An Improved Method for Road Extraction from High-Resolution Remote-Sensing Images that Enhances Boundary Information

Area of Science:

• Remote Sensing
• Computer Vision
• Deep Learning

Background:

• Deep learning methods have advanced road extraction from remote-sensing images.
• Existing methods suffer from spatial feature loss and lack of global context.

Purpose of the Study:

• To propose a novel network, the coord-dense-global (CDG) model, for improved road extraction.
• To address limitations of current deep learning approaches in spatial information and global context.

Main Methods:

• The CDG model incorporates a coordconv module for spatial information preservation.
• It utilizes an improved DenseNet for multi-feature utilization.
• A global attention module is integrated to enhance high-level feature representation.

Main Results:

• CDG demonstrated superior performance on a complex road dataset compared to DeepLabV3+, U-net, and D-LinkNet.
• Achieved mean IoU of 61.90% and mean F1 score of 76.10%, surpassing D-LinkNet.
• Showed improved capability in handling tree occlusion and universality across different datasets.

Conclusions:

• The CDG model offers a significant advancement in road extraction accuracy and robustness.
• Its integrated approach effectively combines spatial, dense, and global contextual information.
• CDG shows strong potential for practical applications in road network mapping from satellite imagery.