Target Detection With Unconstrained Linear Mixture Model and Hierarchical Denoising Autoencoder in Hyperspectral...

Area of Science:

• Remote Sensing
• Signal Processing
• Machine Learning

Background:

• Hyperspectral imagery offers detailed spectral information but faces challenges in target detection due to spectral variability and noise.
• Intraclass dissimilarity and interclass similarity in hyperspectral data hinder accurate identification of subtle nuances.
• Atmospheric effects, illumination variations, and sensor noise introduce significant interference in hyperspectral target detection.

Purpose of the Study:

• To propose a novel hyperspectral target detection method robust to spectral inconsistencies and without strict data distribution assumptions.
• To enhance the accuracy of target detection by effectively suppressing background noise and enhancing target signatures.
• To develop an efficient method for extracting reliable background and target samples for the detection process.

Main Methods:

• A deep-learning-based hierarchical denoising autoencoder is employed to reduce spectral interference.
• A two-step subspace projection is utilized for background suppression and target enhancement.
• A spatial-spectral unified endmember extraction method is developed for generating representative samples.

Main Results:

• The proposed method demonstrates superior performance compared to state-of-the-art detection techniques.
• Validation on four real-world hyperspectral datasets confirms the effectiveness and efficiency of the detector.
• The method successfully alleviates spectral inconsistencies, improving target identification accuracy.

Conclusions:

• The novel deep learning-based approach provides an effective solution for hyperspectral target detection.
• The method's robustness to spectral variability and noise makes it suitable for real-world applications.
• The developed techniques offer advancements in analyzing complex hyperspectral data for substance identification.