Triple-Level Sparsity Awareness for Marine Ship Surveillance Using Satellite Synthetic Aperture Radar
IEEE Transactions on Automation Science and Engineering · 2026 · 23 · 5155-5166
DOI: 10.1109/tase.2026.3661084 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Automation Science and Engineering, 2026. Verfasst von Tianwen Zhang und 1 weiteren Autor:innen. Die Arbeit wurde bisher 58-mal zitiert (Crossref) und nennt 53 Literaturangaben. Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
Most existing methods for marine ship surveillance by satellite synthetic aperture radar (SAR) ignore the sparsity of ships, causing inefficiency and inaccuracy. For this, this paper proposes triple-level sparsity awareness (TLSA) for promising SAR marine ship surveillance. It simultaneously considers the sparsity of SAR imaging and ship scattering. TLSA is implemented from three aspects, i.e., feature-level sparsity awareness (f-LSA), proposal-level sparsity awareness (p-LSA), and token-level sparsity awareness (t-LSA). They are executed sequentially to yield more discriminative sparse ship representations. f-LSA designs a sparse range azimuth multi-head self-attention (SRA-MSA) and a sparse scattering point hyper-graph topology perception (SSP-HGTP) to depict the spatial sparsity of SAR imaging. It considers the Sinc point spread function, and the scattering sparsity caused by the complex physical structure of the ship. f-LSA enables more representative features, so it is known as feature-level. p-LSA learns spatially sparse proposal locations, and specifies the number of proposals for the sparsity awareness at the proposal-level, by a sparse proposal generation network (SPGN). In SPGN, context harmony shuffle attention (CHSA) is designed to expand and refine valuable contextual information. t-LSA yields efficient feature encoding and decoding, by a sparse token prediction network (STPN) and a deformable cross-attention network (DCAN). In this way, only the tokens that are beneficial to the surveillance are selectively retained and updated. Experiments are performed on the satellite SAR data, i.e., RadarSat-1, TerraSAR-X, and Sentinel-1. Results show the validity of f-LSA, p-LSA, and t-LSA, and showcase the state-of-the-art performance of TLSA. Finally, large-scene marine surveillance results reveal the value of TLSA in practical engineering applications. Note to Practitioners—This paper is the first to consider the spatial sparsity induced by the Sinc point spread function in SAR imaging, and the scattering sparsity caused by the complex physical structure of the ship, for preferable SAR marine ship surveillance. Sparsity awareness is implemented at three levels - feature-level, proposal-level, and token-level. Feature level aims to extract sparse features, proposal-level aims to generate sparse proposals, and token-level aims to select more valuable tokens. Our research results indicate that if sparsity is reasonably considered, both surveillance efficiency and accuracy can be significantly improved. In addition, industrial deployment is also carried out on the embedded AI edge device platform NVIDIA Jetson TX2, revealing the automation industrial value of TLSA in practical engineering applications.
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Zhang, T., & Zhang, X. (2026). Triple-Level Sparsity Awareness for Marine Ship Surveillance Using Satellite Synthetic Aperture Radar. IEEE Transactions on Automation Science and Engineering, 23, 5155-5166. https://doi.org/10.1109/tase.2026.3661084
Vancouver
Zhang T, Zhang X. Triple-Level Sparsity Awareness for Marine Ship Surveillance Using Satellite Synthetic Aperture Radar. IEEE Transactions on Automation Science and Engineering. 2026;23:5155-5166. doi:10.1109/tase.2026.3661084
BibTeX
@article{Zhang2026,
author = {Zhang, Tianwen and Zhang, Xiaoling},
title = {Triple-Level Sparsity Awareness for Marine Ship Surveillance Using Satellite Synthetic Aperture Radar},
journal = {IEEE Transactions on Automation Science and Engineering},
year = {2026},
volume = {23},
pages = {5155--5166},
doi = {10.1109/tase.2026.3661084}
}
EUVIMED