VTUAD: An AIS-Auto-Labeled Vessel-Type Underwater Acoustic Dataset Reconstructed from Ocean Networks Canada Archives

VTUAD (rebuilt) is an open reconstruction of the Vessel-Type Underwater Acoustic Dataset (VTUAD) from its public upstream source, the Ocean Networks Canada (ONC) archive (https://data.oceannetworks.ca). Every label is generated automatically by aligning co-archived Automatic Identification System (AIS) messages with hydrophone recordings — no human listening is involved at any stage of labeling. Dataset at a glance:- Recordings: 525 labeled single-vessel scenario clips (509 vessel scenarios + 16 dedicated background recordings)- Duration: ~56.0 hours of 32 kHz single-channel WAV audio- Classes: 12 — 11 AIS-derived vessel types + background noise- Segments: 133,784 labeled log-mel segments (16 kHz, 3 s segments with 1.5 s hop, 128-bin log-mel; train 95,103 / val 20,479 / test 18,202)- Split: frozen, stratified recording-level train/validation/test partition of 369/80/76 clips (70/15/15), pairwise group intersections exactly zero, SHA-256-checksummed- Location: ONC icListen hydrophone ICLISTENAF2523, Strait of Georgia, British Columbia, Canada- Period: May 2016 – May 2017- Per-clip metadata: MMSI, UTC interval, vessel dimensions, and CTD-derived environmental covariates (22 columns) This deposit contains the metadata registry, the frozen split, the complete rebuild pipeline code, the full WAV clips (12 per-class ZIP archives, ~25 GB, since v1.1), and the precomputed log-mel feature store (train split shipped as 7 parts). The WAV audio (~25 GB) can be regenerated from the ONC archive with the released pipeline. Every clip retains its vessel MMSI, so users can optionally enforce vessel-level (hull-level) split isolation — a stronger leakage guard than the released recording-level default. Source data provided by Ocean Networks Canada (data.oceannetworks.ca), used under CC-BY 4.0. Version v1.1 adds the full WAV clips (525 clips, 25 GB) as 12 per-class ZIP archives (vtuad_wav_<class>.zip). SHA-256 checksums: see SHA256SUMS.txt included in this version. The v1.0.0 metadata, splits, features, and code files are unchanged. Version v1.2.0 (registry & feature-store fix release): traceable identifiers (clip_uid = SHA-256 of WAV bytes; segment_uid per 3 s segment); single label source (metadata.csv by relative path) fixing 15 mislabeled clips and restoring 8 train recordings / 1,511 segments lost to a filename collision (segments now 133,784: train 95,103 / val 20,479 / test 18,202); explicit per-clip utc_begin/utc_end; feature store rebuilt with collision-free names and relative-path manifests; frozen split membership unchanged (vtuad_v2.json, corrections documented); reference baseline re-run: accuracy 0.4149 ± 0.0143, macro-F1 0.2576 ± 0.0137 (was 0.3620 ± 0.0058 / 0.2390 ± 0.0141). WAV audio byte-identical to v1.1.0. See CHANGELOG.md inside zenodo_vtuad_v1.2.0.zip. Note: vtuad_features_train.zip (7.40 GB) is shipped as 7 split parts (vtuad_features_train.zip.part01–part07); concatenate them in order (Windows: copy /b part01+...+part07 vtuad_features_train.zip; Unix: cat part0* > vtuad_features_train.zip) and verify against the whole-file SHA-256 in SHA256SUMS.txt — see README_vtuad_features_train_parts.txt. What is new in v1.3.0 (2026-09-23): (1) vtuad_strict_wav_vessel.zip (509 clips; uploaded as three equal byte-split parts .part01-.part03, concatenate in order to restore - see README_vtuad_strict_wav_vessel_parts.txt) and vtuad_strict_wav_background.zip (16 clips): all 525 recordings re-cut to the exact AIS-qualified interval bounds (v1.2.0 released archive-aligned containers; both views describe the same scenarios). strict_interval_wav_SHA256SUMS.txt gives the per-clip SHA-256 ledger for the WAVs inside the archives. (2) registry_anonymized.csv: clip registry with the MMSI column removed; a persistent pseudonymous hull_id retains per-hull grouping for split isolation and provenance auditing. (3) registry_verification_sample30.csv (+ companion README): manual registry-verification ledger of a stratified 30-MMSI sample (covering 9 of 11 vessel classes, 82 of the 525 recordings) cross-checked on 2026-09-23 against the Canadian ISED vessel registry (NMID), VesselFinder, MarineTraffic, and MyShipTracking: 28 confirmed, 2 unresolved (suspected MMSI reuse after the recording window), 0 contradicted. (4) static_field_fill_audit.csv (hull x field fill provenance) and static_field_conflicts.csv (one row per hull) with static_audit_README.md: a full audit of how each released class_code/length/beam was reconstructed from AIS static messages. Note that the conflict counts are a lower bound: only 3 days of raw NMEA survived for direct re-decode, so conflicts outside that evidence window cannot be observed. (5) Documentation refresh: dataset_card.md, CITATION.cff, environment.yml.