Automated data curation for self-supervised learning in underwater acoustic analysis

Burooj Ghani; Hilde I Hummel; Rob van der Mei; Sandjai Bhulai

REVIEW 4 major objections 5 minor 20 references

Automated data curation for self-supervised learning in underwater acoustic analysis

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Automatic curation of raw underwater audio, using AIS ship tracks and hierarchical k-means sampling, lets a self-supervised model beat random sampling of the same data size on ship-type classification.

desk verdict First PAM curation pipeline combining AIS and hierarchical k-means; plausible but single-run evaluation and label-circularity make the headline gain unproven. read the letter →

arxiv 2505.20066 v1 pith:XRJIYVZ4 submitted 2025-05-26 cs.SD cs.AIeess.AS

Hilde I Hummel , Sandjai Bhulai , Burooj Ghani , Rob van der Mei This is my paper · ORCID

classification cs.SDcs.AIeess.AS

keywords self-supervisedlearningdatacurationpassiveacousticmonitoringunderwateracousticsAIShierarchicalk-meansshiptypeclassificationData2Vec

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that unlabeled passive acoustic monitoring data can be curated automatically, without human labels, enough to train self-supervised models that generalize to ship-type classification. It aligns publicly available hydrophone recordings with AIS ship-tracking data, flattens the long-tailed distribution of ship occurrences, and selects diverse audio windows using hierarchical k-means clustering. Training Data2Vec on the resulting curated dataset outperforms training on a random dataset of the same size: accuracy on Deepship rises from 49.16% to 56.72%, and on ShipsEar from 51.98% to 53.11%. The paper positions this as the first automatic passive-acoustic-monitoring data curation pipeline for self-supervised learning, and the implication is that data curation, not just data volume, is what unlocks self-supervised learning on raw underwater recordings.

What carries the argument

The two load-bearing mechanisms are AIS-based balancing and hierarchical k-means audio curation. In the first, each ship's occurrence count in the aligned recordings is computed; a threshold t = 250 at the knee of the skewed distribution splits the data, and ships appearing more often are sampled with lower probability, so the curated AIS portion is flatter than the raw long tail. In the second, 10-second audio windows embedded by a pretrained underwater-audio model are clustered into four levels of sizes [6000, 400, 40, 10] using streaming k-means, and samples closest to cluster centers are retained until the target size N is reached, forcing coverage of the raw distribution. The curated pretraining set is the sum of the AIS-balanced and audio-curated samples, and evaluation uses the Data2Vec self-supervised framework with a momentum-updated teacher.

What would settle it

Run the same Data2Vec comparison with the AIS-to-audio alignment based on modeled transmission loss instead of a fixed 4 km by 4 km square; if the curated-versus-random accuracy gap shrinks, reverses, or disappears, the square assumption is load-bearing. A direct acoustic check is to place a controlled sound source just outside the square at one hydrophone and verify whether it appears in the recordings, or to check a hydrophone in a deep-sound-channel region where ships far beyond the square dominate the spectrum.

Watch

Extended reading notes

Core claim

The central claim is that a fully automated curation pipeline integrating AIS metadata with hierarchical k-means clustering yields a diverse, balanced pretraining set for self-supervised underwater audio analysis. The method counts each ship's occurrence in aligned audio, applies a threshold at the knee of the distribution (t = 250), sub-samples over-represented ships with probability inversely proportional to their incidence, and then curates raw audio by streaming hierarchical k-means with cluster sizes [6000, 400, 40, 10], keeping the 10-second windows closest to cluster centers up to a target size. The resulting dataset, roughly 970 hours, is used to fine-tune Data2Vec, and the curated model beats a randomly sampled model of the same size on both Deepship and ShipsEar. The paper reads the larger Deepship gain as evidence that the acoustic environment of the pretraining data matters for downstream transfer.

Load-bearing premise

The pipeline assumes that a ship's presence inside a fixed 4 km by 4 km square around a hydrophone determines whether it is in the audio, and that ships outside the square are not in the audio; underwater sound propagation actually depends on depth, temperature, and seabed, so the square is only a rough proxy.

