Pixel-Level and Global Similarity-Based Adversarial Autoencoder Network for Hyperspectral Unmixing

Abstract

Hyperspectral unmixing is a critical task in remote sensing, enabling the decomposition of hyperspectral data into their constituent endmembers and abundances. The loss of the traditional unmixing algorithm based on deep learning typically depends on reducing the discrepancy between the original and reconstructed hyperspectral image. However, during the training process, the loss feedback method is relatively simple, resulting highly random unmixing results. Moreover, spatial feature extraction can effectively improve the unmixing effect, but existing spatial feature extraction methods in hyperspectral unmixing still have significant room for improvement. To address these challenges, we propose a novel adversarial autoencoder unmixing network considering pixel-level and global similarity measurements based on a Wasserstein generative adversarial network (WGAN) and a U-shaped transformer-enhanced architecture. The WGAN ensures stable gradient updates through a gradient penalty, maintaining Lipschitz continuity, while the U-shaped network with Swin transformer blocks captures both local and global spatial features. Experiments were conducted on synthetic and real-world hyperspectral datasets. Our method outperformed state-of-the-art approaches, achieving improvement in root mean square error and spectral angle distance (SAD). The SAD is a metric that quantifies the angular difference between the true and estimated endmember spectra, our method improves the mean SAD by at least 8.7% compared to competing algorithms, representing an enhancement in unmixing performance. Notably, the method demonstrated superior robustness in low signal-to-noise ratio scenarios, maintaining high unmixing accuracy. These results highlight the potential of our approach to advance unmixing research by addressing both pixel-level and global similarity constraints, providing a new way for hyperspectral unmixing.