I. Introduction
Hyperspectral image (HSI) is widely used for the area analysis of interest in Earth observation [1], and the classification of HSI is one of the most important applications in practical research [2], [3]. More and more HSI data have been provided for research with the upgrading of remote sensors. Hyperspectral data are composed of one spectral dimension and two spatial dimensions [4], [5], which provide a large number of signatures of objects for classification. However, hundreds of contiguous narrow bands contain not only abundant features but also redundancy and noise [6]. Moreover, the great majority of HSIs are unlabeled because of the expensive, difficult, and time-consuming labeling process. Therefore, the labeled data are limited in practical application, and only small-sized samples are available for supervised classification [7]–[9].