Pith. sign in

REVIEW 4 major objections 4 minor 4 cited by

A nested cross-validation protocol plus an attention CNN cuts whale-call false positives roughly tenfold on unseen Antarctic sites.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

ARPA-N with blocked nested cross-validation is presented as a robust whale-call detector, but the abstract's key 10x false-positive claim is absent from the reported experiments and several stated results contradict the tables.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The abstract promises a 10x false-positive reduction that the experimental section never actually reports; the underlying blocked nested-CV protocol is a reasonable idea, but the paper as written overstates its results. the 4 major comments →

arxiv 2509.04682 v2 pith:WGMPIM5H submitted 2025-09-04 cs.SD cs.AIcs.CVcs.IRcs.LGeess.AS

GetNetUPAM: Ecologically Informed Nested Cross-Validation and Noise-Robust Attention for Marine Bioacoustic Monitoring

Nicholas R. Rasmussen , Rodrigue Rizk , Longwei Wang , KC Santosh This is my paper
classification cs.SD cs.AIcs.CVcs.IRcs.LGeess.AS
keywords marine bioacousticsunderwater passive acoustic monitoringnested cross-validationsite-year blockingblue whale D-call detectionspatial attentionspectrogram classificationmodel stability
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper argues that standard random-split evaluations hide how acoustic detectors fail when deployed on new marine sites. It introduces GetNetUPAM, a two-level cross-validation scheme that holds out whole site-year blocks and uses the inner folds to measure how much a model's scores wobble, rather than to tune hyperparameters. On top of this protocol it builds ARPA-N, a lightweight CNN that uses adaptive pooling and spatial attention to handle irregular spectrogram sizes and suppress noise. The paper reports that ARPA-N under GetNetUPAM beats DenseNet baselines by 14.4% average precision and, on the zero-training Balleny Islands site, reduces false positives per hour by about an order of magnitude at fixed 90% recall. If true, this is a path to detectors whose advertised accuracy survives contact with new environments.

Core claim

The central claim is that ecological structure—site and year—must drive evaluation, and that a stability profile across held-out site-years is more informative than a single accuracy number. GetNetUPAM splits the Antarctic library into site-year blocks; each block is the outer test set, while five inner stratified folds are trained and tested on the remaining data, producing mean and standard deviation for every metric. ARPA-N, the companion architecture, uses CBAM-style spatial attention to learn noise suppression and adaptive resolution pooling to standardize odd-sized spectrograms. Together they are claimed to give consistent detection across environmental regimes and a 14.4% average-prec

What carries the argument

The two load-bearing pieces are GetNetUPAM and ARPA-N. GetNetUPAM is a hierarchical nested cross-validation scheme (Algorithm 1): outer folds are entire site-year recordings, so each test set is an unseen acoustic regime; inner five-fold stratified cross-validation trains several models per outer fold and reports the mean and standard deviation of test metrics, turning fold-to-fold variance into the object of study. ARPA-N is a CNN that adds CBAM spatial attention, which pools feature maps along the channel dimension and applies a 7x7 convolution to produce a spatial mask that suppresses noise, plus adaptive resolution pooling that resizes arbitrary-height/width spectrograms to a fixed 64x64

Load-bearing premise

The load-bearing premise is that three held-out site-years, including just one zero-training site, are enough to estimate how a detector will perform on a new deployment; with such a small outer loop, the reported stability differences could be produced by chance or by per-site annotation noise rather than by environmental regime.

