Deep NMF and Autoencoder: A Comparative Analysis for Hyperspectral Unmixing Using Prisma Real Images

In hyperspectral data, mixed pixels are frequent due to the low-medium spatial resolution of the imaging spectrometer, or to intimate mixing effects. Hence the process of blind hyperspectral unmixing, which separates the pixel spectra into a collection of spectral signatures and a set of fractional abundances, is a mandatory task in hyperspectral image processing. In this study, among models capable of performing linear spectral unmixing, we present a comparative analysis performed on real data acquired from the PRISMA (PRecursore IperSpettrale della Missione Applicativa) hyperspectral satellite between deep nonnegative matrix factorization and artificial neural network autoencoder based methods.