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

Background & Summary

Table 1. Public underwater acoustic datasets most relevant to VTUAD. Durations and class counts as reported by the respective releases; “AIS labels” indicates vessel-type labels derived from Automatic Identification System messages.

Figure 1. The VTUAD labeling concept. Schematic of AIS-based automatic annotation: a vessel inside the 4,000 m inclusion radius broadcasts position and static type messages (a); the co-archived hydrophone recording over the same UTC interval (b) inherits the vessel’s type label when the single-vessel scenario condition holds. Contrast with manual listening-based annotation.

Figure 1. The VTUAD labeling concept. Schematic of AIS-based automatic annotation: a vessel inside the 4,000 m inclusion radius broadcasts position and static type messages (a); the co-archived hydrophone recording over the same UTC interval (b) inherits the vessel’s type label when the single-vessel scenario condition holds. Contrast with manual listening-based annotation.

Methods

Source Infrastructure

All raw material is drawn from the public archive of Ocean Networks Canada (ONC) [13], which operates cabled and autonomous observatories on the Pacific coast of Canada and archives, for the same time ranges, (i) hydrophone audio, (ii) messages from co-located shore-based AIS receivers, and (iii) conductivity–temperature–depth (CTD) casts. The three instrument streams used here are:

• Hydrophones. Ocean Sonics icListen smart hydrophones, device IDs ICLISTENAF2523 (primary) and ICLISTENAF2556, deployed at eight station-periods between May 2016 and March 2024 in and around the Strait of Georgia (deployment coordinates between 49.0808°N–50.0208°N and 123.3282°W–125.2354°W; instrument depths 7–147 m). The released v1 corpus uses the first deployment of ICLISTENAF2523 (station code LSBBL; 49.080927°N, 123.338713°W; 141 m depth; 2016-05-02 to 2017-06-24), in the Fraser River delta shipping approaches of the Strait of Georgia — the same recording region that underlies DeepShip [2] and Oceanship [3]. Audio is archived as single-channel WAV at 32 kHz.
• AIS receiver. Digital Yacht AISNET shore receiver, device DIGITALYACHTAISNET1302-0097-01, co-archived with the hydrophone deployments. Messages are stored as daily plain-text files of timestamped NMEA 0183 sentences (!AIVDM).
• CTD. Sea-Bird SBE 19plus profiler, device SBECTD19p6935, providing temperature, conductivity, pressure, salinity, and derived sound speed; 1,020 daily cast files were downloaded and cleaned for the release window.

The AIS-annotation pipeline

The construction pipeline adapts the open-source ONC dataset-construction code released with the original VTUAD study [11] to the current ONC web API, and consists of six scripted stages (Figure 2). Every rule below is a concrete, inspectable parameter of the released code; no stage involves human listening.

Stage 1 — Deployment query and verification. Hydrophone deployment records (begin/end timestamps, coordinates, depth, station code) are queried from ONC and manually verified for coordinates and date ranges before any download starts. This is the only manual step, and it concerns the instrument metadata, not the labels.

Stage 2 — AIS acquisition and parsing. Raw AIS streams are downloaded per deployment and parsed with a pyais-compatible NMEA 0183/AIS decoder [14]. We retain Class-A position reports (message types 1–3), static-and-voyage data (type 5), and Class-B position/static reports (types 18, 19, 24), extracting per message: MMSI, UTC timestamp, longitude/latitude, speed over ground (SOG), course over ground (COG), true heading, and the type-and-cargo code defined by ITU-R M.1371 [10].

Stage 3 — Static-table propagation and cleaning. Because type-and-cargo codes appear only in the infrequent static messages, we propagate each MMSI’s static label across all of its position-only messages. Messages with invalid coordinates and exact duplicates are dropped. The great-circle distance between each message position and the hydrophone is computed per message. Messages are filtered by a scenario inclusion radius of 4,000 m — matching the largest of the original VTUAD’s 2,000/3,000/4,000 m distance strata [11] — with a hard global ceiling of 15,000 m applied during parsing.

