Feature Extraction Of Hyperspectral Images Based On GAN

In recent years,hyperspectral image classification has gradually become a hot topic in the hyperspectral field,and it plays an important role in national economic development and national security.Hyperspectral images have the characteristics of high dimensionality,high redundancy,and large amount of data.Extracting features of hyperspectral images to improve classification accuracy is a key issue in hyperspectral image classification tasks.The quality of features directly affects the accuracy of classification.However,the phenomenon of same thing but different spectrum and same spectrum of foreign matter of hyperspectral images,the small number of labeled samples,the high cost of sample labeling,and the huge amount of data make the feature extraction and classification of hyperspectral images face many difficulties and challenges.In order to improve the above problems,this thesis combines generative adversarial networks(GAN)and convolutional neural networks(CNN)for feature extraction of hyperspectral images.The main research work of the thesis is as follows:(1)In order to reduce the dependence on labeled samples in hyperspectral images,GAN is introduced on the basis of CNN,and an unsupervised method for extracting features of hyperspectral images is proposed.In order to stabilize the training process of the network and improve the feature representation ability of the discriminator in the generative confrontation network,a gradient penalty term is introduced into the objective function.In the feature extraction stage,for the spectral structure of the hyperspectral image,a channel maximum pooling method is proposed,which can reduce the data dimension while retaining the spectral information of the hyperspectral image as much as possible.Use support vector machine(SVM)and k-nearest neighbor(KNN)methods to classify and test the extracted features.Experimental results on two real data sets show that the proposed method is better than traditional feature extraction methods.(2)Aiming at the difficulties and problems faced by the above-mentioned hyperspectral image feature extraction,focus on the space-spectrum joint feature extraction of hyperspectral images and the improvement of model adaptation capabilities,this thesis propose an unsupervised hyperspectral image space-spectrum feature extraction method based on the generative countermeasure network.This method combines GAN and CNN to get rid of the dependence on hyperspectral labeled samples by means of counter-training.At the same time,it uses the spectral information and spatial information of the hyperspectral image,and uses complementary information to alleviate the phenomenon of same thing but different spectrum and same spectrum of foreign matter.The principal component analysis(PCA)algorithm is used to reduce the dimensionality of the hyperspectral image to reduce the amount of calculation,improve the adaptability of the model to different data sets,and reduce training time and cost.The experimental results on three hyperspectral data sets show that the space-spectrum features obtained by the proposed method have good classification performance.(3)In order to solve the problem of fewer labeled samples for hyperspectral images,this thesis proposes a feature extraction method for hyperspectral images based on sample generation.In this method,we firstly use ACGAN to obtain the generated samples with labels,then mix the generated samples with real samples for CNN training,and finally use the well-trained CNN for the features extraction of the original data.Experiments were conducted on the Indian Pines dataset with extremely uneven sample size.The results show that the features obtained by the proposed method have better classification performance,which proves the effectiveness of sample expansion.