Transferable Contextual Network for Rural Road Extraction from UAV-Based Remote Sensing Images.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Rural road extraction from UAV imagery is complex due to diverse road features and data acquisition challenges.
• High costs, manual annotation needs, and policy limitations hinder rural UAV data collection.

Purpose of the Study:

• To propose a transferable contextual network (TCNet) for enhanced accuracy and transferability in rural road extraction.
• To address challenges in rural road extraction, including data scarcity and complex road environments.

Main Methods:

• Utilized Stable Diffusion for data augmentation to generate diverse training samples.
• Integrated clustered contextual Transformer (CCT), clustered cross-attention (CCA), and CBAM attention for model transferability.
• Developed a Dice-BCE-Lovasz (DBL) loss function to improve segmentation performance on imbalanced datasets.

Main Results:

• TCNet achieved excellent performance on DeepGlobe and road datasets with only 23.67 M parameters.
• Demonstrated outstanding zero-shot transferability to new rural remote sensing datasets.
• Showcased effective performance in Burgundy, France, and Yunnan, China, without fine-tuning.

Conclusions:

• TCNet offers an efficient and accurate solution for rural road extraction using UAV imagery.
• The proposed model overcomes data limitations and demonstrates strong generalization capabilities across diverse regions.