Maritime Object Detection in Synthetic Aperture Radar (SAR) Imagery Using Deep Learning

Abstract:

Maritime object detection using SAR imagery has wide-ranging applications such as tracking iceberg size and location, tracking ships to determine human economic activity, and enhancing ship safety. However, manual interpretation of SAR images is labor intensive. A robust computational model can reduce human effort while increasing both detection speed and accuracy. Our approach uses deep learning, specifically YOLOv4, (Bochkovskiy et al., 2020) an update to the popular YOLO object detection framework (Redmon et al., 2015), to distinguish between low-resolution SAR images of ships, icebergs, islands, gas and oil platforms and terrestrial data. We sought to compare the performance of YOLOv4 to prior versions to determine its improvement in efficiency.

In our approach, we curated and labeled over 14,000 open source SAR images using LabelImg, creating a valuable resource for future researchers in this domain. The dataset includes SAR images with ships, islands, oil platforms and terrestrial data. We evaluated the performance of various algorithms on this dataset. The preliminary YOLOv4 experiments yielded a mean average precision (mAP) of 86%, with a detection testing time of about 20ms per image. Compared to prior research which focused on uni-class detection (Bentes et al., 2016), our approach handles a multitude of classes while maintaining a high accuracy in model performance. Our current research seeks to improve both the size of the dataset and the quality of the detection algorithms.