Superpixel-based graph convolutional neural network for polarimetric synthetic aperture radar image classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Polarimetric Synthetic Aperture Radar (PolSAR) images offer rich scattering information crucial for land cover classification.
• Traditional methods often struggle with complex spatial dependencies and noise in PolSAR data.
• Graph Convolutional Networks (GCNs) show promise but require effective feature extraction and graph construction for PolSAR analysis.

Purpose of the Study:

• To introduce a novel superpixel-based graph convolutional network (SGCN-LG) for enhanced PolSAR image classification.
• To automatically determine optimal superpixel segmentation and neighborhood sizes for graph construction.
• To fuse local and global contextual information for more robust classification results.

Main Methods:

• Automatic superpixel (SP) determination based on pixel standard deviation analysis.
• Construction of a local graph using neighboring SPs with adaptive neighborhood selection.
• Development of a global graph incorporating discriminative information from labeled SPs across the scene.
• Fusion of local and global features for final classification map generation.

Main Results:

• The proposed SGCN-LG model demonstrates superior performance compared to existing PolSAR classification methods.
• Automatic determination of superpixels and neighbors simplifies the model's application.
• Effective fusion of local and global information leads to improved classification accuracy.

Conclusions:

• The SGCN-LG model offers a powerful and effective approach for PolSAR image classification.
• The integration of local and global graph information significantly enhances classification performance.
• This method provides a robust framework for analyzing complex PolSAR data.