Transshipment detection should model an interaction, not two nearby tracks
In brief
- Classic transshipment work operationalises encounters using proximity, low speed, duration and distance from anchorages.
- Those rules are interpretable but cannot by themselves distinguish transshipment from benign rendezvous or waiting.
- A stronger model treats the event as a multi-agent interaction with pre/post encounter trajectories and uncertainty.
Transshipment is a useful example of why maritime intelligence often needs interaction modelling, not single-vessel classification.
Miller et al.’s influential global study operationalised an encounter as a fishing vessel and refrigerated carrier within 500 m for at least 2 h, moving below 2 kn and sufficiently far from an anchorage. The rules are transparent and enabled global-scale analysis over tens of billions of AIS messages.
But proximity is evidence of an encounter—not direct observation of what was transferred.
From threshold event to relational state
I propose to represent each candidate event with:
- fishing-vessel trajectory before/during/after;
- carrier trajectory before/during/after;
- relative range, bearing and velocity;
- joint low-speed duration;
- local traffic density and anchorage/port context;
- vessel type/registry/identity confidence;
- AIS reception quality and gaps;
- optional EO/SAR imagery or operational records.
The latent variable becomes:
| **P(interaction type | vessel A history, vessel B history, context, external evidence).** |
Possible states include benign co-location, queueing, bunkering/service, fishing support, transshipment candidate and unresolved encounter.
Why the pre/post trajectory matters
A carrier that loiters for hours before meeting a fishing vessel and then departs toward port creates a different hypothesis from two vessels briefly co-located in congested traffic.
That suggests a world-model formulation: encode the joint state, then ask whether the observed encounter is likely under normal traffic dynamics. The model should explain which part of the evidence changed the hypothesis.
Causal lens: proximity is not the event
Two vessels becoming co-located is an observed relation; transshipment is a latent interaction with an operational cause and downstream consequences. Context such as port queues, bunkering, service operations or fishing support can confound simple proximity rules.
A stronger model asks: what evidence would change if the interaction had not occurred? Pre/post draught, subsequent route, fishing effort, port call, imagery and observer records provide distinct causal consequences that can strengthen or weaken the transshipment hypothesis.
Evaluation gap
The original transshipment study explicitly notes the difficulty of obtaining independent public transshipment records. That remains the central validation problem.
A modern benchmark could combine:
- AIS-derived candidate encounters;
- known port/anchorage/service locations;
- vessel registry/ownership metadata;
- satellite imagery for a stratified subset;
- human review with explicit uncertainty;
- publicly documented enforcement or observer records when available.
The important methodological point is to keep candidate detection and event verification as separate metrics.
Fisheries impact
Reliable encounter modelling matters beyond surveillance. It can support seafood traceability, effort attribution, monitoring of at-sea transfers and more accurate reconstruction of fishing operations. But the evidence standard must match the claim: AIS can identify a plausible encounter; it does not alone establish illegal activity.