Pith. sign in

REVIEW 3 major objections 5 minor 52 references

15,500 Seconds: Lean UAV Classification Using EfficientNet and Lightweight Fine-Tuning

T0 review · 3 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 A pretrained EfficientNet-B0, fully fine-tuned with three audio augmentations, identifies 31 drone types from sound with 95.95% validation accuracy.

desk verdict A useful but sloppy empirical sweep; the headline accuracy is not trustworthy until the duplicated table entries and the contradictory split description are fixed. read the letter →

arxiv 2506.11049 v4 pith:UVNYLKLU submitted 2025-05-21 cs.LG cs.AI

Andrew P. Berg , Qian Zhang , Mia Y. Wang This is my paper
classification cs.LGcs.AI
keywords UAVaudioclassificationEfficientNetparameter-efficientfine-tuningtransferlearningdataaugmentationmel-spectrogramdeep
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 tries to settle how to classify drone audio accurately when labeled recordings are scarce, comparing pretrained transformers, CNNs, and parameter-efficient fine-tuning (PEFT) on a custom 31-class dataset of 3,100 five-second clips. It claims that full fine-tuning of a small pretrained EfficientNet-B0, aided by three raw-audio augmentations per training sample, reaches 95.95% mean five-fold validation accuracy, ahead of the Audio Spectrogram Transformer and a custom CNN. The result matters because it suggests a lightweight architecture with standard transfer learning can outperform both larger models and specialized fine-tuning tricks on this data scale. The paper positions this as a continuation of the authors' earlier 9-class study, in which the custom CNN had led.

What carries the argument

The vehicle of the result is the pretrained EfficientNet-B0 convolutional network, adapted to accept grayscale mel-spectrograms, trained end to end with cross-entropy loss and Adam at a fixed 0.001 learning rate. The augmentations — time stretch and sine distortion, three per training sample — are applied to raw audio before feature extraction and inflate only the training set, not the validation or test sets. The paper contrasts this full fine-tuning regime against classifier-only linear probing, batchnorm-only tuning, SSF, Ia3, and OFT, with the EfficientNet-B0 rows showing that all of these methods land within roughly a percentage point of each other when augmentations are used.

What would settle it

Record the same 31 drone models at a new location or with a different microphone and run the best EfficientNet-B0 protocol; if validation accuracy drops substantially from 95.95%, the central claim is specific to the original recording conditions. Alternatively, a leave-one-site-out split on the existing data would reveal whether the model relies on site-specific cues.

Watch

Extended reading notes

Core claim

On the authors' own terms, the central discovery is that the combination of a pretrained EfficientNet-B0, full fine-tuning of all parameters, and three augmentations (time stretch combined with sine distortion) applied to raw audio before mel-spectrogram feature extraction yields the best classification performance on their UAV audio dataset, with 95.95% ± 0.61% validation accuracy over five folds. This beats every PEFT configuration and the transformer baseline; among near competitors, AST with SSF fine-tuning and EfficientNet-B7 with three augmentations reach about 95.27% and 95.66%, respectively. The paper interprets this as evidence that, at this dataset scale, lightweight pretrained CNNs can surpass both custom small CNNs and large pretrained transformers, and that full fine-tuning can still be preferable to PEFT when data is sufficient.

Load-bearing premise

The 95.95% figure is treated as drone-identity classification accuracy, but each drone class was recorded at a single site and the recording hardware changed in 2024, so if room acoustics, background noise, or microphone differences act as learnable shortcuts, the accuracy will not generalize beyond this recording protocol.

Editorial extensions

If this is right

  • On this dataset, a roughly 4-million-parameter pretrained CNN is enough to reach about 96% mean validation accuracy, so model scale is not the main driver of performance.
  • Full fine-tuning of EfficientNet-B0 with three augmentations outperforms AST with SSF, Ia3, OFT, and classifier-only tuning on this task, suggesting PEFT is not always necessary at this data scale.
  • Adding three augmentations to the training set improves most models, with EfficientNet-B0 rising from 94.27% to 95.95%.
  • The custom CNN that won in the authors' prior 9-class study no longer leads at 31 classes, indicating that the best architecture depends on dataset scale.
  • Because validation and inference splits are unaugmented, the reported accuracy is not an artifact of inflated test data.

