I. Introduction
Hyperspectral sensors can acquire very rich information with fine spectral resolution and spatial resolution, which makes it possible to implement an accurate classification of land covers. Whereas, the fine spectrum also produces problems for the land-cover classification of HSIs. The high dimensionality of spectral information and few labeled samples are two common challenges in HSI classification, because they can produce the Hughes phenomenon [1]. However, with the development of machine learning methods [2], [3], [4], especially deep learning ones, such as central attention network (CAN) [5], multiarea target attention (MATA) [6], SOT-NET [7], MDA-NET [8], and so on, the Hughes phenomenon was alleviated using some feature extraction techniques. Consequently, the classification performance of HSIs was enhanced to some extent. But, when few labeled samples are available, the improved classification accuracy is very limited [9]. It is well known that labeling samples is difficult and demanding, because it requires a lot of manpower and technology. Thus, small-sample classification of HSIs is an important and urgent task that has attracted widespread attention in HSI processing.