What would settle it

Re-run the same protocol with site labels shuffled (or with a larger set of held-out site-years) and compare ARPA-N versus DenseNet on per-hour false positives at 90% recall; if the roughly 10x reduction does not persist or the variance across outer folds overlaps, the central generalization claim is refuted.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Deployments on a new site can expect roughly 10x fewer false positives per hour at the same 90% recall than a DenseNet-style detector, if the Balleny result generalizes.
  • Reporting mean and standard deviation across site-year folds becomes a way to compare detectors on stability, not just average score.
  • ARPA-N's saliency maps localize calls rather than scattered noise, which enables automated event-to-timestamp mapping for downstream analysis.
  • With only 4.97M parameters and sub-3ms per-sample inference, the detector is plausible for buoy-mounted, resource-limited monitoring.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the same site-year blocked nested design transfers to terrestrial bioacoustics or other rare-event monitoring where spatial and temporal autocorrelation inflate random-split scores.
  • Editorial inference: the false-positive-per-hour claim is stated only in the abstract; a direct table of FP/hour at fixed recall per outer fold would make the operational benefit easy to verify.
  • Editorial inference: with three outer folds, fold-level 'environmental regime' differences are confounded with site-specific annotation style; a larger multi-site benchmark, or a permutation test over site labels, would sharpen the stability claim.
  • Editorial inference: the 14.4% AP gain is reported relative to DenseNet baselines; comparing ARPA-N against other attention-equipped lightweight models under the same protocol would isolate the contribution of CBAM attention from adaptive pooling.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The manuscript introduces GetNetUPAM, a hierarchical nested cross-validation protocol for underwater passive acoustic monitoring (UPAM) in which site-year blocks form outer folds and five-fold stratified cross-validation forms inner folds; the dispersion of inner-loop test metrics is used as a stability measure. The authors also propose ARPA-N, a lightweight CNN combining adaptive resolution pooling with CBAM-style spatial attention, and evaluate it against several baselines on three Antarctic blue-whale D-call site-years (Kerguelen 2015, Casey 2017, Balleny Islands 2015). The abstract claims a 10x reduction in false positives per hour at 90% recall on the zero-support Balleny site, a log2-scale order-of-magnitude drop in variability, and a 14.4% average-precision gain over DenseNet baselines.

Significance. GetNetUPAM addresses a genuine problem: standard random cross-validation can leak site/year structure and overstate generalization in bioacoustic monitoring. If the advertised empirical evidence were provided, the framework plus ARPA-N's lower variance on the zero-support Balleny block would be a useful contribution. The paper also includes an honest disclosure of AI assistance. However, the main quantitative claims in the abstract are either absent from the reported experiments or directly contradicted by Table I. The significance of the work cannot currently be assessed as stated; the central contribution is an evaluation protocol with a promising but unverified performance claim.

major comments (4)
  1. [Abstract & §IV-F, Table I] The abstract's headline claim—'reduces false positives per hour by over an order of magnitude (approximately 10x) at fixed 90 percent recall' on Balleny—is not supported by the experimental section. No FP/hour metric, no decision threshold, and no 90%-recall operating point appear in §IV-F, Table I, or Fig. 5. Moreover, 'consistently improved metrics across folds' is contradicted by Table I: on Balleny, ARPA-N has F1=0.474 (σ=0.023) versus 60sDense F1=0.495 (σ=0.052), and ARPA-N's precision is lower (0.357 vs 0.389). This is the paper's central advertised result and must be either fully reported with operating-point details or withdrawn.
  2. [Abstract & §I vs §IV-F] The abstract promises a 'log2-scale order-of-magnitude drop in variability across all metrics,' but §I states 'up to a three-fold reduction in variability (log 2),' and Table I shows a maximum σ reduction of about 2.3x (Balleny F1: 0.023 vs 0.052). Three-fold is not an order of magnitude. The '14.4% gain in average precision' is also not tied to a specific comparison in the text; from Table I, ARPA-N's AP gains over 60sDense are 8.3% (Kerguelen), 18.7% (Casey), and 23.1% (Balleny), while macro AP is 0.656 vs 0.572 (+14.7%). The abstract should cite the exact comparison and avoid 'order-of-magnitude' unless the 10x FP/hour result is actually computed.
  3. [Algorithm 1 / §III-A, §IV-A] The outer loop has only three folds (three site-years), and only one of them (Balleny) is truly zero-support. With K=3, the stability estimates have very low statistical power; no confidence intervals or significance tests are provided. The claim that GetNetUPAM 'quantifies deployment-relevant stability' is therefore not established. The authors should report fold-wise results with uncertainty, discuss the limited number of environmental regimes, and soften conclusions that generalize beyond the three available sites.
  4. [Acknowledgments] The acknowledgment states that AI assistance included 'contextual extrapolation in the results section.' Given that the headline FP/hour reduction is missing from the results and the abstract conflicts with the body, the authors must clarify which numbers in the abstract and §IV-F were experimentally measured and which were extrapolated. This is a matter of scientific record and directly relevant to assessing the reliability of the central claims.
minor comments (4)
  1. [Fig. 5] The PR curves are not annotated with the 90% recall operating point or the dataset duration needed to convert precision to FP/hour. Adding these annotations would make the abstract's claim checkable.
  2. [Table I] Table I has inconsistent decimal formatting and several entries appear run together, especially in the Balleny and efficiency columns. A cleaner table with aligned columns and consistent significant digits would aid readability.
  3. [§IV-A] The exclusion of Elephant Islands 2013 and inclusion of Elephant Islands 2014 in training are described as choices that 'improved results.' This is not a sufficient justification; the authors should show robustness to this data-selection decision or at least discuss its potential impact on the reported gains.
  4. [§V / reproducibility] The paper claims a reproducible benchmark but gives no code repository, data-processing scripts, or trained-model checkpoints. The public dataset [4] is cited, but exact preprocessing, hyperparameter choices, and threshold selection are not fully specified, making exact replication difficult.

