|
Deep autoencoders have recently been applied to blind hyperspectral unmixing task to estimate endmembers and their corresponding abundances simultaneously. The objective of an original autoencoder is to reconstruct an input data matrix unsupervisedly with an encoder network and a decoder network. For the purpose of spectral unmixing, the activations of the final layer of the encoder and the weights of the decoder form abundances and endmember signatures, respectively; constraints, e.g., abundance non-negativity and abundance sum-to-one, can be imposed. In this paper, we present a novel regularization technique for autoencoder-based hyperspectral unmixing. The basic idea is the inclusion of a generative adversarial network (GAN) joint training objective to condition the decoder to generalize to unseen abundance mixtures. In addition to regularizing the endmember weights of the decoder, this approach has the benefit of explicitly modeling the prior distribution of hyperspectral pixels for a given scene as the abundance output of the generator. The benefit of the proposed strategy is evaluated on synthetic and real data sets, demonstrating that it can produce endmember estimates closer to the ground truth.
|
