Graph Convolutional Network Using Adaptive Neighborhood Laplacian Matrix for Hyperspectral Images with Application to...

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Hyperspectral image classification is crucial for various applications.
• Existing graph convolutional network (GCN) methods often struggle to effectively integrate spatial and spectral information.
• Adaptive feature extraction is needed to capture complex patterns in hyperspectral data.

Purpose of the Study:

• To propose an adaptive neighborhood graph convolutional network (AN-GCN) for improved hyperspectral image classification.
• To enhance the integration of spatial and spectral information using statistical variance.
• To demonstrate the efficacy of AN-GCN in handling variable neighborhood sizes and discriminating subtle differences.

Main Methods:

• Developed an adaptive neighborhood aggregation method based on statistical variance.
• Integrated spatial-spectral information into the adjacency matrix for a single-layer GCN.
• Employed class-conditioned adaptive neighborhood selection criteria.
• Validated the approach on Indian Pines, Houston University, and Botswana Hyperion datasets.

Main Results:

• The proposed AN-GCN significantly improved classification accuracy across multiple datasets.
• Achieved an overall accuracy of 97.88% on the Houston University dataset, a substantial increase from 81.71% (MiniGCN).
• Demonstrated effectiveness in classifying hyperspectral images of rice seeds under high temperature stress.

Conclusions:

• The AN-GCN method offers a superior approach to hyperspectral image classification by adaptively leveraging spatial-spectral features.
• The adaptive neighborhood selection enhances feature extraction compared to fixed window methods.
• AN-GCN shows promise for applications requiring fine-grained discrimination, such as agricultural monitoring under environmental stress.