Reading between the lines

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

  • Because each of the 31 drone classes was recorded at a single site and the recording hardware changed in 2024, the 95.95% figure may partly reflect site acoustics or microphone artifacts rather than drone identity; a cross-location test would reveal how much of the accuracy transfers.
  • A natural extension would be leave-one-site-out cross-validation as a stricter protocol; if accuracy collapses, the paper's central claim is true only for this recording protocol.
  • The near-tie between full fine-tuning and PEFT on EfficientNet-B0 suggests that on slightly larger datasets the PEFT advantage could emerge, making the choice of fine-tuning strategy less important than the base architecture and augmentation policy.
  • The framework could be tested on open UAV audio datasets with multiple recording sites per class to benchmark how much of this result is dataset-specific.
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 evaluates UAV audio classification on a private dataset of 3,100 five-second clips (31 drone classes) using pre-trained CNNs (EfficientNet-B0/B7, ResNet-18/152, MobileNet-V3 small/large), a custom CNN, and the AST transformer, under full fine-tuning and several PEFT regimes (classifier-only, batchnorm, SSF, Ia3, OFT), with and without time-stretch/sine-distortion augmentation. The headline result is 95.95% five-fold cross-validated validation accuracy for full fine-tuning of EfficientNet-B0 with three augmentations per training sample, which the authors interpret as evidence that lightweight CNNs with PEFT and targeted augmentation are effective for small-data UAV audio classification.

Significance. If the empirical claims hold, the paper would provide a useful practical benchmark: a relatively small pre-trained CNN (EfficientNet-B0) outperforming a transformer (AST) and much larger CNNs on a 31-class audio classification task with limited data, while also showing that several PEFT methods approach full fine-tuning performance. The public code repository and W&B logs are commendable strengths, as are the five-year data collection effort and the systematic comparison across many fine-tuning configurations. However, the validity of the central quantitative claim depends on the correctness of the evaluation protocol and the accuracy of the reported numbers, both of which are called into question by the issues below.

major comments (3)
  1. [III.E and IV.C] The split descriptions are internally contradictory. Section III.E states 'Both the training and validation splits have augmented data, whereas the validation and inference splits do not', which is self-contradictory because the validation split is asserted to both have and not have augmented data. Section IV.C then says the five-fold runs use 70% training, 20% 'validation (testing)', and 10% inference, which conflicts with the 60/20/10/10 split described in III.E. If augmented samples, which are derived from training clips via time stretch and sine distortion, are present in the validation split, the reported 95.95% validation accuracy would not measure generalization and would be inflated by training-data leakage. Please state the exact split protocol used for the reported results and confirm that no augmented sample appears in the validation or inference splits.
  2. [Table II] The entries for EfficientNet-B0 with 0 Augs show identical values, 94.27% ± 1.99%, across all four fine-tuning methods (Full, Classifier, Batchnorm, SSF), and the same holds for EfficientNet-B7 with 0 Augs, where all four methods report 93.58% ± 0.50%. Since these methods train different parameter subsets, exact equality of both mean and standard deviation across methods is implausible and indicates a copy or execution error. Because this table is the empirical basis for the paper's conclusions, the authors must verify every entry against the W&B logs and provide a corrected table; the ranking of methods and the headline claim may change once the erroneous entries are fixed.
  3. [Table I and Section III.A] The dataset design confounds drone identity with recording site and hardware. Table I shows that each of the 31 drone classes was recorded at a single collection site, and Section III.A states that the recording hardware changed in 2024 (from an Intel MacBook Air to an M3 MacBook Air). If room acoustics, background noise, or microphone transfer functions differ systematically across sites and sessions, the classifier could be learning site-specific or hardware-specific cues rather than drone identity. The claimed 95.95% validation accuracy would then not generalize to drones recorded under different conditions. Please report site-conditional or hardware-conditional accuracy, or otherwise demonstrate that the model's decisions are driven by drone-specific acoustic signatures rather than by recording environment.
minor comments (5)
  1. [III.A] The text says the dataset contains '3,300 audio files', but Table I and the abstract both total 3,100 files; one of these numbers is a typo.
  2. [III.E] The phrase 'fixed learning rate of 0.001 and reduce it on plateaus by a scaling factor of 0.1' is contradictory; a learning rate that is reduced on plateaus is not fixed. Please clarify the schedule.
  3. [III.G] The phrase 'max average pooling(2D)' is unclear; it should be specified as either max pooling or average pooling.
  4. [III.E] The sentence in Full Fine-tuning ('this is slow and fails and overwrites and has massive forgetting') is grammatically garbled and should be rewritten.
  5. [V] The conclusion states EfficientNet-B0 has 'roughly 4 million parameters', but the standard EfficientNet-B0 has approximately 5.3 million parameters; please verify the number.