Editorial extensions

If this is right

A fully automated curation pipeline can turn large unlabeled passive-acoustic-monitoring archives into pretraining data, removing the manual labeling bottleneck for underwater acoustic analysis.
Self-supervised models trained on curated underwater audio can classify ship types better than the same model trained on a random sample of the same size, with a 7.56-point gain on Deepship.
The smaller gain on ShipsEar indicates that environmental similarity between pretraining and target data affects how much curation helps, so curation and region-matched pretraining should be planned together.
The curated dataset is large enough, about 970 hours, to support downstream tasks such as monitoring marine mammals and assessing sound pollution, not just ship classification.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct ablation separating AIS-driven balancing from audio clustering would say which mechanism carries the accuracy gain; the paper reports only their combination.
If the fixed 4 km square were replaced by a transmission-loss-based alignment, the curation benefit might change; testing this would tell whether the balancing reflects acoustic reality or only the proxy.
The same pipeline could be applied to other self-supervised objectives, such as contrastive learning, which the paper itself suggests may fit stationary underwater sound better than the masking approach.
Because the curated embeddings are learned without labels, they may support few-shot or zero-shot identification of non-ship sounds, for example marine mammal vocalizations, if evaluated on labeled bioacoustic sets.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know this paper is the first to propose an automated data curation pipeline for passive acoustic monitoring (PAM) data, combining AIS ship metadata with hierarchical k-means sampling to build balanced SSL training sets. That framing is new and the pipeline is described clearly: 16 kHz resampling, 10s windows, embeddings from their own contrastive model [13], streaming KMeans with cluster sizes [6000, 400, 40, 10], and an AIS threshold set at the knee of the ship-occurrence distribution. The authors also acknowledge limitations (masking suboptimal for stationary underwater sounds, need for more hydrophones). Credit where due: this is a sensible adaptation of image-curation ideas [10,11] and metadata alignment [4,12] to a domain where labeled data is scarce, and the Deepship gain (49.16% to 56.72%) is nontrivial if it holds.

But it may not hold. The evaluation is a single run per condition with no error bars, and the ShipsEar difference is only 1.13 points, well within typical Data2Vec seed noise. The stress-test note is on target: no replication means the headline result could shrink, reverse, or vanish. There is also a circularity concern the authors do not address: curation balances ship types using AIS, and the only downstream task is ship-type classification, so part of the improvement may be alignment between the curation signal and the evaluation label. The AIS 4 km x 4 km square is a crude proxy for acoustic audibility, but that may be a minor issue for the headline because the PAM clustering branch contributes roughly 93% of the curated sample (323,532 of about 348,553 windows); a poor AIS alignment would weaken the AIS branch but not necessarily collapse the result. More problematic is that the threshold is chosen post hoc from the training distribution, and the random baseline is underspecified.

Overall, the central claim is plausible but not established. The paper is worth a serious referee because it opens a useful direction and the pipeline is concrete, but it needs revision: multiple seeds or at least error bars, an ablation separating AIS and clustering contributions, and ideally a downstream task not so closely tied to the curation signal. I'd bring it to a reading group as a discussion piece, but I wouldn't cite it in my own work until the evidence is stronger. Recommend: send to peer review, but with major revision expected.

Referee Report

4 major / 5 minor

Summary. The paper proposes a fully automated data curation pipeline for self-supervised learning (SSL) on passive acoustic monitoring (PAM) data. The pipeline combines two curation branches: AIS metadata alignment, which samples ship-associated audio to balance ship-type occurrence, and a hierarchical k-means clustering branch that selects diverse audio samples from raw NOAA PAM recordings. The curated dataset D* is then used to fine-tune a Data2Vec model, and its embeddings are evaluated by logistic-regression classification of ship types on the Deepship and ShipsEar benchmarks. The headline results are that the curated model improves accuracy over a randomly curated same-size baseline from 49.16% to 56.72% on Deepship and from 51.98% to 53.11% on ShipsEar (Table 1).

