Automatic Bridge Detection Using SAR Satellite Imagery

Synthetic Aperture Radar (SAR) operates in the microwave region, providing high-resolution images in all weather conditions by penetrating clouds, haze, etc. Detecting bridges in SAR images, particularly in urban areas, is challenging due to similar backscatter values between urban structures and bridges. This study proposes a methodology framework for efficient bridge detection for SAR images using the You Only Look Once (YOLO) deep neural network and compares YOLOv5 and YOLOv8 models. Sentinel-1A (10m resolution) and Earth Observation Satellite-4 (EOS-4) (3m resolution) data are used for training, validation, and testing. YOLOv5s and YOLOv8n models, trained with transfer learning, achieved mAP values of 97.5% and 98.8% with the AdamW optimizer, and 95.4% and 93.4% with the SGD optimizer. The study identifies YOLOv8n with the AdamW optimizer as the most accurate solution for bridge detection in SAR imagery as compared to YOLOv5s.