Hyperspectral target detection based on graph sampling and aggregation network.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Hyperspectral imaging (HSI) offers rich spectral information but presents challenges due to complex pixel structures.
• Traditional target detection methods struggle with the intricate spatial-spectral characteristics of HSI data.
• Graph sampling aggregation networks have been underexplored for hyperspectral target detection.

Purpose of the Study:

• To introduce a novel graph sampling aggregation network for enhanced hyperspectral target detection.
• To leverage graph-based learning for improved feature representation and extraction from HSI.
• To address limitations in handling complex spatial-spectral information in HSI.

Main Methods:

• Feature vector extraction via Principal Component Analysis (PCA) to build adjacency matrices.
• Convolutional operations on HSI using sparse matrix multiplication for node feature propagation.
• Target data extraction using residuals and constraint energy minimization for detection.

Main Results:

• The proposed model achieved an average detection accuracy exceeding 99.8% across seven HSI datasets.
• Demonstrated superior performance compared to existing hyperspectral target detection models.
• Exhibited remarkable adaptability and robustness across datasets with diverse characteristics.

Conclusions:

• The graph sampling aggregation network is highly effective for hyperspectral target detection.
• The model's ability to learn node representations facilitates robust feature extraction.
• The approach offers a promising solution for accurate and adaptable HSI target detection.