Circularity Check

0 steps flagged

No circular derivation: empirical benchmark against external baselines; the unsupported FP/hour headline is a reporting/evidence gap, not a circular step.

full rationale

The paper's contribution is empirical rather than derivational. GetNetUPAM is a cross-validation protocol (Algorithm 1, Sec. III-A) and ARPA-N is a convolutional architecture (Sec. III-D). The quantitative claims in Table I and Fig. 5 are obtained by training on site-year folds and comparing against external baselines (DenseNet [10], ResNet [13], ViT, SurfPerch [36]); none of these baselines' reported numbers are reused as predictions. The only self-citation, [13], is used to justify the 65.536-s window length and as one baseline, but that baseline's performance is re-measured in this paper rather than imported, so it does not make the comparison circular. The saliency/attention analysis (Sec. III-E-a, Sec. IV-H) is post-hoc interpretation, not a fitted quantity renamed as a prediction. The abstract's '10x reduction in FP/hour at 90% recall' cannot be reconstructed from the reported tables or PR curves (no threshold or FP/hour metric is given), and the acknowledgment's 'contextual extrapolation in the results section' is an admitted transparency limitation; these are evidence/reporting defects, not circular reductions. Therefore no circular step is exhibited, and the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper's claims rest on data and design choices rather than derived constants. The most consequential are the post-hoc exclusion of Elephant Islands 2013, the unstated hyperparameters, and the small number of site-years; these would need to be pinned down before the benchmark results can be reproduced or trusted.

free parameters (5)
  • Inner fold count k = 5
    Algorithm 1 fixes k=5 without ablation; the reported sigma values depend directly on this choice, so the 'stability' measurement is not robust to k.
  • Elephant Islands 2013 exclusion = excluded from all experiments
    Section IV-A: the dataset was removed permanently after observing that 'removing it improved results on other datasets.' This is a post-hoc data-selection choice that can bias the benchmark.
  • Negative downsampling factor = 0.5
    Section IV-C: negatives are downsampled by half to control imbalance; the procedure and seed are unspecified, so the exact training set per fold is not reproducible.
  • Adjacent-positive concatenation = up to 3 (60s variant)
    Section IV-B: 'we concatenate up to three adjacent positives for the 60-second variant and unlimited positives for the 4-second variant' is an ad hoc augmentation rule that changes the input distribution and is not justified by prior literature.
  • Additive Gaussian noise variance sigma^2 = unspecified
    Section III-D-a uses O(0)=Pnorm+|N(0,sigma^2)|; the value of sigma^2 is never given, and the absolute-value transform is not explained, making the exact model input ambiguous.
axioms (5)
  • domain assumption Each site-year is a distinct environmental regime, so leave-one-site-year-out tests deployment to unseen regimes.
    Section III-A; with only three held-out site-years, this assumption is thin but underlies the entire generalization claim.
  • domain assumption The Antarctic annotated library annotations are correct for Kerguelen 2015, Casey 2017, and Balleny 2015.
    Section IV-A; the paper relies on these labels as ground truth without re-validating inter-annotator agreement, after disqualifying Elephant 2013 on such grounds.
  • domain assumption The mean and sigma of the five inner-fold test evaluations estimate deployment stability.
    Section III-A/Algorithm 1; the inner folds are not independent because they share the same training pool, so the reported sigma likely understates true model variance.
  • domain assumption Inner-validation binary accuracy is a valid model-selection criterion.
    Section IV-C; best weights are chosen by highest binary accuracy on the validation subset, a common but not universally reliable heuristic for rare-event detection.
  • domain assumption The STFT and windowing parameters preserve the information needed for D-call detection.
    Sections III-B/C; L=256, b=64, 65.536s windows are assumed sufficient for 40-50s calls at 250 Hz.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of GetNetUPAM: Ecologically Informed Nested Cross-Validation and Noise-Robust Attention for Marine Bioacoustic Monitoring." pith.science (2026). https://pith.science/paper/WGMPIM5H

