I. Introduction
Hyperspectral imaging technologies occupy crucial positions in the geoscience field, providing abundant research directions [1], [2]. Compared with classical RGB images, hyperspectral images comprise more discriminant information delivering complex topographic features; however, this is at the expense of hundreds of spectral redundant bands [2], [3]. As an effective solution to these problems, feature extraction maps original spectral features into specific latent spaces by transformation and then generates latent embeddings with the higher discriminant ability and less storage space [4], [5].