Significance. If the reported improvements are robust, the paper would make a useful contribution as the first automatic data curation pipeline specifically designed for underwater SSL, and the integration of AIS metadata with cluster-based sampling is a sensible and novel combination. The evaluation on two public benchmarks is appropriate for the proposed task. However, the strength of the evidence is currently limited by the absence of repeated runs or error bars, the underspecified random baseline, and the unvalidated AIS-range assumption. The paper should be credited for presenting a concrete, reproducible pipeline description and for identifying a real gap in underwater acoustic SSL, but the central quantitative claim needs stronger experimental support before the results can be taken as established.

major comments (4)

[§4.2, Table 1] The headline comparison rests on a single accuracy number per condition, without error bars, confidence intervals, or repeated runs. Data2Vec pretraining is stochastic through masking, batch order, and initialization from a speech checkpoint, so the ShipsEar gap of 1.13 percentage points (53.11% versus 51.98%) is within typical run-to-run variation, and even the larger Deepship gap of 7.56 points has no measure of uncertainty. The authors should report means and standard deviations over at least three to five independent pretraining runs and, ideally, a paired test, because the central claim is precisely that curation improves SSL accuracy.
[§3.5, §4.2] The random baseline is not specified precisely. The phrase 'randomly curated dataset of the same size' does not state whether the baseline is a uniform random sample of 10-second windows from D, whether it preserves the per-hydrophone and temporal distribution of the curated set, or whether it is matched on total duration. Without this specification, the improvement attributed to curation could instead reflect differences in dataset composition, such as an overrepresentation of empty or low-signal windows in the random baseline.
[§3.2, Figure 2] The AIS-PAM alignment assumes that every ship within a fixed 4 km by 4 km square centered on the hydrophone is recorded in the audio and that ships outside the square are irrelevant. Underwater acoustic propagation depends on depth, temperature, bottom topography, and source level, so this fixed square is a crude proxy that needs validation or a sensitivity analysis. Since AIS-based curation is explicitly one of the two key contributions, the choice of range should be justified with acoustic evidence or tested over at least a few range values; the threshold t is also selected visually from the knee of the distribution (§4.1), so a sensitivity analysis would strengthen the claim.
[§3.6, §4.2] The downstream evaluation is ship-type classification, while the AIS curation branch explicitly balances ship types using AIS metadata. The measured improvement on ship-type classification may therefore reflect the curation's direct alignment with the target task rather than a general improvement in representation quality. To support the broader claim that the pipeline benefits 'various tasks such as monitoring marine mammals and assessing sound pollution,' the authors should evaluate on at least one additional downstream task that is not ship-type classification, or otherwise argue why ship-type classification is a representative transfer probe.

minor comments (5)

[§3.5] The target dataset size N is never defined; the reader only learns the final sizes in §4.1. The authors should state N explicitly and explain how it is chosen.
[§3.5] The text says 'The entire dataset D is employed to optimize a hierarchical KMeans model' and then says the model is 'optimized in a streaming manner'; this is potentially contradictory and should be clarified.
[§3.6] The phrase 'completely fine-tuned' is ambiguous: it is not clear whether the logistic-regression classifier operates on fixed embeddings from the fine-tuned model or on embeddings after further fine-tuning of the encoder. Please clarify the exact evaluation protocol.
[§4.2] The statement that the smaller ShipsEar gain is 'likely due to environmental factors' is speculation with no supporting evidence; either temper the claim or provide an analysis, such as source-environment distance measurements or feature visualization.
[Figure 3] The axis label 'Frequency' is unclear; the plot appears to show counts of 10-second windows per individual ship, so the y-axis should be labeled accordingly.

