A Self-Supervised Hyperspectral Unmixing Framework Based on Prior Self-Learning and Irrelevant Endmember Degradation

Abstract

Hyperspectral unmixing is a crucial technique in hyperspectral remote sensing image processing, aiming to separate pure material spectra (endmembers) and their corresponding proportions (abundances) from mixed pixels. Existing nonnegative matrix factorization (NMF) methods suffer from poor interpretability due to the lack of physical priors for endmembers, while sparse unmixing (SU) methods face bottlenecks caused by large scale, highly coherent spectral libraries and the difficulty in imposing the abundance sum to one constraint. To integrate the advantages of both approaches, this article proposes a self-supervised hyperspectral unmixing framework based on prior self-learning and irrelevant endmember degradation. The method couples NMF and SU within a unified optimization framework, establishes a nonlinear mapping between estimated endmembers and library atoms via kernel functions, and designs an abundance map activity measure with a progressive degradation strategy to actively eliminate irrelevant library atoms. A bidirectional abundance prior transfer mechanism is constructed, enabling SU to convey spatial piecewise smoothness to NMF and NMF to convey the sum to one constraint to SU. In addition, an analytical sign-prediction FCLS method is developed, obtaining a closed-form solution via linear system inversion, rigorously satisfying both nonnegativity and sum to one constraints without iterations or nondifferentiable approximations. An ADMM based algorithm is developed to solve the model. Numerous experiments were conducted on both synthetic and real datasets, comparing them with state-of-the-art methods. The experimental results demonstrate the effectiveness and superiority of the proposed method.