A Study Of Sparse Representation And Low-Rank Representation For Hyperspectral Band Selection

Hyperspectral remote sensing has become an important remote sensing method since the 1980 s. Obtained by the spectral imager, hyperspectral images have a high spatial resolution and spectral resolution than conventional remote sensing image, which makes hyperspectral images be widely used in many areas. But there are hundreds of bands in the hyperspectral image, which brings great difficulties for hyperspectral image processing due to the redundant information between the band and band.Therefore, how to reduce the dimensions of hyperspectral data to improve the efficiency and effectiveness of data processing becomes an urgent problem. Feature extraction and feature selection are the mainly methods for data dimensionality usually. The physical meaning of specific bands need to retained for hyperspectral data, so feature selection is a good method for hyperspectral data. It also calls band selection when feature selection is used in hyperspectral data. Based on research on existing methods, this paper proposes some new methods for band selection in hyperspectral images by considering the characteristics of hyperspectral images and combining the related compressed sensing technology. The main contents of this paper are as follows:(1) A hyperspectral image band selection method is proposed based on semi-supervised sparse representation subject to a discriminant constraint. In the selection process, the method takes into account the class separability that can be measured by the intra-class and inter-class distances, Therefore, the method adds the item of discriminant constraint from the intra-class and inter distances to the band spare representation model with the hyperspectral data adopted as the dictionary. With the band subsets selected using the method, pixels of the same class are more compact and pixels of different classes are more dispersed, which is more conducive to classification.(2) A hyperspectral band selection method is proposed using low-rank representation based band clustering. This method aims to select a band subset with small internal correlation and large information amount. To achieve the aim, the method implements low-rank representation for the hyperspectral to remove the intervention of the noise bands. Then with the idea of hierarchical clustering, the method clusters the bands according to the low-rank representation coefficients and selects from each cluster the representative band as the final bands. This method for band selection is self-adaptive and requires no human intervention in determination of the number of bands.(3)A best-match dictionary based hyperspectral band selection method is proposed.This method aims to select the band subset that best represents all the original bands.To achieve the aim, the method uses the selected band subset as the dictionary and implements the graph Laplace constrained group sparse representation for hyperspectral data. Then the method adopts the residual of the representation as the assessment criterion function and uses the function to repeatedly update the dictionary until the dictionary is namely the best-match dictionary and is also the finally selected band subset. This band subset basically contains all useful information of the hyperspectral data and therefore can obtain a good classification result.