Stage 4 — Scenario identification. Deployment-level AIS tracks are merged and linearly interpolated (position and SOG) at 20 s granularity wherever two consecutive reports of the same MMSI are separated by more than 20 s and at most 1,200 s. The interpolated track table is scanned for single-vessel scenarios: intervals during which exactly one vessel remains inside the inclusion radius for at least 30 consecutive minutes. Each scenario is padded by ±1 minute at both ends. Background-noise intervals are defined symmetrically as periods in which no vessel is inside the exclusion radius (inclusion radius + 2,000 m = 6,000 m). The single-vessel requirement is the pipeline’s key label-purity mechanism: it guarantees that within a scenario clip, the dominant radiated noise originates from the labeled vessel, at the price of discarding all multi-vessel traffic.

Stage 5 — Audio download and interval matching. WAV archives covering each scenario interval are downloaded from ONC and matched to AIS intervals with a ±5-minute tolerance. For the v1 release window this produced 1,020 scenario-matched candidate clips, which after the Stage-6 completeness and integrity checks yield the 525 released clips (509 vessel scenarios and 16 dedicated background recordings, labeled via the metadata registry). A note on label snapshots: by registry path the 525 clips are 509 vessel/ + 16 background/ recordings; under the current metadata label column the background class additionally comprises 22 vessel-directory recordings whose propagated static code maps to background (38 background-labeled clips in total; the Data Records section gives the relation to the 53-clip label snapshot that was in force when the feature manifests were generated). Matched audio is split into normalized pieces, grouped by AIS interval, and stored under the corresponding label directory.

Stage 6 — Metadata generation. Each released clip is registered in metadata.csv with: label, duration (s), relative path, sample rate, AIS type-and-cargo code, date, MMSI, vessel length and beam (from static messages), and per-clip averaged CTD covariates (temperature, conductivity, pressure, salinity, sound speed, plus their min–max-normalized versions).

Two design decisions in the pipeline deserve explicit justification. The 20 s interpolation granularity with a 1,200 s gap ceiling balances AIS report rates (Class-A vessels under way report every 2–10 s [10], but shore-receiver capture near the delta is lossy) against the risk of interpolating through a genuine turn or stop: beyond 20 minutes without a report we refuse to invent a position, and the vessel simply drops out of the scenario table. The 30-minute minimum scenario duration is chosen so that every scenario clip supports multiple minutes of contiguous audio after padding and matching — long enough for both segment-level machine learning and clip-level perceptual or propagation analysis — while the ±1 minute padding absorbs clock skew between the AIS receiver and the hydrophone archive, with the ±5-minute audio matching tolerance as a second, coarser safety net. All radii are great-circle distances on the WGS-84 ellipsoid; the zone geometry for radii from 3 to 12 km is published as interactive maps so that users can inspect which traffic lanes fall inside the labeling zone.

Class Taxonomy

Labels follow the AIS type-and-cargo taxonomy [10] rather than an ad-hoc registry: each vessel class corresponds to a numeric AIS code range, so the taxonomy is externally defined and machine-derivable. The 12 released classes are: tug, cargo, passengership, other, background, fishing, pleasurecraft, dredger, sailing, tanker, rescue, and pilotvessel. Two classes need precise definitions. background contains no-vessel intervals as defined in Stage 4: no AIS-tracked vessel inside the 6,000 m exclusion radius for the whole interval. other is the boundary complement on the vessel side: the scenario contains exactly one vessel inside the 4,000 m inclusion radius (so it is not background), but that vessel’s propagated type-and-cargo code is 0 (“not available”, the default of an unconfigured transponder) or falls in a reserved/unassigned range of the ITU-R M.1371 code table, so no finer vessel type can be assigned. other is therefore semantically heterogeneous by construction — it mixes genuine small craft, misconfigured transponders, and rare platform types — and should not be interpreted as a coherent vessel type. Relative to the original VTUAD’s five classes (cargo, tanker, tug, passengership, background noise) [11], the rebuilt corpus resolves seven additional vessel types. The distribution is strongly long-tailed (tug 188 clips vs. pilotvessel 2), reflecting the real traffic composition of the Fraser River approaches; we deliberately keep the tail classes rather than merging them, because hull-level metadata let users define their own aggregation or re-split by MMSI.

