Graph-in-Graph Convolutional Network for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image classification (HSI) is crucial for analyzing Earth's surface.
• Existing Graph Convolutional Network (GCN) methods struggle to fully exploit object information due to feature aggregation.
• Superpixel-based approaches offer a promising alternative for HSI classification.

Purpose of the Study:

• To propose a novel Graph-in-Graph (GiG) model and GiG Convolutional Network (GiGCN) for enhanced HSI classification.
• To address the limitations of feature aggregation in existing GCN-based HSI classification methods.
• To improve the discriminability and robustness of HSI classification using limited labeled samples.

Main Methods:

• Developed a Graph-in-Graph (GiG) representation by constructing internal and external graphs from superpixels.
• Designed a GiG Convolutional Network (GiGCN) incorporating internal and external graph convolutions (EGC) for hierarchical feature extraction.
• Integrated ensemble learning to further enhance the robustness of the GiGCN model.

Main Results:

• The proposed GiGCN model demonstrated effectiveness and feasibility on four benchmark HSI datasets.
• Achieved improved classification accuracy, particularly with limited labeled samples.
• The GiG representation successfully captures both local and global characteristics of ground objects.

Conclusions:

• The GiGCN framework, based on the superpixel viewpoint, is a novel and effective approach for HSI classification.
• The method significantly improves feature extraction and integration across multiple scales.
• The study provides a valuable contribution to the field of hyperspectral image analysis and classification.