Robust Hyperspectral Unmixing With Correntropy-Based Metric.

Area of Science:

• Remote Sensing
• Signal Processing
• Data Analysis

Background:

• Hyperspectral unmixing is vital for analyzing hyperspectral imagery but challenging in unsupervised settings with unknown endmembers and abundances.
• Noise in spectral bands significantly complicates unsupervised hyperspectral unmixing tasks.
• Existing methods often struggle with noise robustness and accurate abundance estimation.

Purpose of the Study:

• To develop a robust unsupervised hyperspectral unmixing model that addresses noise and ensures physical meaningfulness.
• To incorporate a correntropy-based metric and sparsity prior for improved unmixing performance.
• To validate the model's effectiveness on both synthetic and real-world hyperspectral datasets.

Main Methods:

• A novel unsupervised hyperspectral unmixing model utilizing a correntropy-based metric.
• Imposition of nonnegativity constraints on endmembers and abundances for physical interpretability.
• Integration of a sparsity prior to regularize abundance distributions.
• Application of a half-quadratic optimization technique to iteratively reweighted nonnegative matrix factorization with sparsity.

Main Results:

• The model adaptively down-weights noisy spectral bands and emphasizes cleaner ones.
• Sparsity constraints naturally lead to the generation of sparse abundance maps.
• Experimental results on synthetic and real data show superior performance compared to state-of-the-art methods.
• The proposed method demonstrates enhanced robustness in the presence of spectral noise.

Conclusions:

• The developed correntropy-based unsupervised hyperspectral unmixing model offers a robust solution for noisy data.
• The combination of nonnegativity constraints, sparsity prior, and adaptive weighting effectively improves unmixing accuracy.
• This approach provides a significant advancement for various hyperspectral data applications requiring reliable endmember and abundance estimation.