Frozen, Leakage-Safe Splits

The train/validation/test partition is generated once, frozen, and released as vtuad_v1.json (created 2026-08-31, seed 42, stratified by class; test ratio 0.15 and validation ratio 0.15, both measured against the total — a 70/15/15 partition). The split operates at the recording (clip) level: one group is one recording file (525 groups = 525 recordings), all segments of a clip fall entirely into one partition, and a built-in self-check verifies that the three group sets are pairwise disjoint (intersection size exactly 0 for train∩val, train∩test, val∩test). The frozen split is integrity-protected by a SHA-256 checksum (b0ff53acc989bc7cb03f0abc3c41a5765f937c5d83974f0674db182c7fd0fe17). Because every clip retains its MMSI, users can additionally enforce vessel-level (hull-level) isolation (all clips of one hull in one partition) — a stricter guard than the released default and than what the original VTUAD metadata support [11]; note, however, that the frozen split does not merge groups by MMSI, so hull-level isolation is a user-enforced option, not a property of the released default. The resulting partition is 369 / 80 / 76 clips (Table 2).

Unified Machine-Learning Front-End

To make the corpus directly comparable with other public UATR libraries in our companion benchmark [20], we release a unified feature front-end: audio is resampled to 16 kHz, cut into 3 s segments with 1.5 s hop (50% overlap; tail segments shorter than 3 s discarded), and converted to single-channel 128-bin log-mel spectrograms (n_fft = 1024, hop length = 320 samples, i.e. 20 ms; 80 dB dynamic-range floor) computed with librosa [15]. This yields 132,273 labeled segments (train 93,592 / validation 20,479 / test 18,202), distributed per class as in Table 2. We state two counting calibers explicitly, because both appear in the released artifacts: the segment manifests register 133,784 rows (train 95,103), of which 1,511 train rows are duplicate registrations — eight recordings whose numeric file IDs collide across the background/ and vessel/ directories (e.g., background/0.wav vs. vessel/0.wav) carried the same class label under the label snapshot in force at feature-generation time, so their same-named segment files (N_sX.XX.npy) were written to the same path, the later write overwriting the earlier, while the manifest kept both rows. The deduplicated counts (132,273 / 93,592) are the actual on-disk segments and the real training input of the reference classifier and of our companion studies [19,20,21,22]. The front-end also implements LOFAR and DEMON channels, which are not part of the released default but are available in the code.

Figure 2. Construction pipeline. Six-stage flowchart of the fully automatic pipeline (deployment query → AIS parsing → static-table propagation and distance filtering → single-vessel scenario identification → audio–AIS interval matching → metadata generation), annotated with the concrete parameters of each stage (inclusion radius 4,000 m; exclusion radius 6,000 m; interpolation granularity 20 s, gap limits 20–1,200 s; minimum scenario duration 30 min; interval padding ±1 min; audio matching tolerance ±5 min), and with per-stage artifact counts (1,020 candidate clips → 525 released clips).

Figure 2. Construction pipeline. Six-stage flowchart of the fully automatic pipeline (deployment query → AIS parsing → static-table propagation and distance filtering → single-vessel scenario identification → audio–AIS interval matching → metadata generation), annotated with the concrete parameters of each stage (inclusion radius 4,000 m; exclusion radius 6,000 m; interpolation granularity 20 s, gap limits 20–1,200 s; minimum scenario duration 30 min; interval padding ±1 min; audio matching tolerance ±5 min), and with per-stage artifact counts (1,020 candidate clips → 525 released clips).

Figure 3. Deployment geography. (a) Schematic of the Strait of Georgia / Fraser River delta release site showing the ICLISTENAF2523 LSBBL deployment position (49.080927°N, 123.338713°W, 141 m), the 4 km inclusion and 6 km exclusion circles, and representative AIS vessel tracks colored by derived class (tracks decoded from the surviving raw-AIS sample of 2016-09-02; schematic rendering without coastline basemap); (b) the eight registered deployment periods, with the seven further periods available for v2 extension shown in a zoomed inset.

