Hyperspectral Image Classification Based On Domain Adaptation

As one of the basic and key technologies in the field of remote sensing,hyperspectral image classification technology is a key step in the application and analysis of the Earth’s surface.However,the application and promotion of hyperspectral image classification is severely constraints due to the difficulty and cost of obtaining labels for hyperspectral images.Domain adaptation can use source domain hyperspectral image with rich labels to improve the classification effect of target domain hyperspectral image with few or no labels,making it widely used in hyperspectral image classification.However,most existing domain adaptation methods pay little attention to the distance between clusters of different classes.If the distance between clusters is too close,it may cause large-scale misclassification of samples.On the other hand,domain adaptation usually reduces the distribution difference between the source and target domain through adversarial learning.However,adversarial learning may weaken the discriminability of features in the latent space,causing pixels with similar features to be classified into different classes.If fine domain adaptation is performed on this basis,it may exacerbate the misclassification of samples.To address the above problems,this thesis proposes two unsupervised hyperspectral image classification methods by combining the graph convolutional network and the domain adaptation technique,and the main work includes:(1)To solve the problem of too close distance between clusters of different classes in domain adaptation,an unsupervised hyperspectral image classification method based on graph dual adversarial network is proposed.First,the spatial-spectral features of the hyperspectral images are extracted by a graph convolutional network.Then,a prototype adversarial strategy is proposed,which uses the labeled data of source domain to reliably calculate the feature prototypes of different classes.Through the prototype adversarial strategy,the distance between different prototype are is appropriately extended.And overall distribution difference between two domains is reduced by the domain adversarial strategy.Finally,domain adaptation is further performed by reducing the distance between the second-order statistical features of each class of samples in both domains.(2)To solve the problems of low discriminability of features in a latent space,a soft instance-level domain adaptation with virtual classifier for unsupervised hyperspectral image classification method is proposed.First,the spatial-spectral features of the hyperspectral images are extracted by a graph convolutional network.Then a feature similarity metric-based virtual classifier is constructed and the discriminability of the hidden layer features is enhanced by reducing the divergence between the real and virtual classifiers.Finally,to reduce the influence of noisy pseudo-labels,a soft instance-level domain adaptation method is proposed.For each target-domain sample,the confidence coefficients are assigned to its corresponding positive and negative samples in the source domain,and a soft prototype contrastive loss is constructed and minimized to adapt two domains in an instance-level way.Experimental results on the Botswana,Kennedy Space Center,Pavia Center,Pavia University,and Hy RANK datasets verified the effectiveness of the proposed method.There are 23 figures,8 tables and 98 references in this thesis.