CEGAT: A CNN and enhanced-GAT based on key sample selection strategy for hyperspectral image classification

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image classification (HSIC) benefits from deep learning methods like Convolutional Neural Networks (CNNs) and Graph Convolutional Networks (GCNs).
• A significant challenge in HSIC using these models is the scarcity of labeled training samples.
• Existing methods struggle to achieve optimal performance with limited data, hindering practical applications.

Purpose of the Study:

• To propose a novel deep learning architecture, the CEGAT (CNN and Enhanced Graph Attention Network), to improve HSIC performance with limited labeled samples.
• To develop a key sample selection strategy to enhance the efficiency of training on small datasets.
• To reduce spectral redundancy and effectively capture spatial-spectral features for more robust classification.

Main Methods:

• A double branch fusion network combining CNN and an enhanced graph attention network (CEGAT) was developed.
• Key components include a linear discrimination of spectral inter-class slices (LD_SICS) module for spectral redundancy reduction and a spatial spectral correlation attention (SSCA) module.
• The graph attention (GAT) branch utilizes super-pixel segmentation and an enhanced graph attention (EGAT) module, coupled with a key sample selection (KSS) strategy.

Main Results:

• The proposed CEGAT model demonstrates superior classification performance compared to state-of-the-art methods, particularly under conditions with limited labeled samples.
• The LD_SICS and SSCA modules effectively handle spectral redundancy and extract salient spatial-spectral features.
• The KSS strategy significantly enhances the network's ability to learn from scarce labeled data.

Conclusions:

• The CEGAT network provides an effective solution for hyperspectral image classification challenges posed by limited labeled data.
• The integration of CNN, enhanced GAT, and a key sample selection strategy offers a promising direction for future HSIC research.
• The proposed method achieves high classification accuracy, making it suitable for real-world applications with data constraints.