Fractal-Domain Vision Graph Neural Network for Remote Sensing Ground Target Classification

Area of Science:

• Computer Vision
• Graph Neural Networks
• Fractal Signal Processing

Background:

• Traditional vision graph neural networks lack methods to capture complex fractal dynamics in imagery.
• Existing models struggle with fusing spatial and textural information effectively for scene understanding.

Purpose of the Study:

• To propose a novel Fractal-domain Vision Graph Neural Network (FD-ViG) for enhanced remote sensing scene understanding.
• To establish a new paradigm in graph representation learning by integrating fractal dynamics.

Main Methods:

• Developed a Fractal-Domain Learning Module using local Hölder exponents and Singularity Power Spectrum (SPS) for fractal-spatial feature fusion.
• Introduced a Fractal Graph Construction Module combining semantic attention and fractal similarity for adaptive topology generation.
• Implemented a Graph Propagation Module with power-law multi-scale propagation for cross-scale diffusion and texture-structure learning.

Main Results:

• Achieved high accuracies: 91.75% (UCMerced), 89.52% (RSSCN7), and 92.78% (SIRI-WHU).
• Demonstrated consistent improvements over representative vision graph models (ViG, WiGNet, ViHGNN) with a lightweight design (2.6M parameters).
• Showcased competitive or superior performance compared to ResNet-18 and strong generalization on SAR imagery.

Conclusions:

• FD-ViG provides a principled and effective bridge between fractal theory and graph deep learning.
• The model enables interpretable remote sensing scene understanding, particularly for complex textures and structures.
• This work advances graph representation learning for complex visual data analysis.