The Research On Sparse Representation For Hyperspectral Image Classification

Hyperspectral images(HSIs)have been widely used in various fields over the past two decades.However,HSI classification is an important problem.Several researchers have proposed good classification methods.The HSI classification is still a difficult problem because of the high dimensionality of HSIs and the limited availability of training samples.In this paper,the development of HSI and the existing methods are reviewed,and then the applications of sparse representation in HSI classification are discussed.The main contents of the paper are as follows:(1)Joint Sparse Representation Based on Random Subspace for Hyperspectral Image ClassificationThe high dimensionality of HSIs and the limited availability of training samples make such classification methods affected by the Hughes phenomenon.In this paper,we present a new framework for the HSI classification to reduce or eliminate the impact of Hughes phenomenon.Our method mainly involves the following four steps.First,several features are randomly selected from the original spectrum feature space to generate multiple random subspaces.The HSI is divided from a superpixel segmentation technique to acquire a seg-mentation map.Second,the classification method that combines spatial information with spectral information is used to obtain the results in each subspace.We adopt simultaneous orthogonal matching pursuit(SOMP).Third,the ensemble method is employed to select the optimal result.Fourth,the result and the segmentation map are combined by the majority voting to obtain the final result.Comparisons between our method and other classification methods in three real experimental datasets proves that our method yields good results in terms of classification accuracy and robustness.(2)Sparse Representation Based on Dictionary Learning for the Hyperspectral Image ClassificationThere are many sparse representation methods that directly treated training sample as a dictionary.Researchers have shown that more robust dictionaries can be learned from the training set,but the dictionary learning method only takes spectral information into account.As is known by the existing methods,the introduction of spatial information can effectively improve the classification accuracy.Some methods,when introducing spatial information,use a rectangular neighborhood,and it assume that the pixels in the neighborhood are com-posed of the same material.This assumption does not hold in the edge position that is com-posed of different elements of the adjacent position of the pixel.Therefore,a hyperspectral image segmentation method is needed to divide the hyperspectral images into several non-overlapping groups to make the pixels in each group of same material as much as possible.This paper presents a dictionary learning-based method for the hyperspectral image classifi-cation.First,the proposed method divides the hyperspectral image into several superpxiels,and then obtains the dictionary by the context dictionary learning method on the superpixel.Then,we get the sparse representation of each sample in the dictionary.Finally,the support vector machine classifier is constructed on the obtained sparse representation,and predicts the test samples.Comparisons between our method and other classification methods in three real experimental datasets proves that our method yields good performance.