Nonconvex-sparsity and Nonlocal-smoothness Based Blind Hyperspectral Unmixing

Area of Science:

• Remote Sensing
• Signal Processing
• Computer Vision

Background:

• Blind hyperspectral unmixing (HU) is essential for analyzing hyperspectral images (HSI).
• Nonnegative matrix factorization (NMF) is a popular approach for HU, often utilizing abundance sparseness and structural smoothness.
• Existing NMF-based HU methods overlook the non-local smoothness property inherent in natural HSIs.

Purpose of the Study:

• To introduce and investigate the non-local smoothness prior for blind hyperspectral unmixing.
• To develop novel NMF-based HU models incorporating non-local total variation (NLTV) and generalized non-local HSI TV (NLHTV) regularizers.
• To enhance abundance map sparseness using a log-sum regularizer within the NMF framework.

Main Methods:

• Proposed novel blind HU models: NLTV-LSRNMF and NLHTV-LSRNMF.
• Incorporated NLTV and NLHTV regularizers to capture non-local smoothness in HSI.
• Utilized a non-convex log-sum regularizer for abundance sparseness.
• Developed an efficient algorithm based on alternative optimization strategy (AOS) and alternating direction method of multipliers (ADMM) for model solving.

Main Results:

• The proposed NLTV-LSRNMF and NLHTV-LSRNMF models demonstrated superior performance in blind HU.
• Experiments on simulated and real HSI datasets confirmed the effectiveness of the non-local smoothness prior.
• The developed algorithm efficiently solved the proposed non-convex NMF models.

Conclusions:

• Incorporating non-local smoothness priors significantly enhances blind hyperspectral unmixing.
• The proposed NLTV/NLHTV-LSRNMF models offer a more robust and accurate approach to HSI decomposition.
• The study highlights the importance of exploring intrinsic HSI structures for improved data exploitation.