Endmember-Guided Unmixing Network (EGU-Net): A General Deep Learning Framework for Self-Supervised Hyperspectral...

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Hyperspectral unmixing (HU) models struggle with spectral variability (SV) and physically meaningful endmember extraction.
• Existing linear/nonlinear models show limitations in data fitting, reconstruction, and sensitivity to SVs.

Purpose of the Study:

• To develop a general deep learning (DL) approach for HU that addresses limitations of current methods.
• To enhance the accuracy and interpretability of hyperspectral unmixing by incorporating endmember properties.

Main Methods:

• Developed the endmember-guided unmixing network (EGU-Net), a two-stream Siamese deep network.
• EGU-Net utilizes pure endmembers to guide weight correction in the unmixing network, incorporating nonnegativity and sum-to-one constraints.
• The framework supports both pixelwise spectral unmixing and spatial-spectral unmixing using convolutional operators.

Main Results:

• EGU-Net demonstrated superior performance compared to state-of-the-art unmixing algorithms on three datasets.
• The network effectively generalized various spectral variabilities and extracted physically meaningful endmembers.
• Experimental results validated the effectiveness of the proposed spatial-spectral unmixing approach.

Conclusions:

• EGU-Net offers a robust and generalizable deep learning solution for hyperspectral unmixing.
• The endmember-guided approach significantly improves unmixing accuracy and interpretability.
• The framework's adaptability to spatial information modeling opens new avenues for hyperspectral data analysis.