Figure 3. Deployment geography. (a) Schematic of the Strait of Georgia / Fraser River delta release site showing the ICLISTENAF2523 LSBBL deployment position (49.080927°N, 123.338713°W, 141 m), the 4 km inclusion and 6 km exclusion circles, and representative AIS vessel tracks colored by derived class (tracks decoded from the surviving raw-AIS sample of 2016-09-02; schematic rendering without coastline basemap); (b) the eight registered deployment periods, with the seven further periods available for v2 extension shown in a zoomed inset.

Data Records

The dataset is released on Zenodo: https://doi.org/10.5281/zenodo.22274549 (concept DOI 10.5281/zenodo.22274549, always resolving to the latest version; the current release, version v1.1.0: https://doi.org/10.5281/zenodo.22282021 (v1.0.0: https://doi.org/10.5281/zenodo.22274550)). ONC-owned data are licensed under CC-BY 4.0 (reuse and redistribution permitted with ONC attribution); the published package therefore contains the complete WAV audio (twelve class-wise zip archives with a SHA256SUMS.txt checksum manifest), the metadata registry, the frozen split, the precomputed log-mel features with manifests, the zone maps, and all code; the data artifacts are released under CC-BY 4.0 with ONC attribution, and the pipeline code retains the upstream MIT License (code/LICENSE). The WAV clips are redistributable under the same license; since version v1.1.0 they are shipped directly with the record as twelve class-wise zip archives (about 25 GB in total) with a SHA256SUMS.txt checksum manifest, and the released pipeline can still regenerate them from the ONC archive as an independent verification route. If any individual data stream carries partner-shared restrictions, the annotation on the ONC dataset page prevails (verified against the official ONC Data Policy). Layout (paths relative to the release root):

Audio records. Each released clip is a single-channel 32 kHz WAV file containing one single-vessel scenario (vessel clips; median duration 300 s, range 119–2,220 s) or one background interval. Total released audio is approximately 56.0 hours (vessel classes 45.1 h; background 10.8 h). The identical counts of 1,020 scenario-matched WAV candidates and 1,020 daily CTD files are an independent coincidence, not a copy error: the WAV count covers one hydrophone deployment year (2016-05-09 to 2017-05-28; the pipeline log records 1,030 planned downloads, of which 5 failed; the remaining 5-file gap to the 1,020 candidates is filename-level deduplication of the needed-file list — the same archive WAV requested by more than one scenario interval is stored only once — and the final completeness check found all 1,020 needed files present, zero missing), whereas the CTD files are 1,020 non-overlapping daily casts (2016-05-02 to 2024-03-15) from an independent instrument.

Metadata registry. metadata.csv (525 rows, 19 columns) is the authoritative clip-level registry; its fields are defined in Table 3. Vessel-identifying information (MMSI) is retained deliberately: MMSI numbers are broadcast in the clear over public VHF and are required for hull-level split isolation; users who prefer anonymized labels can drop the column without affecting any other field.

Split record. splits/vtuad_v1.json lists the exact file membership of each partition, the splitting parameters, the self-check results (all passed), and the SHA-256 checksum. features/vtuad/manifest_{split}.csv maps every log-mel segment to its source file, recording group, label, and segment offset, so any experiment is fully traceable back to a UTC interval and an MMSI. One integrity caveat follows from the filename collision described in Methods: for the eight recordings whose numeric file IDs collide across the background/ and vessel/ directories, the same-named segment files were overwritten on disk, so the features of 8 of the 369 train recordings are not separately present in the released feature store, and their 1,511 conflicting manifest registrations were dropped from the deduplicated caliber. The companion self-supervised pilot [21] inherits this gap: its recording-identity probe maps only 361 of the 369 train recording IDs (369 − 8).

