Loitering, waiting and anchorage: define the state before training the detector

In brief

• Several influential loitering methods use low speed, course change or trajectory redundancy, but ordinary waiting and anchorage can satisfy the same geometry.
• High reported classification scores can measure agreement with a rule-derived label rather than agreement with independently observed behaviour.
• I propose to model anchorage, waiting, drifting, loitering and encounter states jointly, with context and uncertainty.

The first modelling question is not which classifier to use, but which latent vessel state the observations are intended to identify.

The literature on AIS loitering makes that problem unusually visible. Miller et al. define reefer loitering using low speed and duration away from shore; Zhang et al. model frequent turning and trajectory redundancy; Wijaya and Nakamura combine speed, course change, heading/COG discrepancy and spatial range. These are useful operational abstractions, but they are not the same behavioural concept.

A vessel waiting for a berth, slowly drifting between jobs, swinging at anchor or deliberately remaining in an area can generate similar trajectories.

What the literature shows

Miller et al. use a rule of less than 2 kn for at least 8 h and explicitly acknowledge waiting between jobs as a false-positive mechanism. Their sensitivity analysis shows how strongly estimated loitering hours depend on the chosen speed threshold.

Zhang et al. define loitering around repeated turning in a limited area and use a CNN to classify trajectory shapes. The approach demonstrates that geometry contains useful signal, but anchor swing and port waiting can also create repeated turning.

Wijaya and Nakamura attempt to separate anchoring from loitering with navigational status and a bounding-box criterion before scoring trajectories. Their 2023 evaluation reports 97% accuracy and 92% F-score on 137 vessels labelled by one expert. That is useful evidence of feasibility, but it is still a small, context-specific labelled set.

The methodological implication is that the label definition becomes part of the model. Reported performance is therefore inseparable from the operational definition used to construct the target state.

A latent-state formulation

I propose to define a state space such as:

Anchoredlow translational motion, bounded swing, anchorage/port context.

Waiting / driftinglow speed with slow spatial displacement, often near queues or approaches.

Loiteringpersistent local motion inconsistent with expected route/operation, with intent unresolved.

Encounter / rendezvousstate depends on another vessel's co-location and joint motion.

The model then estimates:

**P(state_t

AIS history, vessel type, port/anchorage context, traffic, observation quality).**

This is closer to a world model than a binary anomaly detector. It separates what happened geometrically from what operational state best explains it.

Why labels need an independent evidence path

A recurring weakness in maritime anomaly literature is circular evaluation: rules define anomalies, then a model learns those labels, and accuracy measures how well it reproduces the rule.

Liu et al. report approximately 99% classification accuracy for kinematic anomalies, but their anomaly labels are generated through rules and hand screening. That is valuable as a scalable labelled dataset, yet the metric is primarily agreement with the constructed taxonomy.

For higher-stakes claims I propose to build a second evidence path: port-call records, anchorage polygons, berth assignments, operational logs, imagery, or reviewed vessel histories. Even partial independent evidence is better than treating the labelling rule as ground truth.

Research direction

The strongest next problem is waiting-aware loitering detection:

1. reconstruct stopping/slow-motion episodes;
2. infer whether the episode is anchorage, queue/waiting, drift or anomalous loitering;
3. condition on port/traffic context and vessel type;
4. retain uncertainty when intent cannot be observed;
5. test across ports and traffic regimes.

This formulation is more demanding than binary loitering detection, but it better matches the operational distinction the model is expected to support.

References

• Miller et al. — Identifying Global Patterns of Transshipment Behavior, Frontiers in Marine Science (2018).
• Zhang et al. — Loitering behavior detection and classification from trajectory shape and CNNs, Ocean Engineering (2022).
• Wijaya & Nakamura — Loitering behavior detection from AIS spatiotemporal characteristics, PeerJ Computer Science (2023).
• Liu, Li & Liu — AIS-based kinematic anomaly classification, Ocean Engineering (2024).