Self-spectral learning with GAN based spectral-spatial target detection for hyperspectral image.

Area of Science:

• Remote Sensing
• Signal Processing
• Machine Learning

Background:

• Hyperspectral target detection faces challenges with single-aspect limitations and redundant spectral band information.
• Existing methods often struggle with effective background suppression and target saliency.

Purpose of the Study:

• To propose a novel spectral-spatial target detection (SSTD) framework in deep latent space.
• To introduce self-spectral learning (SSL) for unsupervised hyperspectral feature extraction, aiming for background suppression and target saliency.
• To identify optimal spectral band subsets without prior knowledge.

Main Methods:

• Developed a spectral-spatial target detection (SSTD) framework utilizing deep latent space and self-spectral learning (SSL).
• Employed a spectral generative adversarial network (GAN) for feature extraction and background suppression.
• Introduced a structure-to-structure selection rule for optimal spectral band subset generation based on structural, contrast, and luminance similarity.
• Combined spatial detection on fused latent features with spectral detection on selected bands and target signatures.

Main Results:

• The proposed SSTD framework demonstrated superior target detection performance compared to conventional methods (CSCR, ACE, CEM, hCEM, ECEM).
• The self-spectral learning approach effectively suppressed background noise and enhanced target saliency.
• The structure-to-structure selection rule successfully identified optimal spectral band subsets for specific targets.

Conclusions:

• The novel SSTD framework offers a new approach to hyperspectral target detection through self-spectral learning.
• The method effectively reduces redundant spectral information and improves detection accuracy.
• This work provides a practical way to identify sensitive spectral bands for target detection.