Extracting ship and heading from Sentinel-2 images using convolutional neural networks with point and vector learning

• Published in: Journal of oceanology and limnology Vol. 43; no. 1; pp. 16 - 28
• Main Authors: Li, Xiunan, Chen, Peng, Yang, Jingsong, An, Wentao, Luo, Dan, Zheng, Gang, Lu, Aiying
• Format: Journal Article
• Language: English
• Published: Heidelberg Science Press 01.01.2025; Springer Nature B.V; State Key Laboratory of Satellite Ocean Environment Dynamics,Second Institute of Oceanography,Ministry of Natural Resources,Hangzhou 310012,China%State Key Laboratory of Satellite Ocean Environment Dynamics,Second Institute of Oceanography,Ministry of Natural Resources,Hangzhou 310012,China%National Satellite Ocean Application Service,Ministry of Natural Resources,Beijing 100081,China; Ocean College,Zhejiang University,Zhoushan 316021,China
• Subjects: Artificial neural networks; Deep learning; Earth and Environmental Science; Earth Sciences; extracts; Feature extraction; Feature maps; head; image analysis; learning; Machine learning; Marine environment; netting; Neural networks; Oceanography; oceans; processing time; Remote sensing; Research Paper; shipping; ships; Target detection; Vectors; Wakes; deep learning; remote sensing; ship detection; heading; Sentinel-2
• ISSN: 2096-5508; 2523-3521
• DOI: 10.1007/s00343-024-3271-1

Obtaining accurate ship positions and headings in remote sensing images plays a crucial role in various applications. However, current deep learning-based methods primarily focus on ship position detection, while the detection of ship wakes relies on traditional non-deep learning approaches, which often underperform in complex marine environments. We proposed a novel, simple, and efficient method called Point-Vector Net. The proposed method leverages convolutional neural networks (CNN) for feature extraction and subsequently integrates multi-scale features to generate high-resolution feature maps. In the final stage, ship positions and headings are represented using a combination of points and vectors. Comparative experiments with results from automatic identification system (AIS) reports demonstrate that our method achieved impressive performance in two-class ship target detection, with an average precision of 96.4%, recall rate of 94.3%, and an F1 score of 95.2%. Notably, the average heading error was 3.3°. The proposed model achieved a practical inference speed (FPS>30), and the average processing time for inferring a large-scale Sentinel-2 remote sensing image was 11.4 s.