Polarimetric SAR Classification Based On Deep Feature Learning

Polarimetric SAR image classification problem is an important research content in the understanding and interpretation of polarimetric SAR images.Its purpose is to determine the category to which each pixel belongs,such as oceans,cities,and forests.Polarimetric SAR image classification is widely used in many fields such as civil and military.For the polarimetric SAR classification problem,the feature extraction and classification method is the key point of the study.For feature extraction,most of the traditional methods are based on the scattering and statistical properties of polarimetric SAR data.These methods rely on the understanding of the polarimetric SAR mechanism and manual design.In recent years,with the continuous development of deep learning methods,polarimetric SAR classification based on deep learning continue to emerge.Based on 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).This paper is based on the research of the generative adversarial network and applies it to the polarimetric SAR classification.The main work is as follows:1.Semi-supervised polarimetric SAR classification method based on the generative adversarial network is to expand the discriminant network into the k classifier.In the unsupervised case,the discrimination network can learn the distribution of the data,and in case of supervision,the discrimination network can learn the category distribution of the data.Through the mutual game between the generation network and the discrimination network(K classifier),the classification effect of the discrimination network is improved.Finally,by using a small amount of labeled samples and making full use of unlabeled data,the accuracy of the classification is improved.Overcoming the high feature richness of the polarimetric SAR image information to be classified in the prior art,it is necessary to explore the insufficiency of the scattering characteristics and statistical characteristics of the polarimetric SAR.2.Generative adversarial network polarimetric SAR image classification method based on Wishart-restricted boltzmann machine(W-RBM).In the semi-supervised polarimetric SAR classification method based on the generative adversarial network,the generated sample is generated by a randomly initialized generation network of random noise as input.Since the distribution of data is not considered in the generation network,so the learning efficiency of the generation network is low.Taking the true samples of the training set as input and pretraining the network via a stack autoencoder or a Wishart-restricted boltzmann machine structure.It can make full use of the unlabeled data and consider the Wishart distribution of polarimetric SAR data to generate the generated samples.When the generation network is enhanced,the convergence speed of the discrimination network is accelerated,and the classification effect of the discrimination network on polarimetric SAR data is indirectly improved.3.Generative adversarial network polarimetric SAR image classification method based on long short-term memory neural network(LSTM).Polarimetric SAR data has spatial information.Each pixel is not completely independent and has a correlation with the pixels in the neighborhood.For the semi-supervised polarimetric SAR classification method based on the generative adversarial network,in order to overcome the problem of not considering neighborhood information when classifying one pixel of polarimetric SAR data,the data of one pixel point and its eight neighboring pixel points in the polar SAR data will be generated as a sequence data.The discriminant network is changed to a multi-layer LSTM,and the sequence data including the neighborhood information is taken as the input of the discriminant network.As a result,the neighborhood information is taken into consideration and the classification accuracy is improved.