I. Introduction
HYPERSPECTRAL imaging has been a popular topic in the remote sensing community [1]. Recent advances in hyperspectral imaging have been successively applied to many real-world applications such as geology, ecology, agriculture, mineral mapping, land cover classification, chemical, environmental monitoring, and military defense [2]. Opposite to conventional multispectral images that contain tens of discrete bands with a broad bandwidth around 100–200 nm, the hyperspectral images usually consist of hundreds of contiguous bands with fine spectral resolutions that are approximately 10 nm. Owing to the wealth of spectral information collected from the advanced hyperspectral imagine sensors, the hyperspectral image systems have greater potential in data exploration [3]. Although the recently developed sensors may achieve a considerably better resolution in the spatial domain, the ground spatial distance (i.e., the spatial distance covered by a pixel) is still high. As a result, a pixel in the hyperspectral image usually captures the mixture of spectral information of different substances. Classifying or quantifying a so-called “mixed pixel” (or subpixel) is a crucial topic in hyperspectral image processing, which is known as spectral unmixing (SU) [4].