Polsar Classification Based On Deep Learning

Polarimetric Synthetic Aperture Radar(Pol SAR)is an active microwave radar imaging technique,it utilizes multi-channel to obtain data and imaging,so Pol SAR data contains more abundant polarization information and richer feature information.The Pol SAR classification has also been an important research direction in the field of radar image understanding and processing.In recent years,deep learning has attracted widespread attention,especially in image classification and recognition.Supported by the National Natural Science Foundation of China(Pol SAR image classification based on co-training and sparse representation,No.61173092),and the National Natural Science Foundation of China(Pol SAR image classification based on generative adversarial network,No.61771379),We proposed a semi-supervised classification method based on AC-GAN(Conditional Image Synthesis With Auxiliary Classifier GANs).Main tasks are as follows:First,Pol SAR classification method based on the semi-supervised learning with AC-GAN network.Because Pol SAR data contains more noise,the model required to be trained must be robust to noise.The Generative Adversarial Net is based on the idea of a zero-sum game,and the generator G and discriminator D of the network are continuously improving in competition with each other,thus the discriminator D obtained is also more robust.In addition,the Pol SAR data is difficult to label,and it is often faced with the problem of small sample size.Therefore,we restructure the AC-GAN network and obtain a new semisupervised network.Experiments have shown that this new network can be trained with a small number of labeled samples and lots of unlabeled samples and can be well applied to Pol SAR classification.Second,Combining multi-scale Convolutional Neural Network(CNN)with the semisupervised AC-GAN.Most of the previous classification methods are based on per-pixel,and rarely consider the neighborhood information of the image.However,the neighborhood information is very important that is very helpful for improving the classification results.Based on the center pixel,we select a fixed-size neighborhood to represent this pixel,utilize the pixel's label as the label of this neighborhood block,and then train the entire network.Through experiments,the use of neighborhood information and multi-scale convolution can better enhance the classification ability of the model.Third,Combining Wasserstein distance with the semi-supervised AC-GAN network.Because the original GAN has the disadvantages of unstable training,insufficient generator diversity and easy collapse of the model,We acquire the causes of these problems by analyzing the cost function of the original GAN.Then the Wasserstein distance is applied to the semi-supervised AC-GAN network.Experiments show that the convergence speed of this model is faster,the training is more stable,and the classification results have been further improved.