Dark shipping: absence of AIS is evidence, not a verdict

In brief

• AIS silence is ambiguous; model the probability that a message should have been observed.
• SAR/EO/RF detections become meaningful through probabilistic correlation, not nearest-neighbour matching.
• The product should rank competing hypotheses with evidence and uncertainty.

AIS silence alone is insufficient to classify a vessel as dark. Receiver geometry, equipment failure, channel congestion, terrestrial coverage gaps and data-pipeline failures can all produce missing observations.

A stronger formulation is:

**P(hypothesis

expected AIS, satellite/RF detections, trajectory history, identity constraints).**

First model the observation process

Before detecting anomalous silence, estimate how likely the system was to receive AIS at that place and time. Features can include receiver history, transponder class, local density, satellite/terrestrial source and recent message cadence.

This separates missing observation from unexpected missing observation.

Identity consistency is a second layer. Impossible jumps, duplicated MMSIs, implausible implied speed and conflicting vessel dimensions can indicate spoofing, identity collision or bad data. These are distinct hypotheses and should remain distinct in the interface.

Multi-sensor association

Non-cooperative sensors add evidence, but they introduce registration and timing uncertainty. A useful association score should include:

• satellite acquisition time and geolocation error;
• predicted vessel position distribution at acquisition time;
• candidate vessel dimensions/type;
• nearby traffic and alternative matches;
• RF emitter evidence where available.

The landmark dark-fleet study by Park et al. combined AIS, optical imagery, radar and night-light data to reveal fishing activity that was not visible in public AIS. Its significance is methodological: different sensing modalities close different observability gaps.

From detector to hypothesis manager

A deep-learning ship detector is only one component. The full system is closer to a probabilistic world model:

1. encode recent vessel/traffic state;
2. predict where cooperative vessels could plausibly be;
3. ingest SAR/EO/RF observations;
4. associate observations to existing hypotheses;
5. create new unmatched-object hypotheses;
6. update probabilities as new evidence arrives.

This naturally supports temporal persistence: one unmatched SAR detection is weak evidence; a coherent sequence of detections with no compatible AIS history is stronger.

Gap reconstruction is not evidence of deliberate shutdown

Recent loitering work increasingly reconstructs missing AIS before classification. Tao et al. use Transformer-based gap reconstruction and then classify loitering patterns. That is technically useful, but the literature review behind this note flags an important epistemic risk: a reconstructed gap explains what motion may have occurred; it does not explain why the signal was missing.

For dark-shipping inference I therefore keep two variables separate:

• latent vessel trajectory during the gap;
• latent cause of missing observation.

The second variable needs receiver coverage, transponder history, regional message density and independent sensor evidence. A trajectory model alone cannot establish deliberate AIS disablement.

Causal lens: model the missingness mechanism

Dark-shipping inference is a missing-data problem with multiple causes: receiver coverage, equipment failure, congestion, deliberate shutdown, spoofing or pipeline failure. Treating every gap as the same event confounds the cause of missingness with the vessel state during the gap.

The useful question is counterfactual: if reception conditions had remained normal, would we still expect the vessel to be unobserved? That separates an anomalous vessel hypothesis from an anomalous sensor/data path.

Evaluation

I propose to separate three tasks:

Sensor detection: ship detection/segmentation precision and recall.
Association: correct AIS↔sensor match rate and calibrated match probability.
Dark-event inference: reviewed event recall, false alerts per area/time, time-to-detection and hypothesis calibration.

Crowded ports, low-reception zones and intentionally withheld AIS should all be represented in testing.

Operational impact

The goal is not to produce an accusatory label. It is to reduce analyst search space while preserving the path from alert to evidence. A good alert explains: what was observed, what was expected, what matched, what did not, and which alternative explanations remain plausible.

References

• Park et al. — Illuminating dark fishing fleets in North Korea, Science Advances (2020).
• Global Fishing Watch — multi-sensor dark-fleet analysis.
• Li, Jiao & Yang — AIS trajectory prediction benchmarking across ML/DL methods (2023).
• Machine learning for AIS-driven maritime research: review of anomaly detection, prediction and DRL directions.
• Tao et al. — Transformer reconstruction and loitering-pattern classification (2025).
• Liu, Li & Liu — rule-derived AIS kinematic anomaly labels (2024).