@misc{pith2026250904682,
  author       = {Pith},
  title        = {Pith review of: GetNetUPAM: Ecologically Informed Nested Cross-Validation and Noise-Robust Attention for Marine Bioacoustic Monitoring},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WGMPIM5H}},
  note         = {Machine review of arXiv:2509.04682}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Deploying reliable bioacoustic monitoring systems requires models that generalize under high-noise, low-SNR conditions and evaluation protocols that expose deployment-relevant failure modes, gaps largely unaddressed in current UPAM practice. Intrinsic noise, variable propagation, and mixed biological and anthropogenic sources induce distribution shifts that conventional models and single-split evaluations obscure, inflating performance and masking instability. We introduce GetNetUPAM, a hierarchical nested cross-validation framework that uses the nested stage to quantify model stability rather than tune for inflated hold-out scores. By partitioning data into site-year blocks, GetNetUPAM preserves ecological heterogeneity and forces each outer fold to represent a distinct environmental regime, preventing overfitting to localized noise or sensor artifacts. Inner stratified folds measure generalization across the full UPAM signal distribution, enforcing strict separation between model development and the outer held-out deployment condition. Using GetNetUPAM, we evaluate the Adaptive Resolution Pooling and Attention Network (ARPA-N), a CNN architecture for irregular spectrogram dimensions. ARPA-N integrates CBAM spatial attention as a learned noise suppressor, producing attention maps that localize true call structure and avoid the global, non-biological cues exploited by standard CNNs on long-window data. Under GetNetUPAM, ARPA-N generalizes robustly across diverse environmental regimes. In the zero-training support Balleny Islands region, it reduces false positives per hour by over an order of magnitude (approximately 10x) at fixed 90 percent recall, yielding consistently improved metrics across folds. These advances provide a reproducible benchmark and move UPAM toward scalable, deployment-reliable ecological monitoring.

Figures

Figures reproduced from arXiv: 2509.04682 by KC Santosh, Longwei Wang, Nicholas R. Rasmussen, Rodrigue Rizk.

Figure 1
Figure 1. Figure 1: Motivation for GetNetUPAM and ARPA-N. (Left) Input spectrograms in UPAM have diverse spectral and temporal resolutions, with colored overlays denoting target whale signals and intrinsic noise sources. Such heterogeneous, odd-dimensional inputs challenge model stability and can conceal performance variance behind a single evaluation score. (Center) The ARPA-N convolutional neural network applies spatial att… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the hierarchical nested cross-validation architecture, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Integrated preprocessing and ARPA-N detection pipeline. Stage 1: raw audio is segmented into overlapping windows and transformed [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: CBAM Spatial Attention Module: Input feature maps are pooled [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Precision Recall Curve. Visualizing the performance of the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparative Analysis of Acoustic Event Detection — The figure illustrates saliency map overlays on human-interpretable spectrograms [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Robust and Clinically Reliable EEG Biomarkers: A Cross Population Framework for Generalizable Parkinson's Disease Detection

    cs.LG 2026-04 conditional novelty 7.0

    A cross-population framework for EEG Parkinson's detection using exhaustive 75 directional evaluations and nested validation shows asymmetric transfer and accuracy up to 94.1% when training diversity increases, suppor...

