Graph Attention Convolutional Autoencoder-Based Unsupervised Nonlinear Unmixing for Hyperspectral Images

Hyperspectral unmixing has received increasing attention as a technique for estimating endmember spectra and fractional abundances of land covers. Encoding high-dimensional hyperspectral data into a low-dimensional latent space to generate reasonable abundances, autoencoder (AE) has shown its great potential and attractive advantages in spectral unmixing. AEs decode abundances back to spectra, which can effectively reflect the general spectral mixing process. However, most existing AE-based unmixing methods often do not fully exploit the spatial information of hyperspectral images, hindering the improvement of unmixing accuracy. This paper proposes a graph attention convolutional autoencoder architecture for hyperspectral unmixing. By incorporating graph attention convolution into AE, the proposed method performs better in leveraging both long-range and short-range spatial information of hyperspectral images. Accurate abundances with global and local spatial consistency can be efficiently learned by the network. Moreover, the decoder is further improved based on the post-polynomial nonlinear mixing (PPNM) model to make the network have stronger physical interpretability to deal with the issue of nonlinear blind unmixing. Experimental results indicate that the proposed method has good unmixing performance. It can reduce the root mean squared error of estimated abundances for synthetic data by over 10% compared to other methods. In experiments with real hyperspectral data, the difference between its unmixing results' accuracy and the best is less than 5%.