Circularity Check

1 steps flagged · score 4.0 of 10

Curation pipeline leans on the authors' own embedding model [13]; the curated-vs-random comparison itself is non-circular.

self citation load bearing [Section 3.4 (Audio preprocessing) and Section 3.5 (PAM data curation), supported by reference [13]]
"From this windowed audio, embeddings were generated using the model presented in [13]. In this research, they state that this model is optimized on a large quantity of unlabeled underwater acoustic data and generates generalized embeddings with dimension 2048. The entire dataset D is employed to optimize a hierarchical KMeans model using the resampling-clustering technique."

The PAM-data-curation branch selects 323,532 of the 348,553 final samples using hierarchical k-means on embeddings produced by [13]. Reference [13] is by three of the four present authors (Hummel, van der Mei, Bhulai), is an unreviewed SSRN preprint, and is invoked only as 'they state that this model is optimized on a large quantity of unlabeled underwater acoustic data.' No external, machine-checked, or code-reproduced validation of those embeddings is offered. Thus the diversity/balance of the dominant PAM branch rests on a self-citation whose own validity is not established; the pipeline's main ingredient reduces to the authors' prior model rather than to an independent first-principles representation.

full rationale

The central empirical claim—that Data2Vec trained on the curated dataset outperforms Data2Vec trained on a random dataset of the same size on Deepship and ShipsEar—is a direct comparison on external benchmarks, with the same SSL framework and dataset size. The curation does not use downstream labels to train the encoder; AIS metadata is used only to balance the pretraining pool, and the final classification is a linear probe on the learned Data2Vec embeddings. That part of the derivation is self-contained and not forced by construction. The absence of error bars or repeated runs is a reproducibility and statistical-validity concern, not a circularity step. However, the PAM-data-curation branch depends entirely on embeddings from [13], a self-citation by three of the four authors, presented without independent verification. Since this branch supplies roughly 93% of the curated samples, the load-bearing quality of the largest curation component is justified by the authors' own prior work, warranting a moderate circularity score. No other circular steps were identified in the paper.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The pipeline depends on several hand-picked numbers: the AIS threshold t=250, the cluster size ladder, the 4 km alignment square, and an implicit target dataset size. The core algorithm is borrowed from image curation literature, while the acoustic domain assumptions about AIS range and embedding quality are not independently verified.

free parameters (4)

AIS threshold t = 250
Chosen by visually identifying the knee of the ship-occurrence distribution in the aligned dataset (Section 4.1, Figure 3). This is a free parameter fitted to the data.
Hierarchical cluster sizes = [6000, 400, 40, 10]
No justification provided; these four levels determine the granularity of sampling and are selected by the authors.
AIS alignment range = 4 km x 4 km square
Assumed range of the hydrophone; no acoustic modeling supports this choice (Section 3.2).
Target dataset size N = 323,532 PAM samples (plus 25,021 AIS samples)
The target size is not explicitly stated; the resulting sample counts are reported in Section 4.1.

assumptions (4)

standard math Standard k-means and its streaming approximation converge to a useful clustering of the embedding space.
The hierarchical k-means algorithm is treated as a reliable tool (Sections 3.5 and [14]).
domain assumption Ships within the 4 km square are the dominant acoustic sources in the aligned 10-second windows.
The entire AIS curation step depends on this geometric alignment (Section 3.2).
domain assumption Uniform distribution over the support of the raw PAM data is the right target for a diverse dataset.
The objective stated in Section 3.5 is to approach a uniform distribution without empirical support that it improves downstream tasks.
domain assumption The embedding model of [13] produces features that represent acoustic diversity suitable for clustering.
The curation pipeline relies entirely on these embeddings, which are computed by a model from the same authors that is not described in detail.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Automated data curation for self-supervised learning in underwater acoustic analysis." pith.science (2026). https://pith.science/paper/XRJIYVZ4