Circularity Check

0 steps flagged · score 2.0 of 10

Empirical benchmark paper; the central accuracy claim is a cross-validated measurement rather than a derived quantity, and no step reduces to its inputs by construction.

full rationale

The central claim is that EfficientNet-B0 with full fine-tuning and three raw-audio augmentations achieved 95.95% validation accuracy in 5-fold cross-validation. This is an empirical measurement reported from a training protocol, not a quantity derived from a fitted parameter or from a definition that presupposes the result. There is no equation in the paper that makes the accuracy equal to an input by construction, and no fitted parameter is renamed as a prediction; the cross-validation protocol is the standard setting for such a benchmark claim. The main self-citations are the custom CNN architecture and the augmentation recipe inherited from the authors' prior paper [1]. These are design choices and prior empirical findings, not load-bearing justifications that force the reported accuracy: the current paper independently compares 0-augmentation and 3-augmentation regimens in Table II, and the augmentation scope is additionally anchored to an external survey [23]. The paper does contain internal inconsistencies that bear on the trustworthiness of the empirical claim, but they are not circularity. Section III.E states 'Both the training and validation splits have augmented data, whereas the validation and inference splits do not,' which contradicts itself and, combined with Section IV.C's description of a fixed 10% inference split and 20% validation (testing) split, raises a possible validation-augmentation leakage concern. Additionally, Table II lists identical accuracy and standard deviation values across distinct fine-tuning methods for EfficientNet-B0 '0 Augs' and EfficientNet-B7 '0 Augs', which indicates a reporting error. These are data-integrity and correctness risks, not instances of the paper's conclusions being equivalent to its inputs by definition; therefore the circularity score remains at the minor self-citation level rather than higher.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper is an empirical sweep, not a derivation. Its load-bearing choices are the augmentation recipe inherited from the authors' own prior work, the transfer-learning assumption, and the unexamined recording-site confound. No new entities are introduced.

free parameters (5)
  • Learning rate = 0.001
    Fixed at 0.001 and reduced by factor 0.1 on plateau (Section III.E); not tuned per model in the final cross-validation runs.
  • Effective batch size = 16 (8 with 2 accumulation steps)
    Fixed batch size 8 with 2 accumulation steps (Section III.E).
  • Weight decay (AdamW) = 0.01
    Used only for transformer models (Section III.E).
  • Number of augmentations = 3
    Chosen based on the authors' prior 9-class study [1], not re-benchmarked on the 31-class dataset (Section III.D).
  • Augmentation types = time stretch and sin distortion
    Inherited from prior work [1] as the 'optimal combination' without fresh ablation on this dataset (Section III.D).
assumptions (5)
  • standard math Mel-spectrogram and Fourier transform representation is adequate for UAV audio classification
    Used to convert raw waveforms to model inputs (Section III.C); standard but a modeling choice.
  • domain assumption ImageNet-pretrained CNN features transfer to drone-audio mel-spectrograms despite domain shift
    All pretrained CNNs rely on this; cited in [36] but not tested against random initialization in this paper.
  • domain assumption Each drone class's identity, not recording site or equipment, is the main discriminative signal
    Each of the 31 classes was recorded at a single site and partly with different hardware (Table I, Section III.A); the paper never stratifies or tests unseen locations.
  • domain assumption The 5-fold protocol with a fixed 10% inference split yields an unbiased estimate
    Section IV.C; the inference split is never swapped into training, so reported variance reflects only the 70/20 rotation, not the full data.
  • domain assumption Augmentations applied to raw audio then feature extraction preserve class information
    Section III.D; no verification that augmented clips remain recognizable as their source class.

