AIS, the global ship identification standard, is vulnerable to outages, coverage gaps, and deliberate deactivation, highlighting the need for independent ship identification methods. Optical imaging satellites offer a global, non-compliance-dependent solution. Paired with deep neural networks trained on satellite imagery of ships, it has become possible to determine the identity of specific vessels, based on their unique visual signatures. This enables re-identification, even when cooperative signals like AIS are unavailable or unreliable. Our paper builds on previous work with neural networks for ship identification, and presents an approach based on contrastive self-supervised learning. Self-supervised learning allows for existing, unlabeled, and freely available satellite imagery datasets with ships, to be leveraged for model training. Using these self-supervised models to initialize ship identification training results in almost 32% higher accuracy compared to baseline models. In one case equivalent to doubling the labeled training data. This lowers the threshold for optical ship identification from space by reducing dependence on large labeled datasets. This scalability is crucial for making space-based ship identification viable for global maritime situational awareness.
Nezhad et al. (Tue,) studied this question.