Hyperspectral Image Classification Based On Unsupervised Domain Adaptation

As an important part of remote sensing image,hyperspectral image contains rich spectral and spatial information.Using spatial-spectral information can help people better understand the earth.Hyperspectral image classification is one of the common applications of hyperspectral data.It is designed to recognize the environment by classifying the hyperspectral image cell by cell.However,the similarity of spectral information between bands of hyperspectral data,the complexity of overall spatial information and the high labeling cost hinder the development of the application of hyperspectral image classification.To solve these problems,this thesis uses deep learning and domain adaptation technology to classify hyperspectral images.The main work includes:1.Aiming at the problem of high labeling cost,lack of labeled samples and large probability distribution of data between different hyperspectral images,combined with unsupervised domain adaptation technology,a hyperspectral image classification method based on deep adapted features alignment is proposed.The covariance and centroid alignment between deep adapted features of the same class from source and target domain are used to realize data distribution adaptation in two domains.Firstly,a convolution deep adaptation network is used to extract the deep adapted features from two domains.Then,in order to align data distribution of the two domains more accurately,the covariance and centroid alignment between the features of same class from two domains are realized by matrix transformation and transformation operations,while keeping the manifold structure of the features from source domain unchanged,and the alignment process is carried out iteratively.Finally,a basic classifier is trained by the covariance and centroid aligned features and labels from source domain,and the predictions of target data are obtained by this classifier.2.Combined with the characteristics “combining image with spectrum” of hyperspectral image,aiming at the problem that its spectral and spatial information is difficult to represent,a hyperspectral image classification method based on threedimensional subdomain adaptation network is proposed.Combining three-dimensional convolution operation and unsupervised domain adaptation technology,this method completes domain adaptation classification between different hyperspectral images by feature adaptation and classifier adaptation.Firstly,the spatial spectral features from source and target domain are extracted by three-dimensional convolution neural network with strong representation ability.Furthermore,spatial spectral features from two domains are input into the feature adaptation layer for features adaptation to reduce the maximum mean discrepancy of its subdomains,and align the covariance of its subdomains to obtain deep adapted features from two domains.Finally,a classifier adaptation module is used to classifier adaptation,and then the deep adapted features from target domain are input into target domain classifier to obtain the prediction.Experiments are carried out on Botswana,Kennedy Space Center,Houston and HyRANK datasets.The experimental results show that compared with the classical unsupervised domain adaptation methods,the two hyperspectral image classification methods proposed in this thesis have better classification performance in the application of cross domain hyperspectral image classification.There are 26 figures,14 tables and 110 references in this thesis.