Satellite Detection of Moving Vessels in Marine Environments

Natalie Fridman, Doron Amir, Yinon Douchan, Noa AgmonAAAI 2019 · DOI

Introduction

Monitoring Exclusive Economic Zones (EEZs) is a critical yet daunting task for maritime authorities. While the Automatic Identification System (AIS) provides real-time location data for civilian vessels, it is frequently unreliable due to spoofing, hardware failure, or intentional deactivation in remote areas. To verify vessel identities or detect illegal activities, visual confirmation via Electro-Optic (EO) or Synthetic Aperture Radar (SAR) satellites is required. However, the high cost and narrow footprint of satellite imagery create a significant "needle in a haystack" challenge.

The Predictive Challenge

Satellite tasking involves a significant "data cycle" delay. By the time a satellite is positioned to capture an image, a vessel may have traveled far from its last known position, expanding the uncertainty zone to as much as 17,000 $km^2$. Covering such an area with high-resolution imagery is economically unfeasible. The KINGFISHER system addresses this by predicting vessel locations over 1–24 hour windows, drastically reducing the required search area.

The KINGFISHER Framework

The core of the KINGFISHER approach is the transition from simple linear extrapolation to a sophisticated movement graph based on historical AIS data.

Graph-Based Modeling

The system processes years of historical data into "leg segments"—straight-line paths between turning points—which are then clustered into nodes representing physical maritime locations. This transforms the open ocean into a network of established "maritime highways."

Vessel movement graph constructed from historical AIS data across the Indian Ocean and Southeast Asia.

Second Order Markov Chains (SOMC)

To capture vessel "intent," the framework employs a Second Order Markov Chain. Unlike standard Markov models that only consider the current state, the SOMC considers the previous state ($N_a \to N_b \to N_c$). This provides the necessary context to predict which route a vessel will take at a junction, significantly improving accuracy over first-order models.

Schematic of a Second Order Markov Chain where transitions depend on the sequence of previous nodes.

Multi-Agent Simulation and Mapping

To account for velocity uncertainty, the system runs 10,000 virtual "agents" through the movement graph. Each agent's speed is drawn from a Gaussian distribution derived from historical data for specific edges. The resulting 2D probability map is integrated with the Satellite Tool Kit (STK) to optimize tasking for the highest probability of detection.

Results and Economic Impact

The KINGFISHER model was validated using 24 months of AIS data from 100 vessels. Key findings include:

  • Superior Accuracy: The History-Graph model achieved a 67% success rate in 24-hour prediction windows, compared to 47% for standard straight-line models.
  • Massive Area Reduction: The model reduced the required satellite coverage area by over 99% compared to naive search methods.
  • Operational Validation: The system successfully tracked high-interest targets, such as the Russian research vessel YANTAR, maintaining contact over 12-hour uncertainty periods with minimal imagery.

Conclusion

By leveraging the inherent structure of maritime traffic through graph theory and Second Order Markov Chains, the KINGFISHER system transforms satellite monitoring from an expensive gamble into an operationally viable tool. Future work aims to extend prediction windows to 48 hours and improve handling for vessels entering unfamiliar regions.

Summary updated

BibTeX
@inproceedings{fridman2019satellite,
  author = {Natalie Fridman and Doron Amir and Yinon Douchan and Noa Agmon},
  title = {{Satellite Detection of Moving Vessels in Marine Environments}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33019452},
}
Figure