@misc{pith2026250520066,
  author       = {Pith},
  title        = {Pith review of: Automated data curation for self-supervised learning in underwater acoustic analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XRJIYVZ4}},
  note         = {Machine review of arXiv:2505.20066}
}

read the original abstract

The sustainability of the ocean ecosystem is threatened by increased levels of sound pollution, making monitoring crucial to understand its variability and impact. Passive acoustic monitoring (PAM) systems collect a large amount of underwater sound recordings, but the large volume of data makes manual analysis impossible, creating the need for automation. Although machine learning offers a potential solution, most underwater acoustic recordings are unlabeled. Self-supervised learning models have demonstrated success in learning from large-scale unlabeled data in various domains like computer vision, Natural Language Processing, and audio. However, these models require large, diverse, and balanced datasets for training in order to generalize well. To address this, a fully automated self-supervised data curation pipeline is proposed to create a diverse and balanced dataset from raw PAM data. It integrates Automatic Identification System (AIS) data with recordings from various hydrophones in the U.S. waters. Using hierarchical k-means clustering, the raw audio data is sampled and then combined with AIS samples to create a balanced and diverse dataset. The resulting curated dataset enables the development of self-supervised learning models, facilitating various tasks such as monitoring marine mammals and assessing sound pollution.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 17 canonical work pages

[13]

Self-supervised learning-for un- derwater acoustic signal classification with mixup,

Q. Xu, J. Jiang, K. Xu, Y . Dou, C. Gao, B. Zhu, K. You, and Q. Zhu, “Self-supervised learning-for un- derwater acoustic signal classification with mixup,” IEEE Journal of Selected Topics in Applied Earth Ob- servations and Remote Sensing , vol. 17, pp. 3530– 3542, 2024

work page 2024
[1]

Automated data curation for self-supervised learning in underwater acoustic analysis

INTRODUCTION The increasing levels of sound pollution threaten the preservation of ocean ecosystems, necessitating the moni- toring of underwater sounds [1]. Passive Acoustic Mon- itoring (PAM) systems are globally deployed and col- lect a vast amount of diverse underwater sound record- ings. The complexity of the marine environment, com- bined with the l...

work page Pith review arXiv 2025
[2]

With no large, curated underwater acoustic dataset publicly avail- able yet, [7] and [8] proposed pretraining on AudioSet applying a mix-up strategy

RELA TED WORK 2.1 SSL in underwater acoustics Several previous studies have shown the potential of SSL in the automatic recognition of ship types [1]. With no large, curated underwater acoustic dataset publicly avail- able yet, [7] and [8] proposed pretraining on AudioSet applying a mix-up strategy. Next, a Swin Transformer encoder is optimized with maske...

work page
[3]

All these PAM audio recordings are combined into the result- ing dataset D

METHODS 3.1 PAM data The raw audio data is collected from NOAA 1 , selecting hydrophones which started recording in 2023 or later. All these PAM audio recordings are combined into the result- ing dataset D. In total, this combination covers the dura- tion of 8 years, 6 months, 9 days, 15 hours, 19 minutes, and 49 seconds from 11 individual hydrophones. Th...

work page 2023
[4]

Figure 3 illustrates the number of 10-second audio windows per individual ship, revealing a skewed distribution

RESULTS 4.1 Data curation For the AIS curation method, the optimal threshold value t is defined to correspond to the knee of the skewed dis- tribution of Ds [4]. Figure 3 illustrates the number of 10-second audio windows per individual ship, revealing a skewed distribution. The optimal threshold, aligning with the knee of the distribution, is around 250. ...

work page 2025
[5]

The study demonstrates that curation is a key aspect in extracting accurate SSL model representations from unlabeled un- derwater recordings

CONCLUSION This work describes the first automatic data curation pipeline to curate large web-scraped PAM data. The study demonstrates that curation is a key aspect in extracting accurate SSL model representations from unlabeled un- derwater recordings. Although this work focuses on data curation, more research is still required on SSL methods applied to ...

work page
[6]

A survey on machine learning in ship radiated noise,

H. I. Hummel, R. D. van der Mei, and S. Bhulai, “A survey on machine learning in ship radiated noise,” Ocean Engineering, vol. 298, p. 117252, 2024

work page 2024
[7]

