Spectral Variability Aware Blind Hyperspectral Image Unmixing Based on Convex Geometry.

Area of Science:

• Remote Sensing
• Signal Processing
• Computer Vision

Background:

• Hyperspectral image unmixing is crucial for interpreting hyperspectral data.
• Traditional linear unmixing methods often overlook intra-class spectral variability.
• Existing algorithms based on convex geometry may lack robustness when variability is present.

Purpose of the Study:

• To analyze the robustness of classical hyperspectral unmixing tools under spectral variability.
• To propose an improved unmixing chain addressing limitations of linear models.
• To validate the extended linear mixing model for hyperspectral data.

Main Methods:

• Analysis of intrinsic dimensionality, endmember extraction, and pixel purity under a prescribed variability model.
• Development of an integrated unmixing chain based on an extended linear mixing model.
• Validation using simulated and real-world hyperspectral datasets.

Main Results:

• Classical convex geometry-based unmixing tools show limitations when intra-class spectral variability is significant.
• The proposed extended linear mixing model and integrated unmixing chain demonstrate improved performance.
• The approach effectively handles spectral variability in both simulated and real hyperspectral data.

Conclusions:

• Traditional hyperspectral unmixing techniques require adaptation to account for spectral variability.
• The proposed extended linear mixing model offers a more robust framework for hyperspectral image analysis.
• The integrated unmixing chain provides a promising solution for accurate unmixing in the presence of intra-class variability.