Label snapshots. Table 2 is computed under the current metadata.csv label column; the feature manifests were generated under an earlier snapshot. The background row comprises the 16 dedicated background/ recordings plus 22 vessel/-directory recordings whose current label is background (38 = 28/5/5 across train/val/test). Between the feature-manifest snapshot and the current label snapshot, 15 vessel/-directory recordings were relabeled from background to vessel classes: vessel-directory background recordings 37 → 22, background-labeled clips 53 → 38. Feature segments are keyed by filename, so the relabeling does not affect the segment store. Under the unified caliber, the classes unaffected by the collision yield ≈2,375–2,396 deduplicated segments per hour (range boundaries subject to rounding of the duration column; nominal 2,400 = 3 s / 1.5 s hop, minus about one truncated tail segment per recording); the deduplicated ratios of tug/cargo/pleasurecraft (≈2,330–2,350 segments/h) sit slightly lower because the 1,511 collision-overwritten segments belong to those classes under the current labels.

Known limitations. The v1 release is one hydrophone at one station (ICLISTENAF2523, Strait of Georgia) over one window (May 2016–May 2017); other sites, other years, and seasonal balance are not sampled. Three classes — pilotvessel, rescue, tanker — have zero test recordings under the frozen split (Table 2) and support only few-shot or metadata-level use. The filename collision described above removes the on-disk features of 8 of the 369 train recordings; those clips remain in the audio release and the split, and their segments can be regenerated with the released front-end if needed. Evaluation-related limitations (weak labels, segment-level correlation, the mixed other class) are discussed in Usage Notes.

Table 2. Class taxonomy and released counts; all columns are computed under the current metadata.csv label column and the frozen split. Recordings = scenario clips; segments = deduplicated 3 s log-mel segments on disk after the unified front-end (the manifests register 1,511 further duplicate train rows from the filename collision described in Methods; registered totals are 133,784 = 95,103/20,479/18,202). Unique MMSI counted over vessel classes. AIS code = type-and-cargo code range of ITU-R M.1371 [10] most frequently associated with the class in the registry. Label-snapshot note: the label snapshots differ (37 → 22; 53 → 38; 15 recordings relabeled); the reconciliation is in the Data Records section. Segment counts are deduplicated (133,784 → 132,273).

Table 2. Class taxonomy and released counts; all columns are computed under the current metadata.csv label column and the frozen split. Recordings = scenario clips; segments = deduplicated 3 s log-mel segments on disk after the unified front-end (the manifests register 1,511 further duplicate train rows from the filename collision described in Methods; registered totals are 133,784 = 95,103/20,479/18,202). Unique MMSI counted over vessel classes. AIS code = type-and-cargo code range of ITU-R M.1371 [10] most frequently associated with the class in the registry. Label-snapshot note: the label snapshots differ (37 → 22; 53 → 38; 15 recordings relabeled); the reconciliation is in the Data Records section. Segment counts are deduplicated (133,784 → 132,273).

Class	AIS code (typical)	Recordings (train/val/test)	Total recordings	Segments	Duration (h)	Unique MMSI
tug	31/32/52 (towing)	131 / 28 / 29	188	43,970	18.83	56
cargo	70–79 (cargo)	68 / 14 / 15	97	20,023	8.52	45
passengership	60–69 (passenger)	50 / 11 / 10	71	13,049	5.47	10
background	—	28 / 5 / 5	38	25,923	10.82	—
other	0 / unclassifiable	27 / 7 / 6	40	8,440	3.53	8
fishing	30	24 / 5 / 5	34	8,446	3.53	21
pleasurecraft	36/37	23 / 5 / 4	32	6,567	2.82	18
dredger	33	6 / 1 / 1	8	1,712	0.72	1
sailing	36	4 / 1 / 1	6	1,354	0.57	5
tanker	80–89	3 / 2 / 0	5	1,315	0.55	4
rescue	51	3 / 1 / 0	4	1,116	0.47	3
pilotvessel	50	2 / 0 / 0	2	358	0.15	1
Total		369 / 80 / 76	525	132,273	55.97	172 (hulls; no MMSI appears in two classes)

Table 3. Metadata registry (metadata.csv) field dictionary.

Table 3. Metadata registry (metadata.csv) field dictionary.