  2. Mechanistic Interpretability of LLM Jailbreaks via Internal Attribution Graphs

    cs.CR 2026-07 conditional novelty 6.0

    On Llama-2-7B, path-rerouting magnitude in paired transcoder attribution graphs correlates with jailbreak success (r=0.461), while static node metrics and top-feature ablations do not.

  3. Explainable Novel Category Discovery in Semantic Concept Space

    cs.CV 2026-07 conditional novelty 6.0

    xNCD routes novel category discovery through a CLIP-aligned concept bottleneck, matching strong NCD baselines while producing intrinsic cluster- and instance-level concept explanations.

  4. Learning to Transmit: Volatility-Aware Predictive Communication for Energy-Efficient IoT Networks

    cs.IT 2026-07 conditional novelty 5.0

    Sensors using volatility-aware studentized residuals plus RLS online adaptation transmit up to 94.7% less IoT data while keeping reconstruction MAE at 0.35°C.

Reference graph

Works this paper leans on

41 extracted references · 37 canonical work pages · cited by 4 Pith papers

  1. [1]

    Identifying the world’s most climate change vulnerable species: A systematic trait-based assessment of all birds, amphibians and corals,

    W. B. Foden, S. H. M. Butchart, S. N. Stuart, J.-C. Vi ´e, H. R. Akc ¸akaya, A. Angulo, L. M. DeVantier, A. Gutsche, E. Turak, L. Caoet al., “Identifying the world’s most climate change vulnerable species: A systematic trait-based assessment of all birds, amphibians and corals,”PLOS ONE, vol. 8, no. 6, p. e0065427, 2013. [Online]. Available: https://journ...

    work page doi:10.1371/journal.pone.0065427 2013

  2. [2]

    A vulnerability assessment of fish and invertebrates to climate change on the northeast u.s. continental shelf,

    J. A. Hare, W. E. Morrison, M. W. Nelson, M. M. Stachura, E. J. Teeters, R. B. Griffiset al., “A vulnerability assessment of fish and invertebrates to climate change on the northeast u.s. continental shelf,”PLOS ONE, vol. 11, no. 2, p. e0146756, 2016. [Online]. Available: https://journals.plos.org/plosone/article?id=10.1371/ journal.pone.0146756

    work page 2016

  3. [3]

    Shaped by their environment: 11 Variation in blue whale morphology across three productive coastal ecosystems,

    D. R. Barlow, K. C. Bierlich, W. K. Oestreich, G. Chiang, J. W. Durban, J. A. Goldbogen, D. W. Johnston, M. S. Leslie, M. J. Moore, J. P. Ryan, and L. G. Torres, “Shaped by their environment: 11 Variation in blue whale morphology across three productive coastal ecosystems,”Integrative Organismal Biology, vol. 5, no. 1,

    work page

  4. [4]

    An open access dataset for developing automated detectors of antarctic baleen whale sounds and performance evaluation of two commonly used detectors,

    B. S. Miller, B. S. Miller, K. M. Stafford, I. Van Opzeeland, D. Harris, F. Samaran, A. ˇSirovi´c, S. Buchan, K. Findlay, N. Balcazar, and et al., “An open access dataset for developing automated detectors of antarctic baleen whale sounds and performance evaluation of two commonly used detectors,”Scientific Reports, vol. 11, no. 1, 2021

    work page 2021

  5. [5]

    Seeing biodiversity: perspectives in machine learning for wildlife conservation,

    D. Tuia, B. Kellenberger, S. Beery, B. R. Costelloe, S. Zuffi, B. Risse, A. Mathis, M. W. Mathis, F. van Langevelde, T. Burghardtet al., “Seeing biodiversity: perspectives in machine learning for wildlife conservation,”Nature Communications, vol. 12, no. 1, pp. 1–13, 2021

    work page 2021

  6. [6]

    Ecoacoustic indices in marine ecosystems: a review on recent developments, challenges, and future directions,

    M. Minello, L. Calado, and F. C. Xavier, “Ecoacoustic indices in marine ecosystems: a review on recent developments, challenges, and future directions,”ICES Journal of Marine Science, vol. 78, no. 9, pp. 3066–3074, November 2021

    work page 2021

  7. [7]

