Entropic Descent Archetypal Analysis for Blind Hyperspectral Unmixing

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Hyperspectral unmixing aims to identify constituent materials within a pixel.
• Traditional methods often require pure pixels, which are rare in real-world data.
• Archetypal analysis offers a promising framework for unmixing by representing endmembers as combinations of image pixels.

Purpose of the Study:

• Introduce a new algorithm for blind hyperspectral unmixing using archetypal analysis.
• Address the limitation of requiring pure pixels in existing methods.
• Develop an efficient and robust unmixing solution.

Main Methods:

• Developed a novel algorithm based on archetypal analysis for hyperspectral unmixing.
• Employed an entropic gradient descent strategy for improved solution quality and GPU implementation.
• Incorporated an ensembling mechanism and model selection for hyper-parameter robustness.

Main Results:

• The proposed method outperforms traditional archetypal analysis and state-of-the-art matrix factorization and deep learning techniques.
• Achieved efficient GPU implementations due to the entropic gradient descent strategy.
• Demonstrated robustness to hyper-parameter choices through ensembling and model selection.

Conclusions:

• The novel archetypal analysis algorithm provides a superior approach to blind hyperspectral unmixing.
• The method is efficient, robust, and does not require pure pixels.
• An open-source PyTorch implementation is available for further research and application.