Semi-Supervised Multiscale Dynamic Graph Convolution Network for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel in hyperspectral image (HSI) classification but require substantial labeled data.
• Conventional CNNs struggle with capturing long-range spatial context in HSIs due to fixed kernel shapes.
• Existing methods are often limited by the need for large labeled datasets and inability to capture diverse spatial relationships.

Purpose of the Study:

• To develop a novel hyperspectral image classification method that overcomes the limitations of supervised learning and fixed kernel structures.
• To introduce a dynamic multiscale graph convolutional network (GCN) capable of learning pixel representations dynamically and capturing both local and global contextual information.
• To reduce the dependency on large labeled datasets for effective HSI classification.

Main Methods:

• Constructing a region-level graph using superpixel segmentation and metric learning.
• Applying a dynamic pixel-level feature update strategy to the graph's adjacency matrix for adaptive representation learning.
• Expanding the graph convolutional network into a multiscale architecture to capture hierarchical spatial information.
• Utilizing graph learning principles for semi-supervised classification.

Main Results:

• The proposed Dynamic Multiscale Graph Convolutional Network (DMSGer) classifier demonstrates robust performance in HSI classification across four public datasets.
• DMSGer effectively captures both pixel-level and region-level information simultaneously within graph convolution layers.
• The multiscale expansion and dynamic feature updates contribute to improved classification accuracy by leveraging comprehensive contextual information.
• The semi-supervised approach significantly alleviates the burden of acquiring extensive labeled HSI data.

Conclusions:

• The DMSGer model offers a powerful and flexible approach for hyperspectral image classification, outperforming existing methods.
• The dynamic and multiscale nature of the GCN allows for more effective learning of complex spatial-spectral features in HSIs.
• This work presents a promising direction for semi-supervised learning in HSI analysis, reducing data acquisition costs.
• The developed method provides a robust solution for HSI classification tasks demanding high accuracy and efficiency.