how reviews work

0 comments
Cite this review

Pith. "Pith review of 15,500 Seconds: Lean UAV Classification Using EfficientNet and Lightweight Fine-Tuning." pith.science (2026). https://pith.science/paper/UVNYLKLU

@misc{pith2026250611049,
  author       = {Pith},
  title        = {Pith review of: 15,500 Seconds: Lean UAV Classification Using EfficientNet and Lightweight Fine-Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UVNYLKLU}},
  note         = {Machine review of arXiv:2506.11049}
}
read the original abstract

As unmanned aerial vehicles (UAVs) become increasingly prevalent in both consumer and defense applications, the need for reliable, modality-specific classification systems grows in urgency. This paper addresses the challenge of data scarcity in UAV audio classification by expanding on prior work through the integration of pre-trained deep learning models, parameter-efficient fine-tuning (PEFT) strategies, and targeted data augmentation techniques. Using a custom dataset of 3,100 UAV audio clips (15,500 seconds) spanning 31 distinct drone types, we evaluate the performance of transformer-based and convolutional neural network (CNN) architectures under various fine-tuning configurations. Experiments were conducted with five-fold cross-validation, assessing accuracy, training efficiency, and robustness. Results show that full fine-tuning of the EfficientNet-B0 model with three augmentations achieved the highest validation accuracy (95.95), outperforming both the custom CNN and transformer-based models like AST. These findings suggest that combining lightweight architectures with PEFT and well-chosen augmentations provides an effective strategy for UAV audio classification on limited datasets. Future work will extend this framework to multimodal UAV classification using visual and radar telemetry.

Figures

Figures reproduced from arXiv: 2506.11049 by the authors.

Figure 1
Figure 1. Audio Analysis of DJI Tello Drone Feature extraction is a staple in Audio Classification [21] and greatly improves the performance of models in audio tasks. Further, there is precedent for using mel-spectrograms as used in previous UAV audio classification [22]. D. Data Augmentations Data augmentations vary in their application and methods. We leverage empirical literature [23] to narrow our scope of relevant UAV au… view at source ↗
Figure 2
Figure 2. AST visualization. Photo taken from [37] [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. SSF Sweeps of Resnet152 C. 5-fold Cross Validation Our final Results and metrics are derived from our 5- fold Cross Validation runs. For our purposes we used both validation and inference splits in the cross validation. Opting for 5-fold cross validation. A fixed 10% of the data is used for inference at the end of a fold’s model training. The other 90% is used in the Training and testing the model’s performance. Whe… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 31 canonical work pages

  1. [1]

    A. P. Berg, M. Y . Wang, Q. Zhang ”4,500 Seconds: Small data training approaches for deep UA V audio classification,” Proceedings 14th Int. Conf. Data Sci., Technol. Appl. (DATA), 2025. [Online] https://www. insticc.org/Primoris/Resources/PaperPdf.ashx?idPaper=ScI3f83+ah4=

    work page 2025

  2. [2]

    Zaman, M

    K. Zaman, M. Sah, C. Direkoglu, and M. Unoki, ”A survey of audio classification using deep learning,” IEEE Access, vol. 11, pp. 106620–106649, 2023, doi: 10.1109/ACCESS.2023.3318015

    arXiv 2023

  3. [3]

    A. F. R. Nogueira, H. S. Oliveira, J. J. M. Machado, and J. M. R. S. Tavares, ”Transformers for urban sound classification—A comprehen- sive performance evaluation,” *Sensors*, vol. 22, no. 22, p. 8874, 2022, doi: 10.3390/s22228874

    work page doi:10.3390/s22228874 2022

  4. [4]

    Berg, *UA V Classification* [Computer software], Available: https: //github.com/AndrewPBerg/UA VClassification

    A. Berg, *UA V Classification* [Computer software], Available: https: //github.com/AndrewPBerg/UA VClassification

    work page

  5. [5]

    Berg, Weights & Biases Project Logs [Online] Available: https:// wandb.ai/andberg9-self/projects

    A. Berg, Weights & Biases Project Logs [Online] Available: https:// wandb.ai/andberg9-self/projects

    work page

  6. [6]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. K ¨opf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, ”PyTorch: An imperative style, high- performance deep learning library,” arXiv preprint arXiv:1912.01703, Dec. 2019....

