Generation of artificial training data for spectral unmixing by modelling spectral variability using Gaussian random variables

Abstract

A stochastic method how artificial training data for spectral unmixing can be generated from real pure spectra is presented. Since the pure spectra are modelled as Gaussian random vectors, spectral variability is also considered. These training data can in turn be used to train an artificial neural network for spectral unmixing. Non-negativity and sum-to-one constraints are enforced by the network architecture. The approach is evaluated using real mixed spectra and achieves promising results with the used datasets.