Pith. sign in

REVIEW 3 major objections 5 minor 25 references

Acoustic Classification of Maritime Vessels using Learnable Filterbanks

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims that learning spectral filters from raw audio, rather than fixing them, lets one model classify ships accurately across varying recording distances, reaching 96.63% test accuracy on the combined VTUAD benchmark where the…

desk verdict A credible SOTA claim on VTUAD that needs the train/test split documented before the 12-point gain can be taken at face value. read the letter →

arxiv 2505.23964 v1 pith:Z3BCQJP2 submitted 2025-05-29 cs.SD cs.LGeess.AS

classification cs.SDcs.LGeess.AS
keywords underwateracousticspassiveacousticmonitoringvesselclassificationlearnablefilterbanksGaborfiltersattentionpoolingenvironmentalsensorfusionVTUADdataset
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

Passive acoustic ship classification usually fails when the same vessel is recorded at different distances: fixed spectrogram features that work well for close-range recordings lose discriminative power as range increases. The paper claims that a classifier which learns its own spectral filters from raw audio—the CATFISH model—solves this problem on the VTUAD benchmark, reaching 96.63% test accuracy on the combined three-range scenario, compared with 84.13% for the original fixed-filter benchmark. The claim matters because real maritime surveillance must handle vessels approaching and receding from a hydrophone, and existing single-scenario models drop by double digits when scenarios are mixed. If true, this would mean that learnable front-ends, not just better backbones or more data, are the key to consistent underwater vessel recognition.

What carries the argument

The load-bearing mechanism is the learnable Gabor filterbank front-end, which replaces a fixed spectrogram with a bank of filters whose parameters are trained together with the rest of the network from raw 1-second waveforms. Because the filter shapes, center frequencies, and bandwidths adapt to the data, the model can emphasize the frequency bands that survive long-distance propagation; the paper shows the active filter range shrinking from roughly channels 0–60 in the closest scenario to roughly 0–10 in the farthest, while those lowest channels stay discriminative in all scenarios. Attention pooling and optional fusion of environmental sensor variables are secondary mechanisms that each add about five percentage points of accuracy, and the ablations show they are largely interchangeable.

What would settle it

Run the VTUAD benchmark's original split procedure and train CATFISH under a strictly transit-disjoint split, checking whether any vessel passage contributes clips to both training and test. If the combined accuracy falls to roughly the 91–92% level observed without attention or environmental fusion, the claimed 12-point gain over the 84.13% benchmark would be attributable to split leakage rather than to the learnable filterbank.

Watch

Extended reading notes

Core claim

CATFISH is an end-to-end vessel classifier whose front-end consists of learnable Gabor filters: each filter has trainable center frequency, bandwidth, Gaussian pooling width, log-gain, and per-band normalization, all optimized jointly with a compact convolutional embedding and a classification head. On the VTUAD multi-scenario task—four vessel classes plus a background class, recorded in three inclusion/exclusion radius scenarios—the paper reports 96.63% test accuracy on the combined test set, a 12.5 percentage point gain over the 84.13% benchmark. Ablations show the learnable front-end alone accounts for 91.59%; adding either 2D attention pooling or the five-parameter environmental metadata (conductivity, temperature, depth, salinity, sound velocity) raises accuracy above 96%, and adding both gives the top result. The paper's analysis of learned filters shows that the tug-versus-background discriminative signal contracts from a wide band of filter channels at close range to the lowest few channels at the farthest scenario, which the authors interpret as the model concentrating on propagation-invariant low-frequency cues.

Load-bearing premise

The fairness of the reported 12-point gain rests on the assumption that the combined-scenario train/test split is exactly the one used for the original 84.13% benchmark, with no 1-second clips from the same vessel passage crossing the train/test boundary; if the split differs or leaks, the comparison is not valid.

Editorial extensions

If this is right

  • A single CATFISH model trained on all three scenarios stays above 96% when evaluated on each scenario separately, so distance variation no longer requires per-scenario retraining.
  • The learnable front-end supplies most of the gain: replacing the fixed filterbank with trainable filters lifts combined accuracy from 84.13% to 91.59% before any attention or metadata is added.
  • Either 2D attention pooling or CTDSV environmental fusion closes nearly all of the remaining gap (96.23% and 96.32% respectively), so the two mechanisms are alternative routes to distance robustness.
  • Tug/background confusion drops from 14% and 6% in the original benchmark to 2.2% and 1.9%, which addresses the main confound identified in the multi-scenario task.
  • The narrowing of discriminative filter channels with distance indicates that the model learns a spectral-compression rule rather than a single fixed template, which is why combined-scenario accuracy remains high.