A survey on self-supervised learning: Algorithms, applications, and future trends,

J. Gui, T. Chen, J. Zhang, Q. Cao, Z. Sun, H. Luo, and D. Tao, “A survey on self-supervised learning: Algorithms, applications, and future trends,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 46, no. 12, p. 9052–9071, 2024

work page 2024

Show all 20 references

[8]

Audio self-supervised learning: A survey,

S. Liu, A. Mallol-Ragolta, E. Parada-Cabaleiro, K. Qian, X. Jing, A. Kathan, B. Hu, and B. W. Schuller, “Audio self-supervised learning: A survey,” Patterns, vol. 3, no. 12, p. 100616, 2022

2022
[9]

Dinov2 meets text: A unified framework for image-and pixel-level vision-language alignment,

C. Jose, T. Moutakanni, D. Kang, F. Baldassarre, T. Darcet, H. Xu, D. Li, M. Szafraniec, M. Ramamon- jisoa, M. Oquab, et al., “Dinov2 meets text: A unified framework for image-and pixel-level vision-language alignment,” arXiv preprint arXiv:2412.16334, 2024

2024 arXiv
[10]

Deepship: An underwater acoustic benchmark dataset and a separable convolution based autoencoder for classification,

M. Irfan, Z. Jiangbin, S. Ali, M. Iqbal, Z. Masood, and U. Hamid, “Deepship: An underwater acoustic benchmark dataset and a separable convolution based autoencoder for classification,” Expert Systems with Applications, vol. 183, p. 115270, 2021

2021
[11]

Shipsear: An underwater vessel noise database,

D. Santos-Dom ´ınguez, S. Torres-Guijarro, A. Cardenal-L´opez, and A. Pena-Gimenez, “Shipsear: An underwater vessel noise database,” Applied Acoustics, vol. 113, pp. 64–69, 2016

2016
[12]

Self-supervised learning–based under- water acoustical signal classification via mask mod- eling,

K. Xu, Q. Xu, K. You, B. Zhu, M. Feng, D. Feng, and B. Liu, “Self-supervised learning–based under- water acoustical signal classification via mask mod- eling,” The Journal of the Acoustical Society of Amer- ica, vol. 154, no. 1, pp. 5–15, 2023

2023
[14]

Masking hierarchical tokens for underwater acoustic target recognition with self-supervised learning,

S. Feng, X. Zhu, and S. Ma, “Masking hierarchical tokens for underwater acoustic target recognition with self-supervised learning,”IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 32, pp. 1365–1379, 2024

2024
[15]

DINOv2: Learning Robust Visual Features without Supervision,

M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haz- iza, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.-W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatut, A...

2024
[16]

Automatic data curation for self-supervised learning: A clustering-based ap- proach,

H. V . V o, V . Khalidov, T. Darcet, T. Moutakanni, N. Smetanin, M. Szafraniec, H. Touvron, C. Couprie, M. Oquab, A. Joulin, et al., “Automatic data curation for self-supervised learning: A clustering-based ap- proach,” arXiv preprint arXiv:2405.15613, 2024

2024 arXiv
[17]

Acav100m: Auto- matic curation of large-scale datasets for audio-visual video representation learning,

S. Lee, J. Chung, Y . Yu, G. Kim, T. Breuel, G. Chechik, and Y . Song, “Acav100m: Auto- matic curation of large-scale datasets for audio-visual video representation learning,” in Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pp. 10274–10284, 2021

2021
[18]

The computation of generalized embeddings for underwater acoustic target recogni- tion using contrastive learning,

H. I. Hummel, A. Gansekoele, R. D. van der Mei, and S. Bhulai, “The computation of generalized embeddings for underwater acoustic target recogni- tion using contrastive learning,” Available at SSRN: http://dx.doi.org/10.2139/ssrn.5112948, 2025

doi
[19]

