Unsupervised Space-spectral Joint Feature Extraction Of Hyperspectral Images Based On Deep Learning

With the continuous development of space technology and imaging technology,the resolu-tion of hyperspectral images becomes higher and higher.which also brings great challenges to the feature extraction of hyperspectral images.Traditional hyperspectral image feature extraction methods are concerned with spectral information,and the extracted features have poor effect.In recent years,the method based on deep learning has been widely used because it can extract the spatial and spectral features of hyperspectral images simultaneously.How-ever,most feature extraction methods based on deep learning are supervised.Supervised feature extraction methods need a large number of labeled samples for model training,while the labels of hyperspectral images are pixel-level,such labels are difficult to obtain.Aim-ing at the problem that the traditional hyperspectral image feature extraction method fails to take into account the spatial neighborhood infonnation and the problem that the supervised feature extraction method is not applicable due to the limited labels of the hyperspectral im-age,three unsupervised feature extraction algorithms for hyperspectral images are proposed based on two ideas of unsupervised learning:"the idea of encoding and decoding" and "the idea of generative adversarial".The deep learning network proposed in this paper solves the problem that the traditional hyperspectral image feature extraction method fails to take into account the spatial neighborhood information,while the unsupervised feature extraction idea solves the problem that the supervised feature extraction method is not applicable.(1)Aiming at the problem that the traditional hyperspectral image feature extraction method fails to take into account the spatial neighborhood information and the problem that the supervised feature extraction method is not applicable,this paper first proposes an unsu-pervised feature extraction algorithm for hyperspectral images based on symmetric two-dimensional full convolution,which is proposed based on "the idea of encoding and de-coding".Based on autoencoder,the algorithm implements an end-to-end full convolution network,which consists of multiple convolution layers and deconvolution layers.The appli-cation of convolution can learn the spatial neighborhood information and spectrum informa-tion of sample points simultaneously,which overcomes the defects of traditional hyperspec-tral image feature extraction methods,while the application of deconvolution can restore the features to the original input.In the feature extraction process after the training,pooling and combination operations are used to process the original features.Pooling can enhance the robustness of features,and it also can reduce feature dimensions,which can avoid the Hughes phenomenon.The combination of multi-level features can improve the expression ability of the final features.It is worth noting that the process of network training and feature extraction is unsupervised,which solves the problem of unsuitability of supervised feature extraction methods.Experiments on hyperspectral image dataset show that the proposed algorithn has good performance.(2)Two-dimensional convolution can only perform convolution in spatial dimension,while three-dimensional convolution performs convolution not only in spatial dimension but also in spectral dimension.To solve this problem,this paper introduces three-dimensional con-volution into the unsupervised feature extraction framework described above.Hyperspectral images are often a large three-dimensional data block,and three-dimensional convolution is very suitable for this three-dimensional input of hyperspectral images,because it can make full use of spectral information.Therefore,this paper proposes an unsupervised feature extraction algorithm for hyperspectral images based on symmetric three-dimensional full convolution network,which is also implemented based on "the idea of encoding and decod-ing".While solving the problems that the traditional hyperspectral image feature extraction method fails to take into account the spatial neighborhood information and the supervised feature extraction method is not applicable,it also improves the effect of the features ex-tracted.Experiments prove that the effect of three-dimensional convolution is better.(3)The generative adversarial network can generate very real images after training,but the deep network based on "the idea of encoding and decoding" can not output such images,which shows that the generative adversarial network can obtain more details of input samples so as to generate clearer images.To solve this problem,an unsupervised feature extraction algorithm for hyperspectral images based on generative adversarial network is proposed,which is based on the "the idea of generative adversarial".By using Wasserstein distance to improve the original generative adversarial network and redesign the cost function,the problems of unstable training and difficult convergence of the original generative adversarial network are solved.The improved generative adversarial network consists of a generator and a discriminator.Unsupervised training of the network is accomplished through a zero-sum game between the generator and the discriminator.For the feature extraction part of the network,we use pooling and combination operations to process the original features to enhance the effect of the features.The classification experiments on real data sets verify the effectiveness of the algorithm.