Reading between the lines

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

  • The paper leaves implicit that the same learnable front-end could replace fixed filterbanks in other passive acoustic tasks with variable source-sensor geometry, such as marine mammal call detection or sonar target classification; the distance-robustness result suggests the benefit is largest exactly where signal-to-noise ratio varies most.
  • Because attention pooling and CTDSV fusion are interchangeable, a testable extension is to check whether the metadata branch is simply helping the model select the same low-frequency cues—if the lowest filter channels are ablated, CTDSV should no longer compensate.
  • The reported robustness is for one geographic region, so a natural extension is to evaluate CATFISH on hydrophone data from a different site or season, where thermocline shifts and seabed composition change the propagation conditions; the authors themselves flag this as a limitation.
  • A practical extension for safety-critical port monitoring would be to add calibrated uncertainty estimates, since the paper reports no confidence calibration and high accuracy on a single region is not evidence of reliable performance on novel vessel classes.
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.

Referee Report

3 major / 5 minor

Summary. The paper presents CATFISH, an end-to-end classifier for underwater vessel acoustics using a learnable Gabor filterbank front-end followed by an EfficientNet-B0 backbone, 2D attention pooling, and optional CTDSV sensor fusion. On the VTUAD benchmark, the authors report 96.63% test accuracy on the combined three-scenario task, exceeding the prior 84.13% benchmark by 12.5 percentage points, and provide ablations separating the contributions of the learnable frontend, attention pooling, and CTDSV fusion. The authors also analyze the learned filters to show that discriminative activity shifts to low-frequency channels as the source-sensor distance increases.

Significance. If the comparison is valid, the result is a significant advance on a realistic multi-scenario underwater vessel classification task, with practical relevance for passive acoustic monitoring. The paper ships code, uses a public dataset, and the ablation design is sensible, with the learnable-frontend-only variant already outperforming the fixed-filter benchmark by 7.5 percentage points. The main caution is that the headline comparison depends on the unstated train/test split protocol, and all numbers are single runs without uncertainty estimates.

major comments (3)
  1. [Results, Implementation and training, Table 1] The central claim of a 12-percentage-point gain over the 84.13% benchmark rests on the comparability of the train/test split, but the manuscript only states that the methodology follows Domingos et al. (2022) without specifying the split ratio, whether the split is at the level of clips or continuous recording sessions, or how stratification by class and scenario was performed. Because VTUAD consists of 1-second clips cut from continuous recordings of vessel passages, a random clip-level split can place highly correlated clips from the same passage on both sides of the train/test boundary, inflating accuracy through near-duplicate memorization. The authors should report the exact split protocol, confirm that it matches Domingos et al., and ideally add a session-grouped or passage-grouped evaluation to show that the reported accuracy is not an artifact of temporal leakage.
  2. [Results, Table 2 and Discussion] All accuracies in Tables 1 and 2 are from single runs. The differences among the main model (96.63%), attention-only (96.23%), CTDSV-only (96.32%), and max-pooling plus CTDSV (96.32%) are small and may be within run-to-run variability. The discussion's conclusion that either attention pooling or CTDSV fusion can inject the necessary information is therefore not yet supported. Please report mean and standard deviation over multiple seeds, or provide per-seed results, at least for the combined-scenario accuracy.
  3. [Methods, Model] The CTDSV fusion branch is not described in sufficient detail. The text says only that normalized environmental variables are processed by a small feed-forward branch and concatenated with the audio embedding, but it does not state the branch architecture, the normalization method, or whether the CTDSV values are constant per scenario or vary per clip. Since the CTDSV ablation is credited with a substantial accuracy gain, the authors should specify these details and also check whether the CTDSV variables are confounded with the recording scenario (e.g., depth or sound velocity differing systematically between S1/S2/S3), which would change the interpretation of the fusion ablation.
minor comments (5)
  1. [Introduction, paragraph 1] There are several typographical issues: 'LOF AR lines' should be 'LOFAR lines', 'T able 1' and 'T able 2' have spacing errors, and 'PUA VC' should be written consistently as 'PUAVC' or 'passive underwater acoustic vessel classification' throughout the manuscript.
  2. [Results, Equations (2) and (3)] The averaging denominators are written as N_T T and N_B T, but N_T, N_B, and T are not defined explicitly; please define the number of clips per class and the number of time frames.
  3. [Results, Figures 4 and 5] Figures 4 and 5 should include labeled colorbars and axis units; the text refers to 'red and blue regions' but the color map is not described in the caption.
  4. [Methods, Dataset and challenge] The paper should state the audio sample rate, the number of clips per class per scenario, and the total train/test sizes, as these are needed to assess class imbalance and reproduce the experiments.
  5. [References] Reference 22 appears to contain a formatting artifact ('Proceedings of the Conference48'); please correct the bibliographic entry.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central 96.63% test-accuracy claim is an empirical result from a model trained and evaluated on held-out VTUAD data, with the comparison benchmark taken from external work; ablations decompose the contributions rather than renaming fitted inputs.

