Research On HSI Supervised/Semi-supervised Classification Based On Generative Adversarial Networks

Hyperspectral remote sensing image contains the information of spatial structure,location information and spectral characteristics of objects.As a basic technology widely used in remote sensing technology,hyperspectral image classification is one of the important ways for people to obtain landmark information.The characteristics of massive data,high-dimensional,small samples of hyperspectral remote sensing usually bring opportunities and challenges to image classification.On the one hand,the traditional methods of hyperspectral classification have weak ability to extract features from hyperspectral images.On the other hand,in practical applications,the conventional hyperspectral sample data sets have fewer types of labels,and the additional cost of obtaining labeled information is higher.Based on this,in order to improve the classification accuracy of hyperspectral image,on the one hand,the hyperspectral image is preprocessed by Gabor filter,and then the convolution neural network is used to extract deep features to complete image classification.On the other hand,the generative adversarial network model is improved suitably,and the supervised and semi-supervised modesl for classification are proposed for the first time.This model can not only effectively improve the feature extraction ability of hyperspectral images,but also effectively combine a large number of unlabeled samples to further improve the classification accuracy by semi-supervised method.The main contents of this paper include:Firstly,under the condition of few labeled samples,this paper combines Gabor filter with deep convolution neural network to pre-process the image effectively,and alleviates the over-fitting phenomenon of deep neural network with fewer training samples.Secondly,in order to further improve the ability of feature extraction of hyperspectral images,a method of generative adversarial network based on deep convolution neural network is proposed for the first time.This method can not only utilize the spectral information of hyperspectral images,but also effectively combine its spatial spectral information.The model is further trained and optimized by the adversarial action between generator and discriminator,which can effectively improve the classification accuracy compared with traditional methods.On the other hand,for the first time,this paper accomplishes the semi-supervised classification task of hyperspectral images by generative adversarial network,and sends the unlabeled samples into the trained DenseNet network model to predict the labels of the samples.Then the unlabeled samples and prediction labels are jointly sent into the Semi-GAN with the original labeled sample data set to train the model.This method can be used to modify the decision plane obtained by lacked labeled samples,and improve the classification accuracy deeply,besides,the problem of scarcity of labeled samples can be effectively alleviated.