwise convolutions.The standard deconvolutions are replaced by depthwise deconvolutions to form the Contourlet DepthwiseDCGAN model.Finally,five groups of data and different comparison experiments were used to verify that this method not only reduces the complexity of the model,avoids the risk of overfitting,but also improves the ability of feature representation and further improves the classification accuracy.