Field	Type	Definition
label	string	Class label (Table 2)
duration_sec	float	Clip duration in seconds
path	string	Relative path of the WAV file
sample_rate	int	Sampling rate (32,000 Hz for all clips)
class_code	float	AIS type-and-cargo code of the labeling message [10]
date	int	Scenario date, YYYYMMDD (UTC)
MMSI	float	Maritime Mobile Service Identity of the scenario vessel
length, beam	float	Vessel length / beam (m), from AIS static messages (0 = not reported)
t1, c1, p1, sal, sv	float	Clip-averaged CTD covariates: temperature (°C), conductivity (S/m), pressure (dbar), salinity (PSU), sound speed (m/s)
*_norm	float	Min–max-normalized versions of the five CTD covariates

Figure 4. Class distribution. (a) Number of scenario clips per class on a log scale, stacked by split (train/val/test); (b) released audio hours per class. The figure documents the long-tailed design (tug 188 clips / 18.83 h vs. pilotvessel 2 clips / 0.15 h) and the retained tail classes.

Figure 4. Class distribution. (a) Number of scenario clips per class on a log scale, stacked by split (train/val/test); (b) released audio hours per class. The figure documents the long-tailed design (tug 188 clips / 18.83 h vs. pilotvessel 2 clips / 0.15 h) and the retained tail classes.

Figure 5. Temporal and durational coverage. (a) Histogram of clip durations (median 300 s, range 119–2,220 s); (b) monthly clip counts over the release window (May 2016–May 2017), vessel vs. background; (c) clip-averaged CTD sound speed vs. temperature scatter, colored by class, showing the environmental spread covered by the corpus.

Figure 5. Temporal and durational coverage. (a) Histogram of clip durations (median 300 s, range 119–2,220 s); (b) monthly clip counts over the release window (May 2016–May 2017), vessel vs. background; (c) clip-averaged CTD sound speed vs. temperature scatter, colored by class, showing the environmental spread covered by the corpus.

Figure 6. Representative log-mel spectrograms. A 4 × 3 grid of ≈10 s log-mel excerpts, one per class (background included), drawn from the test partition where available (tanker and rescue from validation, pilotvessel from training — these classes have no test segments), illustrating inter-class spectral diversity (e.g., tonal blade-rate lines for passengership, broadband cavitation for tug) and within-class variability across hulls.

Figure 6. Representative log-mel spectrograms. A 4 × 3 grid of ≈10 s log-mel excerpts, one per class (background included), drawn from the test partition where available (tanker and rescue from validation, pilotvessel from training — these classes have no test segments), illustrating inter-class spectral diversity (e.g., tonal blade-rate lines for passengership, broadband cavitation for tug) and within-class variability across hulls.