    arXiv 1912

  7. [7]

    [Online]

    PyTorch Foundation, PyTorch Documentation, 2025. [Online]. Avail- able: https://docs.pytorch.org/docs/stable/index.html

    work page 2025

  8. [8]

    Available: https://huggingface.co/docs/transformers/en/index

    Hugging Face, Transformers Documentation, 2025 [Online]. Available: https://huggingface.co/docs/transformers/en/index

    work page 2025

Show all 52 references
  1. [9]

    Available: https: //huggingface.co/docs/peft/en/index

    Hugging Face, PEFT Documentation, 2025 [Online]. Available: https: //huggingface.co/docs/peft/en/index

    2025

  2. [10]

    Avail- able: https://lightning.ai/docs/pytorch/stable/

    Lightning AI, Pytroch Lightning Documentation, 2025 [Online]. Avail- able: https://lightning.ai/docs/pytorch/stable/

    2025

  3. [11]

    Weights & Biases Documentation, 2025 [Online]

    Weights & Biases. Weights & Biases Documentation, 2025 [Online]. Available: https://docs.wandb.ai/

    2025

  4. [12]

    Docker Documentation, 2025 [Online]

    Docker. Docker Documentation, 2025 [Online]. Available: https://docs. docker.com/

    2025

  5. [13]

    [Online]

    PyTorch Foundation, TorchAudio Documentation, 2025. [Online]. Avail- able: https://docs.pytorch.org/audio/stable/index.html

    2025

  6. [14]

    [Online]

    Audiomentations Documentation, 2025. [Online]. Available: https:// iver56.github.io/audiomentations/

    2025

  7. [15]

    [Online]

    TorchMetrics Documentation, 2025. [Online]. Available: https:// lightning.ai/docs/torchmetrics/stable/

    2025

  8. [16]

    [Online]

    Matplotlib Documentation, 2025. [Online]. Available: https://matplotlib. org/stable/index.html

    2025

  9. [17]

    [Online]

    scikit-learn Documentation, 2025. [Online]. Available: https://scikit-learn.org/stable/

    2025

  10. [18]

    [Online]

    NumPy Documentation, 2025. [Online]. Available: https://numpy.org/ doc/

    2025

  11. [19]

    [Online]

    Librosa Documentation, 2025. [Online]. Available: https://librosa.org/ doc/latest/index.html

    2025

  12. [20]

    S. S. Stevens, J. V olkmann, and E. B. Newman, ”A scale for the measurement of the psychological magnitude pitch,” *J. Acoust. Soc. Am.*, vol. 8, no. 3, pp. 185–190, 1937. [Online]. Available: https://pubs.aip.org/asa/jasa/article/8/3/185/673917/ A-Scale-for-the-Measurement-of...

    1937

  13. [21]

    Wolf-Monheim, ”Spectral and rhythm features for audio classi- fication with deep convolutional neural networks,” *arXiv preprint arXiv:2410.06927*, Oct

    F. Wolf-Monheim, ”Spectral and rhythm features for audio classi- fication with deep convolutional neural networks,” *arXiv preprint arXiv:2410.06927*, Oct. 2024. [Online]. Available: https://arxiv.org/abs/ 2410.06927

    2024 arXiv

  14. [22]

    Y . Wang, Z. Chu, I. Ku, E. C. Smith, and E. T. Matson, ”A large-scale UA V audio dataset and audio-based UA V classification using CNN,” in *Proc. 2022 Sixth IEEE Int. Conf. Robotic Comput. (IRC)*, Naples, Italy, Dec. 2022, pp. 186–189, doi: 10.1109/IRC55401.2022.00039

    2022

  15. [23]

    K ¨ummritz, ”The sound of surveillance: Enhancing machine learning- driven drone detection with advanced acoustic augmentation,” *Drones*, vol

    S. K ¨ummritz, ”The sound of surveillance: Enhancing machine learning- driven drone detection with advanced acoustic augmentation,” *Drones*, vol. 8, no. 3, p. 105, Mar. 2024. [Online]. Available: https://doi.org/10. 3390/drones8030105

