Deep Spatial-Spectral Joint-Sparse Prior Encoding Network for Hyperspectral Target Detection.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Deep learning excels in hyperspectral target detection but often lacks interpretability due to black-box architectures.
• Existing methods struggle to explicitly incorporate domain knowledge, limiting understanding of feature extraction processes.

Purpose of the Study:

• To propose a novel deep learning network, the Joint-Spatial-Spectral Prior Encoding Network (JSPEN), for interpretable hyperspectral target detection.
• To embed domain knowledge into a neural network architecture for improved accuracy and explicit interpretability.

Main Methods:

• Developed an adaptive joint spatial-spectral sparse model (AS2JSM) to capture spatial-spectral correlations in hyperspectral images (HSIs).
• Designed an optimization algorithm and simulated its iterative process within the JSPEN architecture, ensuring each module has a clear function.
• Enabled end-to-end training of JSPEN to automatically learn sparse properties of HSIs for accurate background and target feature characterization.

Main Results:

• JSPEN demonstrates explicit interpretability by mapping network modules to optimization algorithm steps.
• The method effectively mines spatial-spectral correlations, leading to improved data representation accuracy.
• Experimental results confirm the effectiveness and high accuracy of the proposed JSPEN for hyperspectral target detection.

Conclusions:

• JSPEN offers a novel, interpretable approach to hyperspectral target detection by integrating domain knowledge.
• The network's design facilitates intuitive analysis and understanding of its operational mechanisms.
• The proposed method achieves superior performance in accuracy and effectiveness for hyperspectral target detection tasks.