1. Introduction
Hyperspectral remote sensing combines spatial geometry imaging with spectroscopy. In recent years, it has been widely used in the fields of environmental monitoring, precision agriculture and defense security. The development of high spectral-spatial resolution make the hyperspectral data volume increase sharply, the high-dimensional data contains redundant information which may cause computational complexity and lead to the Hughes phenomenon, so, it is very necessary to reduce the amount of data and save resources. The feature extraction and band selection are the two main methods for reducing the dimension of HSIs. Recently, many band selection methods (BS) are proposed because it can preserve the original features of remote sensing data, such as, a Particle Swarm optimization (PSO)-based band selection is proposed in [1], Fishers linear discriminant analysis (LDA) [2], a supervised method based on the rough set theory for hyperspectral band selection is presented in [3], Wang et al. [4] introduce a manifold ranking (MR)-based band selection method. And also, a lot of methods based on spectral information have been proposed in[5–8]. But most band selection methods have disadvantages in the identification and discrimination of classes.