    2024

  16. [24]

    A. P. Berg. Augmentations Colab Notebook,

  17. [25]

    M. Iman, H. R. Arabnia, and K. Rasheed, ”A review of deep transfer learning and recent advancements,” *Technologies*, vol. 11, no. 2, p. 40, Feb. 2023. [Online]. Available: https://doi.org/10.3390/ technologies11020040

    2023

  18. [26]

    A. Mao, M. Mohri, and Y . Zhong, ”Cross-entropy loss functions: Theoretical analysis and applications,” arXiv preprint arXiv:2304.07288, Apr. 2023. [Online]. Available: https://arxiv.org/abs/2304.07288

    2023 arXiv

  19. [27]

    D. P. Kingma and J. Ba, ”Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, Dec. 2014. [Online]. Available: https: //arxiv.org/abs/1412.6980

    2014 arXiv

  20. [28]

    Loshchilov and F

    I. Loshchilov and F. Hutter, ”Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, Nov. 2017. [Online]. https://arxiv.org/ abs/1711.05101

    2017 arXiv

  21. [29]

    Alain and Y

    G. Alain and Y . Bengio, ”Understanding intermediate layers using linear classifier probes,” *arXiv preprint arXiv:1610.01644*, Nov. 2018. [Online]. Available: https://arxiv.org/abs/1610.01644

    2018 arXiv

  22. [30]

    Z. Qiu, W. Liu, H. Feng, Y . Xue, Y . Feng, Z. Liu, D. Zhang, A. Weller, and B. Sch ¨olkopf, ”Controlling text-to-image diffusion by orthogo- nal finetuning,” arXiv preprint arXiv:2306.07280, Jun. 2023. [Online]. Available: https://arxiv.org/abs/2306.07280

    2023 arXiv

  23. [31]

    H. Liu, D. Tam, M. Muqeeth, J. Mohta, T. Huang, M. Bansal, and C. Raffel, ”Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning,” arXiv preprint arXiv:2205.05638, May 2022. [Online]. Available: https://arxiv.org/abs/2205.05638

    2022 arXiv

  24. [32]

    D. Lian, D. Zhou, J. Feng, and X. Wang, ”Scaling & shifting your features: A new baseline for efficient model tuning,” arXiv preprint arXiv:2210.08823, Oct. 2022. [Online]. Available: https://arxiv.org/abs/ 2210.08823

    2022 arXiv

  25. [33]

    A. P. Berg. UA V Classification SSF PEFT code, 2025 [Online]https://github.com/AndrewPBerg/UA VClassification/blob/ master/src/models/ssf adapter.py

    2025

  26. [34]

    Frankle, D

    J. Frankle, D. J. Schwab, and A. S. Morcos, ”Training BatchNorm and only BatchNorm: On the expressive power of random features in CNNs,” arXiv preprint arXiv:2003.00152, Mar. 2020. [Online]. Available:https: //arxiv.org/abs/2003.00152

    2003 arXiv

  27. [36]

    Palanisamy, D

    K. Palanisamy, D. Singhania, and A. Yao, ”Rethinking CNN models for audio classification,” arXiv preprint arXiv:2007.11154, Jul. 2020. [Online]. Available: https://arxiv.org/abs/2007.11154

    2007 arXiv

  28. [37]

    Gong, Y .-A

    Y . Gong, Y .-A. Chung, and J. Glass, ”AST: Audio Spectrogram Trans- former,” arXiv preprint arXiv:2104.01778, Apr. 2021. [Online]. Avail- able: https://arxiv.org/abs/2104.01778

    2021 arXiv

  29. [38]

    Russakovsky et al., ”ImageNet large scale visual recognition chal- lenge,” arXiv preprint arXiv:1409.0575, Sep

    O. Russakovsky et al., ”ImageNet large scale visual recognition chal- lenge,” arXiv preprint arXiv:1409.0575, Sep. 2014. [Online]. Available: https://arxiv.org/abs/1409.0575

    2014 arXiv

  30. [39]

    J. F. Gemmeke et al., ”Audio Set: An ontology and human-labeled dataset for audio events,” in *Proc. 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)*, New Orleans, LA, USA, 2017, pp. 776–780, doi: 10.1109/ICASSP.2017.7952261. [Online]. Av...

