Abstract
Bridge detection from Synthetic Aperture Radar (SAR) is of great significance for infrastructure management, disaster prevention, and navigation automation. Although high-resolution SAR is increasingly accessible, exploiting Sentinel-1 remains advantageous due to its global coverage and accessibility. However, low spatial resolution of 20 m in Sentinel-1 products poses challenges in detecting small and indistinct bridges. To enhance bridge detection in low-resolution Sentinel-1 imagery, we propose a novel architecture that incorporates CycleGAN as an edge detector. CycleGAN generates detailed boundaries using 5 m land use maps, with enhanced bridge saliency. Since bridges often connect distinct edges like riverbanks and roads, this edge information assists Poly-YOLOv8, our chosen detector, in accurately localizing bridges. Our approach then integrates edge information through feature fusion module and feature alignment loss. Accordingly, the proposed SAR-EdgeYOLO achieves a precision of 98.2%, recall of 91.7%, and mAP of 94.8% at IoU 0.5, marking improvements of 3.9, 3.8, and 3.1 percentage points respectively over the baseline. The results demonstrate that CycleGAN-aided edge extraction effectively addresses the limitations of low-resolution remote sensing data, improving bridge detection accuracy and multi-class localization. This research can further contribute to the advancement of time-series infrastructure monitoring with wide applicability and higher accuracy.
Funding Information
  • Institute of Information & Communications Technology Planning & Evaluation
  • Korea government [MSIT (Ministry of Science and ICT)] (2021-0-01343)
  • Artificial Intelligence Graduate School Program (Seoul National University)], and Korea Research Institute for defense Technology planning and advancement
  • Korea government [DAPA (Defense Acquisition Program Administration)] (KRIT-CT-22-040)
  • Heterogeneous Satellite constellation based ISR (Intelligent Surveillance and Reconnaissance) Research Center, 2022

This publication has 34 references indexed in Scilit: