MG-RoadNet: Road Segmentation Network for Remote Sensing Images Based on Multi-Receptive Field Graph Convolution

This repository provides the official implementation of MG-RoadNet, a multi-receptive field graph convolutional network for road segmentation in remote sensing images.

MG-RoadNet is designed to preserve road connectivity while retaining fine-grained road details, especially in complex scenarios with heavy occlusions (e.g., trees, buildings) and diverse road morphologies.

📄 Paper:
MG-RoadNet: Road Segmentation Network for Remote Sensing Images Based on Multi-Receptive Field Graph Convolution

🧠 Network Architecture

MG-RoadNet follows a U-shaped encoder–decoder architecture with ResNet-style residual blocks as the backbone.

⚙️ Implementation Details

• Framework: PyTorch 1.11.0
• Python Version: 3.8
• CUDA: 11.3
• Optimizer: Adam
• Loss Function: Binary Cross Entropy (BCE)
• Learning Rate: 1e-4
• Weight Decay: 1e-4
• Batch Size: 4
• Training Epochs: 200

📌 Citation

If you find this repository useful, please consider citing our paper:

@article{Song2025,
  author  = {Song, Runtian and Shi, Fan and Du, Guikang and Zhang, Xinpeng and Jiang, Cheng},
  title   = {{MG-RoadNet}: Road Segmentation Network for Remote Sensing Images Based on Multi-Receptive Field Graph Convolution},
  journal = {Signal, Image and Video Processing},
  volume  = {19},
  number  = {8},
  pages   = {679},
  year    = {2025},
  issn    = {1863-1711},
}