Unsupervised Hyperspectral Unmixing via Nonlinear Autoencoders

This work derives a novel unsupervised neural network-based scheme for unmixing hyperspectral pixels. A novel autoencoder structure was combined with a kernelization layer, mapping the mixed pixels in a higher dimensional space for easier separability, along with a novel cross-product layer to account for nonlinear mixing mechanisms. K-means clustering is utilized to estimate endmembers, and radial basis functions (RBFs) are employed to measure distances in a kernelized space to estimate abundances that provide a preliminary pixel unmixing stage to be enhanced by an autoencoder structure. A novel layer is introduced, which accounts for nonlinear mixing terms by forming pertinent cross products across the abundances of each mixed pixel. This enables accurate reconstruction of the mixed pixel adhering to different nonlinear mixing models while using the abundances and endmembers' estimates obtained via the decoding stage weights. The novel network structure is flexible since not only can it accommodate higher degree cross products of abundances and the corresponding endmember weights but it can also consider other mixing models, such as the postpolynomial nonlinear mixing (PPNM) model. Extensive testing across semisynthetic and real-world datasets shows that the proposed method, while being highly versatile in structure, also outperforms recent state-of-the-art unmixing methods.

Automated summary

This study introduces a novel unsupervised neural network-based scheme for unmixing hyperspectral pixels, incorporating autoencoder structure, kernelization layer, cross-product layer, K-means clustering, and radial basis functions. The proposed method is highly versatile and outperforms recent state-of-the-art unmixing methods in extensive testing across semisynthetic and real-world datasets.