full rationale

The paper's central claim is an empirical test-set accuracy: CATFISH is trained with a joint objective over frontend and classifier parameters (Eq. 6) and evaluated on the VTUAD multi-scenario test set, reporting 96.63% versus the external Domingos et al. (2022) benchmark of 84.13%. This is not a derivation from assumed constants, and no equation defines the predicted accuracy in terms of a fitted parameter or a self-cited result. The learnable Gabor filterbank is inherited from LEAF and EfficientLEAF, which are external prior works, not the authors' own, and the paper does not invoke a uniqueness theorem or self-citation to force its architectural choices. The ablation study further separates the contributions: the model without attention pooling and without CTDSV still reaches 91.59%, while adding either mechanism reaches above 96%, showing that the main result is not simply a renamed fit. The filter activation analysis (Eqs. 1-5) is post hoc and descriptive, not load-bearing for the accuracy claim. The statement that the methodology follows the original benchmark from Domingos et al. (2022) concerns protocol reuse; the absence of explicit train/test split and leakage-control details is a reproducibility and validity concern, not circularity, because the held-out evaluation still breaks any definitional equivalence between training input and reported result. No self-citation, no fitted-input-as-prediction, and no equation-level circular reduction were found. The appropriate finding is therefore no significant circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The model is a standard deep-learning classifier trained on a labeled benchmark; there are no derived constants or invented entities. The key implicit premises are the correctness and split-integrity of the VTUAD dataset and the usual assumptions of gradient-based training.

free parameters (2)
  • Number of Gabor filters K = 96
    Chosen by hand as an architectural hyperparameter; all reported experiments use 96 filters.
  • Number of training epochs = 40
    Chosen by hand; reported for all models, with no early stopping or learning-rate schedule details.
assumptions (3)
  • domain assumption VTUAD dataset labels and scenario assignments are correct and consistent with the original benchmark.
    The evaluation and comparison assume the dataset's annotations and the benchmark's train/test split as defined by Domingos et al. (2022) are used without modification. Introduced in Methods: 'This methodology follows the original benchmark from Domingos et al. (2022).'
  • domain assumption The 1-second clips in training and test splits do not share the same vessel passage (no temporal leakage).
    The paper does not describe how clips are split; if clips from the same vessel event appear in both train and test, the accuracy would be inflated. This premise is implicit in the evaluation.
  • standard math Standard deep learning training assumptions (differentiability, stochastic gradient descent convergence) hold.
    Relies on backpropagation and optimizer convergence, standard in the field.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Acoustic Classification of Maritime Vessels using Learnable Filterbanks." pith.science (2026). https://pith.science/paper/Z3BCQJP2

@misc{pith2026250523964,
  author       = {Pith},
  title        = {Pith review of: Acoustic Classification of Maritime Vessels using Learnable Filterbanks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z3BCQJP2}},
  note         = {Machine review of arXiv:2505.23964}
}
read the original abstract

Reliably monitoring and recognizing maritime vessels based on acoustic signatures is complicated by the variability of different recording scenarios. A robust classification framework must be able to generalize across diverse acoustic environments and variable source-sensor distances. To this end, we present a deep learning model with robust performance across different recording scenarios. Using a trainable spectral front-end and temporal feature encoder to learn a Gabor filterbank, the model can dynamically emphasize different frequency components. Trained on the VTUAD hydrophone recordings from the Strait of Georgia, our model, CATFISH, achieves a state-of-the-art 96.63 % percent test accuracy across varying source-sensor distances, surpassing the previous benchmark by over 12 percentage points. We present the model, justify our architectural choices, analyze the learned Gabor filters, and perform ablation studies on sensor data fusion and attention-based pooling.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 22 canonical work pages

  1. [1]

    Hildebrand, J. A. Anthropogenic and natural sources of ambient noise in the ocean.Marine Ecology Progress Series 395, 5–20 (2009)

  2. [2]

    & Wolf, S

    Badiey, M., Mu, Y., Lynch, J., Apel, J. & Wolf, S. Temporal and azimuthal depen- dence of sound propagation in shallow water with internal waves. IEEE journal of oceanic engineering 27, 117–129 (2002)

  3. [3]

    C., Santos, P

    Domingos, L. C., Santos, P. E., Skelton, P. S., Brinkworth, R. S. & Sammut, K. A sur- vey of underwater acoustic data classification methods using deep learning for shoreline surveillance. Sensors 22, 2181 (2022)

  4. [4]

    & Yang, L

    Wu, Y., Yang, Y., Tao, C., Tian, F. & Yang, L. Robust underwater target recogni- tion using auditory cepstral coefficients in OCEANS 2014-TAIPEI (2014), 1–4

  5. [5]

    Li, Y., Li, Y., Chen, X. & Yu, J. A novel feature extraction method for ship-radiated noise based on variational mode decompo- sition and multi-scale permutation entropy. Entropy 19, 342 (2017)