    Calls reveal population structure of blue whales across the southeast indian ocean and the southwest pacific ocean,

    N. E. Balcazar, J. S. Tripovich, H. Klinck, S. L. Nieukirk, D. K. Mellinger, R. P. Dziak, and T. L. Rogers, “Calls reveal population structure of blue whales across the southeast indian ocean and the southwest pacific ocean,”Journal of Mammalogy, vol. 96, no. 6, pp. 1184–1193, 2015. [Online]. Available: https://doi.org/10.1093/jmammal/gyv126

    work page doi:10.1093/jmammal/gyv126 2015

  8. [8]

    Sounding the call for a global library of underwater biological sounds,

    M. J. G. Parsons, T.-H. Lin, T. A. Mooney, C. Erbe, F. Juanes, M. Lammers, S. Li, S. Linke, A. Looby, S. L. Nedelec, I. Van Opzee- land, C. Radford, A. N. Rice, L. Sayigh, J. Stanley, E. Urban, and L. Di Iorio, “Sounding the call for a global library of underwater biological sounds,”Frontiers in Ecology and Evolution, vol. 10, pp. 1–15, February 2022

    work page 2022

  9. [9]

    The past, present, and future of underwater passive acoustic monitoring,

    J. Hildebrand, S. Wiggins, S. Baumann-Pickering, K. Frasier, and M. A. Roch, “The past, present, and future of underwater passive acoustic monitoring,”The Journal of the Acoustical Society of America, vol. 155, no. Supplement 3, p. A96, 2024. [Online]. Available: https://doi.org/10.1121/10.0026934

    work page doi:10.1121/10.0026934 2024

  10. [10]

    Deep learning algorithm outperforms experienced human observer at detection of blue whale d-calls: A double-observer analysis,

    B. S. Miller, S. Madhusudhana, M. G. Aulich, and N. Kelly, “Deep learning algorithm outperforms experienced human observer at detection of blue whale d-calls: A double-observer analysis,” Remote Sensing in Ecology and Conservation, vol. 9, no. 1, p. 104–116, 2022

    work page 2022

  11. [11]

    Rapid assessment of vessel noise events and quiet periods in glacier bay national park and preserve using a convolutional neural net,

    S. M. Haver, K. B. Gustafson, and C. M. Gabriele, “Rapid assessment of vessel noise events and quiet periods in glacier bay national park and preserve using a convolutional neural net,” Environmental Data Science, vol. 2, p. e14, 2023

    work page 2023

  12. [12]

    Deep learning in marine bioacoustics: a benchmark for baleen whale detection,

    E. Schall, I. I. Kaya, E. Debusschere, P. Devos, and C. Parcerisas, “Deep learning in marine bioacoustics: a benchmark for baleen whale detection,”Remote Sensing in Ecology and Conservation, vol. 10, no. 2, pp. 131–143, 2024

    work page 2024

  13. [13]

    Deepwhalenet: Climate change-aware fft-based deep neural network for passive acoustic monitoring,

    N. Rasmussen, R. Rizk, O. Matoo, and K. Santosh, “Deepwhalenet: Climate change-aware fft-based deep neural network for passive acoustic monitoring,”International Journal of Pattern Recognition and Artificial Intelligence, vol. 38, no. 14, p. 2459014, 2024. [Online]. Available: https://doi.org/10.1142/S0218001424590146

    work page doi:10.1142/s0218001424590146 2024

  14. [14]

    Cross-validation strategies for data with tem- poral, spatial, hierarchical, or phylogenetic structure,

    D. R. Roberts, V. Bahn, S. Ciuti, M. S. Boyce, J. Elith, G. Guillera- Arroita, S. Hauenstein, J. J. Lahoz-Monfort, B. Schr¨oder, W. Thuiller, and D. I. Warton, “Cross-validation strategies for data with tem- poral, spatial, hierarchical, or phylogenetic structure,”Ecography, vol. 40, no. 8, pp. 913–929, 2017

    work page 2017

  15. [15]