    2017

  31. [40]

    K. He, X. Zhang, S. Ren, and J. Sun, ”Deep residual learning for image recognition,” arXiv preprint arXiv:1512.03385, Dec. 2015. [Online]. Available: https://arxiv.org/abs/1512.03385

    2015 arXiv

  32. [41]

    Available: https://pytorch.org/vision/main/models/resnet.html

    TorchVision, ”ResNet — TorchVision main documentation,” [Online]. Available: https://pytorch.org/vision/main/models/resnet.html

  33. [42]

    A. G. Howard et al., ”MobileNets: Efficient convolutional neural net- works for mobile vision applications,” arXiv preprint arXiv:1704.04861, Apr. 2017. [Online]. Available: https://arxiv.org/abs/1704.04861

    2017 arXiv

  34. [43]

    Howard et al., ”Searching for MobileNetV3,” arXiv preprint arXiv:1905.02244, May 2019

    A. Howard et al., ”Searching for MobileNetV3,” arXiv preprint arXiv:1905.02244, May 2019. [Online]. Available: https://arxiv.org/abs/ 1905.02244

    1905 arXiv

  35. [44]

    Available: https://pytorch.org/vision/main/models/mobilenetv3

    TorchVision, ”MobileNet V3 — TorchVision main documentation,” [On- line]. Available: https://pytorch.org/vision/main/models/mobilenetv3. html

  36. [46]

    Available: https://pytorch.org/vision/main/models/efficientnet.html

    TorchVision, ”EfficientNet — TorchVision main documentation,” [On- line]. Available: https://pytorch.org/vision/main/models/efficientnet.html

  37. [47]

    Semenyuk, I

    V . Semenyuk, I. Kurmashev, A. Lupidi, D. Alyoshin, L. Kurma- sheva, and A. Cantelli-Forti, ”Advance and Refinement: The Evolution of UA V Detection and Classification Technologies,” arXiv preprint arXiv:2409.05985, Sep. 2024. [Online]. Available: https://arxiv.org/abs/ 2409.05985

    2024 arXiv

  38. [48]

    Md. M. Rahman, S. Siddique, M. Kamal, R. H. Rifat, and K. D. Gupta, ”UA V (Unmanned Aerial Vehicles): Diverse Applications of UA V Datasets in Segmentation, Classification, Detection, and Track- ing,” arXiv preprint arXiv:2409.03245, Sep. 2024. [Online]. Available: https://arx...

    2024 arXiv

  39. [49]

    Lialin, V

    V . Lialin, V . Deshpande, X. Yao, and A. Rumshisky, ”Scaling Down to Scale Up: A Guide to Parameter-Efficient Fine-Tuning,” arXiv preprint arXiv:2303.15647, Nov. 2024. [Online]. Available: https://arxiv.org/abs/ 2303.15647

    2024 arXiv

  40. [50]

    M. Y . Wang, D. C. Ramirez, E. Noonan, M. Linn, and Q. Zhang, ”A Comprehensive Dataset and Visualization Tool for Drone Acoustic Signatures,” in *Proc. 2024 Artificial Intelligence x Humanities, Education, and Art (AIxHEART)*, Laguna Hills, CA, USA, Sep. 2024, pp. 13–17, doi: ...

    2024

  41. [51]

    M. Y . Wang, Z. Chu, I. Ku, E. C. Smith, and E. T. Matson, ”A 15-Category Audio Dataset for Drones and an Audio-Based UA V Classification Using Machine Learning,” *International Journal of Semantic Computing*, vol. 18, no. 2, pp. 257–272, 2024, doi: 10.1142/S1793351X24300048. ...

    2024 doi

  42. [2019]

    Available: https://arxiv.org/abs/1905.11946

    [Online]. Available: https://arxiv.org/abs/1905.11946

    1905 arXiv

  43. [2020]

    Available: https://arxiv.org/abs/2010.11929

    [Online]. Available: https://arxiv.org/abs/2010.11929

    2010 arXiv

  44. [2025]

    [Online]https://colab.research.google.com/drive/ 1bl4RTQd7ENnMYEc4thwBwtocF-q1NYp2?usp=sharing]

Pith tools

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