Hyperspectral image is a cube data with rich spatial information and spectral information.Because the hyperspectral sensors collect imagery simultaneously in hundreds of narrow and contiguously spaced spectral bands,the hyperspectral data generally have tens or hundreds of bands.Although these bands provide rich image information,it also brings a lot of problems and challenges,such as huge amount of data,high computing costs and storage costs,time-consuming.Therefore,it is very necessary to reduce the dimension of hyperspectral data effectively.In this paper,we mainly study the dimension reduction method of hyperspectral data,and propose solutions from the two aspects of band selection and feature extraction to realize the dimensionality reduction of the original hyperspectral data.The main work of this paper is summarized as follows:(1)We proposed a band selection method of hyperspectral image based on joint sparse representation.Because sparse representation of a single band,ignoring the similarity between the bands,the similar bands may be represented by different bands.In order to avoid this problem,we put forward the method of this chapter.The similarity between the bands was measured by calculating the spectral angle,each band finds a few bands that are closest to it as a similar cluster,using the whole hyperspectral image as a dictionary to joint sparse representation.Find the band that contributes the most to the reconstructed similar band group.At last,the band appears most frequently was selected.The method is validated on the real data sets,and the result is superior to the classical band selection methods.(2)We proposed a band selection method of hyperspectral image based on improved sparse subspace clustering.The relationship between the bands is measured by sparse representation and low-rank representation from the two aspects of global and local,and then the similarity between the transform domain bands is measured by spectral information divergence(SID).Next,spectral clustering is used to label the similar bands into several groups.A band with a largest entropy of groups as the representative band is selected.The method proposed in this chapter not only used the sparse representation to measure the local linear structure of the original band,but the low rank is a good measure of the global relationship between the bands,what's more,it is robust to the noise.Themethod is validated on several real data sets,and the effect is superior to the classical band selection methods.(3)We proposed a dimensionality reduction algorithm of hyperspectral image based on semi-supervised sparse graph construction.Because the semi-supervised approach has the advantage of using a small number of training samples and a large number of unlabeled samples to improve the performance of the classifier.The sparse graph can preserve the discriminant information of the samples very well,besides,it is very easy to construct the sparse graph,the sparse coefficients obtained from the labeled samples are used to construct the inner and interclass graphs.Since the KNN graph can measure the local information of the sample very well while reduce the computational complexity,the KNN graph is used to measure the essential structural attributes of the unlabeled samples.The method is validated on several real data sets,and the effect is superior to other methods. |