  6. [6]

    & Pena-Gimenez, A

    Santos-Dom ´ ınguez, D., Torres-Guijarro, S., Cardenal-L´ opez, A. & Pena-Gimenez, A. ShipsEar: An underwater vessel noise database. Applied Acoustics 113, 64–69 (2016). 1https://github.com/Jotels/CATFISH 7

  7. [7]

    & Lee, K

    Choi, J., Choo, Y. & Lee, K. Acoustic classi- fication of surface and underwater vessels in the ocean using supervised machine learning. Sensors 19, 3492 (2019)

  8. [8]

    C., Santos, P

    Domingos, L. C., Santos, P. E., Skelton, P. S., Brinkworth, R. S. & Sammut, K. An inves- tigation of preprocessing filters and deep learning methods for vessel type classifica- tion with underwater acoustic data. IEEE Access 10, 117582–117596 (2022)

Show all 25 references
  1. [9]

    Gao, Y., Chen, Y., Wang, F. & He, Y. Recognition method for underwater acous- tic target based on DCGAN and DenseNet in 2020 IEEE 5th International Conference on Image, Vision and Computing (ICIVC) (2020), 215–221

  2. [10]

    De BA Barros, R. E. & Ebecken, N. F. Devel- opment of a ship classification method based on Convolutional neural network and Cyclo- stationarity Analysis. Mechanical Systems and Signal Processing 170, 108778 (2022)

  3. [11]

    Irfan, M. et al. DeepShip: An underwater acoustic benchmark dataset and a separable convolution based autoencoder for classifica- tion. Expert Systems with Applications 183, 115270 (2021)

  4. [12]

    & Bengio, Y

    Ravanelli, M. & Bengio, Y. Interpretable convolutional filters with sincnet. arXiv preprint arXiv:1811.09725 (2018)

  5. [13]

    Zeghidour, N., Teboul, O., Quitry, F. D. C. & Tagliasacchi, M. LEAF: A learnable fron- tend for audio classification. arXiv preprint arXiv:2101.08596 (2021)

  6. [14]

    & Gutenbrunner, G

    Schl¨ uter, J. & Gutenbrunner, G. Effi- cientleaf: A faster learnable audio frontend of questionable use in 2022 30th European signal processing conference (EUSIPCO) (2022), 205–208

  7. [15]

    & Glass, J

    Gong, Y., Chung, Y.-A. & Glass, J. Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778 (2021)

  8. [16]

    Xu, K. et al. Self-supervised learning–based underwater acoustical signal classification via mask modeling. The Journal of the Acoustical Society of America 154, 5–15 (2023)

  9. [17]

    Kuperman, W. A. & Roux, P. Underwater acoustics. Springer Handbook of Acoustics, 157–212 (2014)

  10. [18]

    L., Ramakrishnan, R., Williams, S

    Ferguson, E. L., Ramakrishnan, R., Williams, S. B. & Jin, C. T. Convolutional neural networks for passive monitoring of a shallow water environment using a single sensor in 2017 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) (2017), 2657–2661

  11. [19]

    & Duan, J

    Li, Y., Xiao, Q., Hu, K., Fang, Y. & Duan, J. Enhancing Underwater Acoustic Signal Classification with CAM++ and Change Point Features in 2024 IEEE 13th Data Driven Control and Learning Systems Con- ference (DDCLS) (2024), 2253–2258

  12. [20]

    Nathala, S. S. et al. Vessel Type Classifica- tion Utilizing Underwater Acoustic Data and Deep Learning in 2024 IEEE 19th Confer- ence on Industrial Electronics and Applica- tions (ICIEA) (2024), 1–6

  13. [21]

    Wei, W. et al. Underwater vessel sound recognition based on multi-layer feature and attention mechanism. Scientific Reports 15, 11239 (2025)

  14. [22]

    & Liebetrau, J

    Gourishetti, S., Grollmisch, S., Abeßer, J. & Liebetrau, J. Potentials and Challenges of AI-based Audio Analysis in Industrial Sound Analysis in Proceedings of the Conference48 (2022)

  15. [23]

    & Liu, J

    Li, J., Wang, B., Cui, X., Li, S. & Liu, J. Underwater acoustic target recognition based on attention residual network.Entropy 24, 1657 (2022)

  16. [24]

    Tan, M. & Le, Q. Efficientnet: Rethink- ing model scaling for convolutional neu- ral networks in International conference on machine learning (2019), 6105–6114

  17. [25]

    Vaswani, A. et al. Attention is all you need. Advances in neural information processing systems 30 (2017). Acknowledgements Authors acknowledge financial support from the Danish National Research Foundation with the Pioneer Center for Accelerating P2X Materials Discovery (CAPe...

Pith tools

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