A manifold-based adversarial autoencoder with Fourier convolution for hyperspectral unmixing

Abstract

Hyperspectral unmixing aims to decompose each subpixel into their pure endmembers and the corresponding proportions. But existing deep autoencoder-based hyperspectral unmixing methods often suffer from obstacles like endmember variability, local respective fields and insufficient use of inner structure. In the manuscript, we build a manifold-based Fourier adversarial autoencoder which regards generative adversarial mechanism as a utilization of prior information. This method combines manifold learning with adversarial autoencoder in order to promote the performance of hyperspectral unmixing. Specifically, firstly, in order to preserve local manifold structure, we add a discriminator to the autoencoder which uses the covariance matrices of a superpixel as real samples while covariance matrices of the abundance as fake samples; secondly, we add a regularization term of Laplacian eigenmap at the loss of autoencoder to in-depth abbreviate autoencoder solution space; thirdly, Fast Fourier Convolution modules are used to enhance multi-scale information fusion. At last, comparative experiments are conducted on three popular datasets, including Jasper, Urban4 and Samson, to validate the effectiveness of the proposed method.