    Attention based convolutional neural network with multi- frequency resolution feature for environment sound classification,

    Y. Zhang, Y. Zhang, Y. Zhang, Y. Wang, Y. Wang, and Z. Wang, “Attention based convolutional neural network with multi- frequency resolution feature for environment sound classification,” Computational Intelligence and Neuroscience, vol. 2022, pp. 1–12,

    work page 2022

  16. [16]

    A robust method to automatically detect fin whale acoustic presence in large and diverse passive acoustic datasets,

    E. Schall and C. Parcerisas, “A robust method to automatically detect fin whale acoustic presence in large and diverse passive acoustic datasets,”Journal of Marine Science and Engineering, vol. 10, no. 12, p. 1831, 2022

    work page 2022

  17. [17]

    Variability in the performance of the spectrogram correlation detector for north-east pacific blue whale calls,

    A. ˇSirovi´c, “Variability in the performance of the spectrogram correlation detector for north-east pacific blue whale calls,”Bioa- coustics, vol. 25, no. 2, p. 145–160, 2015

    work page 2015

  18. [18]

    Vocalisations of antarctic blue whales, balaenoptera musculus intermedia, recorded during the 2001/2002 and 2002/2003 iwc/sower circumpolar cruises, area v, antarctica,

    S. Rankin, D. Ljungblad, C. Clark, and H. Kato, “Vocalisations of antarctic blue whales, balaenoptera musculus intermedia, recorded during the 2001/2002 and 2002/2003 iwc/sower circumpolar cruises, area v, antarctica,”J. Cetacean Res. Manage., vol. 7, no. 1, p. 13–20, 2023

    work page 2001

  19. [19]

    Automatic detection and clas- sification of baleen whale social calls using convolutional neural networks,

    J. H. Rasmussen and A. ˇSirovi´c, “Automatic detection and clas- sification of baleen whale social calls using convolutional neural networks,”The Journal of the Acoustical Society of America, vol. 149, no. 5, p. 3635–3644, 2021

    work page 2021

  20. [20]

    Wavelet-based feature extraction with hidden markov model classification of antarctic blue whale sounds,

    O. P. Babalola and D. Versfeld, “Wavelet-based feature extraction with hidden markov model classification of antarctic blue whale sounds,”Ecological Informatics, vol. 80, p. 102468, 2024

    work page 2024

  21. [21]

    Cbam: Convolutional block attention module,

    S. Woo, J. Park, J.-Y. Lee, and I. S. Kweon, “Cbam: Convolutional block attention module,”arXiv preprint arXiv:1807.06521, 2018

    Pith/arXiv arXiv 2018

  22. [22]

    Consistent cross-validatory model-selection for de- pendent data: hv-block cross-validation,

    J. Racine, “Consistent cross-validatory model-selection for de- pendent data: hv-block cross-validation,”Journal of Econometrics, vol. 99, no. 449, pp. 39–61, 2000

    work page 2000

  23. [23]

    blockcv: An r package for generating spatially or environmentally separated folds for k-fold cross-validation of species distribution models,

    R. Valavi, J. Elith, J. J. Lahoz-Monfort, and G. Guillera-Arroita, “blockcv: An r package for generating spatially or environmentally separated folds for k-fold cross-validation of species distribution models,”Methods in Ecology and Evolution, vol. 10, no. 2, pp. 225– 232, 2019

    work page 2019

  24. [24]

    Ethical considerations and unanticipated consequences associated with ecological forecasting for marine resources,

    A. J. Hobday, J. R. Hartog, J. P. Manderson, K. E. Mills, M. J. Oliver, A. J. Pershing, and S. Siedlecki, “Ethical considerations and unanticipated consequences associated with ecological forecasting for marine resources,”ICES Journal of Marine Science, vol. 76, no. 5, pp. 1244–1256, 2019. [Online]. Available: https://doi.org/10.1093/icesjms/fsy210

    work page doi:10.1093/icesjms/fsy210 2019

  25. [25]

    Ethical ecosurveillance: Mitigating the potential impacts on humans of widespread environmental monitoring,

