Hyperspectral image unmixing using autoencoder cascade

Abstract

Hyperspectral image unmixing is the process of estimating pure source signals (endmemebers) and their proportions (abundances) from highly mixed spectroscopic images. Due to model inaccuracies and observation noise, unmixing has been a very challenging problem. In this paper, we exploit the potential of using autoencoder to tackle the unmixing challenges. Two important facts are considered in the algorithm: first, the observation noise in the hyperspectral image generally exists and largely affects the unmixing results; second, the mixing process contains sparsity priori which should be considered to assist the endmember extraction. The proposed autoencoder cascade concatenates a marginalized denoising autoencoder and a non-negative sparse autoencoder to solve the unmixing problem which implicitly denoises the observation data and employs the self-adaptive sparsity constraint. The algorithm is tested on a set of synthetic mixtures and a real hyperspectral image. The experimental results demonstrate the proposed algorithm outperforms several advanced unmixing approaches in highly noisy environment.