References

1. Santos-Domínguez, D.; Torres-Guijarro, S.; Cardenal-López, A.; Pena-Gimenez, A. ShipsEar: An Underwater Vessel Noise Database. Applied Acoustics 2016, 113, 64–69. [CrossRef]
2. Irfan, M.; Jiangbin, Z.; Ali, S.; Iqbal, M.; Masood, Z.; Hamid, U. DeepShip: An Underwater Acoustic Benchmark Dataset and a Separable Convolution Based Autoencoder for Classification. Expert Systems with Applications 2021, 183, 115270. [CrossRef]
3. Li, Z.; Xiang, S.; Yu, T.; Gao, J.; Ruan, J.; Hu, Y.; Liu, T.; Fu, Y. Oceanship: A Large-Scale Dataset for Underwater Audio Target Recognition. In Proceedings of the International Conference on Intelligent Computing (ICIC 2024); Springer: Singapore, 2024; pp. 475–486. arXiv:2401.02099.
4. Du, X.; Hong, F. QiandaoEar22: A High-Quality Noise Dataset for Identifying Specific Ship from Multiple Underwater Acoustic Targets Using Ship-Radiated Noise. EURASIP Journal on Advances in Signal Processing 2024, 2024, 96. arXiv:2406.04354.
5. Sayigh, L.; Daher, M.A.; Allen, J.; Gordon, H.; Joyce, K.; Stuhlmann, C.; Tyack, P. The Watkins Marine Mammal Sound Database: An Online, Freely Accessible Resource. Proceedings of Meetings on Acoustics 2016, 27, 040013. [CrossRef]
6. Xie, K.; Yang, J.; Qiu, K. A Dataset with Multibeam Forward-Looking Sonar for Underwater Object Detection. Scientific Data 2022, 9, 739. [CrossRef]
7. NOAA Office of National Marine Sanctuaries and U.S. Navy. SanctSound Raw Passive Acoustic Data; NOAA National Centers for Environmental Information, 2021. [CrossRef]
8. Torralba, A.; Efros, A.A. Unbiased Look at Dataset Bias. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2011; pp. 1521–1528.
9. Hummel, H.I.; Bhulai, S.; van der Mei, R.; Ghani, B. Linear Probing Enables Ship-Radiated Noise Recognition with Pretrained Audio Embeddings. Ecological Informatics 2026, 95, 103709. [CrossRef]
10. International Telecommunication Union. Technical Characteristics for an Automatic Identification System Using Time-Division Multiple Access in the VHF Maritime Mobile Frequency Band; Recommendation ITU-R M.1371-5; ITU: Geneva, Switzerland, 2014.
11. Domingos, L.C.F.; Santos, P.E.; Skelton, P.S.M.; Brinkworth, R.S.A.; Sammut, K. An Investigation of Preprocessing Filters and Deep Learning Methods for Vessel Type Classification With Underwater Acoustic Data. IEEE Access 2022, 10, 117582–117596. [CrossRef]
12. Yuan, H.; Wang, W. Real-Time Underwater Acoustic Target Detection with DEIMv2: Benchmarking, Weighted-Box-Fusion Test-Time Augmentation, and Domain Shift Analysis on the UATD Dataset. Preprints 2026, 202609.0449.v1. (companion paper). [CrossRef]
13. Ocean Networks Canada data archive. Available online: https://data.oceannetworks.ca (accessed 2026).
14. pyais: AIS message decoding in Python. Available online: https://github.com/M0r13n/pyais (the released package uses the pyais-compatible fork lpais 0.0.2, see code/requirements.txt).
15. McFee, B.; Raffel, C.; Liang, D.; Ellis, D.P.W.; McVicar, M.; Battenberg, E.; Nieto, O. librosa: Audio and Music Signal Analysis in Python. In Proceedings of the 14th Python in Science Conference (SciPy), 2015; pp. 18–24.
16. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016; pp. 770–778.
17. Park, D.S.; Chan, W.; Zhang, Y.; Chiu, C.-C.; Zoph, B.; Cubuk, E.D.; Le, Q.V. SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition. In Proceedings of Interspeech 2019; pp. 2613–2617.
18. Ainslie, M.A.; McColm, J.G. A Simplified Formula for Viscous and Chemical Absorption in Sea Water. The Journal of the Acoustical Society of America 1998, 103, 1671–1672. [CrossRef]
19. Yuan, H.; Wang, W. Unsupervised Domain Adaptation for Cross-Water Underwater Acoustic Target Recognition: A Four-Corpus Benchmark on Oceanship, QiandaoEar22, an AIS-Auto-Labeled VTUAD Reconstruction, and ShipsEar. Manuscript in preparation (companion paper), 2026.
20. Yuan, H.; Wang, W. When Underwater Acoustic Recognition Fails Across Datasets: A Cross-Dataset Benchmark Revealing Zero-Transfer and Label Shift. Manuscript in preparation (companion paper), 2026.
21. Yuan, H.; Wang, W. Self-Supervised Pretraining for Underwater Acoustic Target Recognition: A Three-Corpus SimCLR Pilot with a Sobering Verdict. Manuscript in preparation (companion paper), 2026.
22. Yuan, H.; Wang, W. Open-Set Recognition for Underwater Acoustic Target Recognition: A Six-Protocol Pilot Exposing Confidence Inversion Across Water Areas. Manuscript in preparation (companion paper), 2026.