I. Introduction
Hyperspectral image (HSI) contains hundreds of channels captured by imaging spectrometer under narrow and continuous electromagnetic spectrums covering from visible light to infrared light [1]–[3]. In HSI intelligent interpretation, each pixel is considered as a high-dimensional vector, which provides rich characteristic information of target objects [4]–[6]. Since different target objects contain different characteristic information, the type of each pixel can be recognized according to the differences of high-dimensional vectors [7], [8]. However, there is a large amount of redundant information in high-dimensional HSI data, which will have an adverse effect on the discriminant performance of features for HSI classification task [9]–[11].