I. Introduction
Since hyperspectral images (HSIs) usually have hundreds of channels that are the continuous and narrow frequency bands of reflected radiation from a specific surface, HSIs have been widely applied in many areas [1]–[5]. This article mainly focuses on the classification task that aims at assigning every pixel in HSIs to a certain land-cover type. There are two major characteristics of HSIs that must be considered for HSI classification: 1) abundant spectral information that makes the accurate pixelwise classification possible and 2) strong spatial correlation that reflects the spatial distribution of different land types.