Band selection-based collaborative representation for anomaly detection in hyperspectral images

Hyperspectral images with high spectral resolution contain hundreds and even thousands of spectral bands and convey abundant spectral information to distinguish subtle spectral differences, especially between similar materials, thereby providing unique advantage for target detection. At the same time, hyperspectral images cause large number of bands and redundant information between adjacent bands. The high-dimensional data structure frequently reduces the separability between the anomaly and background classes of hyperspectral images. This study proposed a Band Selection-based Collaborative Representation (BSCR) method for hyperspectral anomaly detection to overcome these shortcomings. In BSCR, we first selected hyperspectral bands via an optimal clustering framework and obtained a set of representative bands. The separability between the anomaly and background classes enhanced. Then, we reconstructed each pixel in the image through collaborative representation. We obtained a large residual when reconstructing the anomaly pixel via collaborative representation and achieved a large output value for an anomaly pixel because of the enhanced separability between the anomaly and background classes, thereby improving the separation of the anomaly class from the background class. The proposed algorithm was tested on synthetic and real hyperspectral images. The experimental results of three hyperspectral images show that the proposed BSCR demonstrates outstanding detection performance in the receiver operating characteristic curve, area under the curve value, and separability map compared with other state-of-the-art detectors. BSCR has an improved discriminative ability to separate the target from the background. Compared with several traditional anomaly detection algorithms, BSCR enhances the separability between the target and background through band selection and can effectively separate the anomaly class from the background class in hyperspectral images. Algorithms, such as PCAroCRD, which also conduct collaborative representation after the dimensionality reduction of hyperspectral images, can remove certain anomaly pixels and make the background construction stable. However, the dimensionality reduction mode of such algorithms will change the original signal of the image, thereby making it smooth and increasing the difficulty to distinguish between anomaly and background classes. The dimensionality reduction method of the band selection in BSCR can enhance the separability between the target and background without changing the original signal of the image, thereby making it easy to distinguish when detecting anomaly pixels and enabling sufficient discriminative ability to separate the target from the background. © 2020, Science Press. All right reserved.