Vessel Identity · Dataset integrity

Evaluate passages, not repeated clips.

Analyze groups chronologically adjacent recordings from the same MMSI into vessel encounters. The model can then be tested on genuinely separate passages instead of neighbouring clips that share the same machinery state, range and recording conditions.

Analyze · Vessel Identity library
Analyze Vessel Identity library showing MMSI groups, encounter IDs and model controls
Each recording exposes its dataset split and automatically assigned encounter identifier.

Why encounters matter

Random clip splitting can produce deceptively high accuracy.

A long vessel passage may yield hundreds of short audio windows. If those windows are randomly divided between training and testing, the model sees nearly identical examples on both sides. It may learn the encounter’s range, background noise or hydrophone response rather than a reusable vessel signature.

Encounter grouping treats the passage as the independent experimental unit. Clips remain available individually for listening and selection, but their shared session is visible and can be kept together when constructing model partitions.

Current rule: recordings for one MMSI separated by no more than 30 minutes are assigned to the same encounter. A larger time gap starts a new encounter.

Automatic workflow

From timestamps to leakage-aware evaluation.

Order by vessel and time

Labelled recordings are grouped by MMSI and placed in chronological order using their recording timestamps.

Assign encounter IDs

Contiguous recordings receive stable, readable identifiers such as 209345000-E0001.

Expose every clip

The split and encounter ID appear directly beside each file in the Vessel Library.

Detect split leakage

Analyze counts encounters whose clips cross training, validation or test partitions.

Protect verification

Any held-out encounter that also contains a training clip is automatically excluded from batch verification.

Report by encounter

Verification reports both clip accuracy and majority-vote accuracy across independent encounters.

Interpretation

What the library tells the analyst.

IndicatorMeaningRecommended response
Encounter countNumber of temporally independent sessions available for a vessel.Prefer vessels represented across several passages and operating conditions.
Cross-split encounterAt least one session appears in more than one dataset partition.Reassign the complete encounter to one partition before formal evaluation.
Many clips, few encountersHigh file volume may come from only one or two passages.Collect additional independent encounters rather than subdividing existing audio.
No timestampThe recording cannot be reliably joined to adjacent files by time.Treat it conservatively as a separate encounter until provenance is known.

Scope and limitations

Encounter grouping improves the test design; it does not prove vessel identity.

The grouping rule uses MMSI and time. It does not yet infer CPA, vessel operating state or competing traffic. Incorrect AIS association, simultaneous vessels and repeated environmental conditions can still bias results.

For operational work, review unusual sessions, retain independent deployment periods and compare performance across distance, speed, hydrophone and environment whenever those fields are available.

Designed to evolve

The encounter identifier is part of the library record, leaving room for later AIS-track association and manual encounter editing without changing the overall model workflow.

Build evaluation around independent vessel passages.

We can demonstrate encounter-aware library organisation using a representative vessel dataset.

Plan a demonstration