A Survey of Deep Learning Road Extraction Algorithms Using High-Resolution Remote Sensing Images

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Artificial Intelligence

Background:

• Road networks are crucial for transportation, economic development, and disaster management.
• Automatic road extraction from high-resolution remote sensing images is a challenging but vital task.
• Deep learning models have become prominent tools for analyzing remote sensing data.

Purpose of the Study:

• To systematically review and summarize deep-learning-based techniques for automatic road extraction.
• To classify deep learning models based on their supervision levels (fully, semi-, and weakly supervised).
• To provide a comprehensive overview and future outlook on deep learning applications in road extraction.

Main Methods:

• Review of recent literature on deep learning for road extraction.
• Classification of deep learning models based on data labeling strategies.
• Analysis of model performance and applicability in various scenarios.

Main Results:

• Deep learning models significantly advance automatic road extraction accuracy and efficiency.
• Categorization reveals distinct advantages and challenges for supervised, semi-supervised, and weakly supervised approaches.
• The review synthesizes a broad range of deep learning techniques applied to this domain.

Conclusions:

• Deep learning is a powerful methodology for automatic road extraction from remote sensing data.
• Further research is needed to optimize models for diverse conditions and improve label efficiency.
• The field is rapidly evolving, with significant potential for future applications in geospatial intelligence.