Classification Of Hyperspectral Images Based On Graph Model

With the increasing maturity of optics and photonics,hyperspectral technology has been significantly developed.Hyperspectral images,which consist of hundreds of contiguous bands and contain large amounts of useful information,can be easily acquired.Over the past few decades,hyperspectral image classification has played an important role in various fields,such as military target detection,vegetation monitoring,and disaster prevention and control.Up to now,diverse kinds of approaches have been proposed for classifying the pixels of a hyperspectral image.However,hyperspectral data labeling is expensive,so we choose semisupervised and unsupervised learning for classification.Considering that the hyperspectral data has a high spatial-spectral relevance,and how to express and learn this relevance is an important prerequisite for the follow-up effective analysis.The graph model,especially the graph deep network which has attracted much attention in recent years,is a powerful method to express the complex relationship between samples and can effectively express the spatial-spectral relevance of hyperspectral data.In this context,this paper mainly researches the classification of hyperspectral images based on graph model,which includes:A semi-supervised classification algorithm based on label constrained elastic network graph is proposed.The graph semi-supervised algorithm combines a small number of labeled samples with a large amount of unlabeled data to learn.The given label information is fully utilized,and the label constraint matrix is formed by the constraint propagation between the vertices.Then,for each vertex,the pixels that meets the label constraint are adaptively selected as the representation dictionary of the underlying vertex.The nearest neighbors of each vertex can be found by selecting the mostly related vertices in its elastic network representation upon the associated dictionary,and the semi-supervised classification of hyperspectral images is realized based on the constructed graph model.An unsupervised classification algorithm based on embedding representation of graph convolution network is proposed.It addresses the problems of hyperspectral data with high dimensions,insufficient training samples,and high spatial resolution.Firstly,the hyperspectral image is pre-segmented by the super-pixel segmentation algorithm,and the graph representation model of the super-pixel vertices is constructed by its elastic network representation.Then,this paper further studies the deep learning method under the graph model,the unsupervised graph convolution network is used to learn the embedding of each vertex in the graph,obtains the better low-dimensional feature representation,and finally realizes clustering through K-Means algorithm.