Web-scale k-means clustering,

D. Sculley, “Web-scale k-means clustering,” in Pro- ceedings of the 19th international conference on World wide web , (New York, NY , USA), pp. 1177– 1178, 2010. 11th Convention of the European Acoustics Association M´alaga, Spain • 23rd – 26th June 2025 •

2010
[20]

Data2vec: A general framework for self-supervised learning in speech, vision and lan- guage,

A. Baevski, W.-N. Hsu, Q. Xu, A. Babu, J. Gu, and M. Auli, “Data2vec: A general framework for self-supervised learning in speech, vision and lan- guage,” inInternational conference on machine learn- ing, (Maryland, USA), pp. 1298–1312, 2022

2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [13]

Self-supervised learning-for un- derwater acoustic signal classification with mixup,

Q. Xu, J. Jiang, K. Xu, Y . Dou, C. Gao, B. Zhu, K. You, and Q. Zhu, “Self-supervised learning-for un- derwater acoustic signal classification with mixup,” IEEE Journal of Selected Topics in Applied Earth Ob- servations and Remote Sensing , vol. 17, pp. 3530– 3542, 2024

work page 2024

[2] [1]

Automated data curation for self-supervised learning in underwater acoustic analysis

INTRODUCTION The increasing levels of sound pollution threaten the preservation of ocean ecosystems, necessitating the moni- toring of underwater sounds [1]. Passive Acoustic Mon- itoring (PAM) systems are globally deployed and col- lect a vast amount of diverse underwater sound record- ings. The complexity of the marine environment, com- bined with the l...

work page Pith review arXiv 2025

[3] [2]

With no large, curated underwater acoustic dataset publicly avail- able yet, [7] and [8] proposed pretraining on AudioSet applying a mix-up strategy

RELA TED WORK 2.1 SSL in underwater acoustics Several previous studies have shown the potential of SSL in the automatic recognition of ship types [1]. With no large, curated underwater acoustic dataset publicly avail- able yet, [7] and [8] proposed pretraining on AudioSet applying a mix-up strategy. Next, a Swin Transformer encoder is optimized with maske...

work page

[4] [3]

All these PAM audio recordings are combined into the result- ing dataset D

METHODS 3.1 PAM data The raw audio data is collected from NOAA 1 , selecting hydrophones which started recording in 2023 or later. All these PAM audio recordings are combined into the result- ing dataset D. In total, this combination covers the dura- tion of 8 years, 6 months, 9 days, 15 hours, 19 minutes, and 49 seconds from 11 individual hydrophones. Th...

work page 2023

[5] [4]

Figure 3 illustrates the number of 10-second audio windows per individual ship, revealing a skewed distribution

RESULTS 4.1 Data curation For the AIS curation method, the optimal threshold value t is defined to correspond to the knee of the skewed dis- tribution of Ds [4]. Figure 3 illustrates the number of 10-second audio windows per individual ship, revealing a skewed distribution. The optimal threshold, aligning with the knee of the distribution, is around 250. ...

work page 2025

[6] [5]

The study demonstrates that curation is a key aspect in extracting accurate SSL model representations from unlabeled un- derwater recordings

CONCLUSION This work describes the first automatic data curation pipeline to curate large web-scraped PAM data. The study demonstrates that curation is a key aspect in extracting accurate SSL model representations from unlabeled un- derwater recordings. Although this work focuses on data curation, more research is still required on SSL methods applied to ...

work page

[7] [6]

A survey on machine learning in ship radiated noise,

H. I. Hummel, R. D. van der Mei, and S. Bhulai, “A survey on machine learning in ship radiated noise,” Ocean Engineering, vol. 298, p. 117252, 2024

work page 2024

[8] [7]

A survey on self-supervised learning: Algorithms, applications, and future trends,

J. Gui, T. Chen, J. Zhang, Q. Cao, Z. Sun, H. Luo, and D. Tao, “A survey on self-supervised learning: Algorithms, applications, and future trends,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 46, no. 12, p. 9052–9071, 2024

work page 2024