    N. Young, R. J. Lennox, J. R. Bennett, D. G. Roche, and S. J. Cooke, “Ethical ecosurveillance: Mitigating the potential impacts on humans of widespread environmental monitoring,”People and Nature, vol. 4, no. 4, pp. 830–840, 2022. [Online]. Available: https://doi.org/10.1002/pan3.10327

    work page doi:10.1002/pan3.10327 2022

  26. [26]

    Hyperparameter tuning and performance assessment of statistical and machine-learning algorithms using spatial data,

    P. Schratz, J. Muenchow, E. Iturritxa, J. Richter, and A. Brenning, “Hyperparameter tuning and performance assessment of statistical and machine-learning algorithms using spatial data,”Ecological Modelling, vol. 406, pp. 109–120, 2019

    work page 2019

  27. [27]

    The precision-recall plot is more informative than the roc plot when evaluating binary classifiers on imbalanced datasets,

    T. Saito and M. Rehmsmeier, “The precision-recall plot is more informative than the roc plot when evaluating binary classifiers on imbalanced datasets,”PLOS ONE, vol. 10, no. 3, p. e0118432, 2015. [Online]. Available: https://doi.org/10.1371/journal.pone.0118432

    doi:10.1371/journal.pone.0118432 2015

  28. [28]

    A comparative analysis of windowing approaches in dense sensing environments,

    A. McCallum, C. Nugent, I. Cleland, and P. McCullagh, “A comparative analysis of windowing approaches in dense sensing environments,”Proceedings, vol. 2, no. 19, p. 1245, 2018

    work page 2018

  29. [29]

    A unified approach to time–frequency representations and generalized spectrograms,

    E. Cordero, G. Giacchi, and L. Rodino, “A unified approach to time–frequency representations and generalized spectrograms,” Journal of Fourier Analysis and Applications, vol. 31, no. 9, pp. 1–28, 2025

    work page 2025

  30. [30]

    Activation functions in deep learning: A comprehensive survey and benchmark,

    S. R. Dubey, S. K. Singh, and B. B. Chaudhuri, “Activation functions in deep learning: A comprehensive survey and benchmark,” Neurocomputing, vol. 503, p. 92–108, Sep 2022

    work page 2022

  31. [31]

    Image lightness rescaling using sigmoidal contrast enhancement functions,

    G. J. Braun and M. D. Fairchild, “Image lightness rescaling using sigmoidal contrast enhancement functions,”Journal of the Imaging Science and Technology, vol. 54, no. 4, p. 040501, 2010

    work page 2010

  32. [32]

    Learning deep features for discriminative localization,

    B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning deep features for discriminative localization,” inProceed- ings of the IEEE conference on computer vision and pattern recognition, 2016, p. 2921–2929

    work page 2016

  33. [33]

    Deepening into the suitability of using pre- trained models of imagenet against a lightweight convolutional neural network in medical imaging: an experimental study,

    L. Alzubaidiet al., “Deepening into the suitability of using pre- trained models of imagenet against a lightweight convolutional neural network in medical imaging: an experimental study,”PeerJ Computer Science, vol. 7, p. e715, 2021

    work page 2021

  34. [34]

    Comparison of pre-trained cnns for audio classification using transfer learning,

    A. Guzhov, F. Raue, J. Hees, and A. Dengel, “Comparison of pre-trained cnns for audio classification using transfer learning,” Sensors, vol. 10, no. 4, p. 72, 2021

    work page 2021

  35. [35]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in3rd International Conference for Learning Representations, 2015. [Online]. Available: https://doi.org/10. 48550/arXiv.1412.6980

    Pith/arXiv arXiv doi:10.48550/arxiv.1412.6980 2015

  36. [36]

    Leveraging tropical reef, bird and unrelated sounds for superior transfer learning in marine bioacoustics,

    B. Williams, B. van Merri ¨enboer, V. Dumoulin, J. Hamer, E. Triantafillou, A. B. Fleishman, M. McKown, J. E. Munger, A. N. Rice, A. Lillis, C. E. White, C. A. D. Hobbs, T. B. Razak, K. E. Jones, and T. Denton, “Leveraging tropical reef, bird and unrelated sounds for superior transfer learning in marine bioacoustics,”

    work page

  37. [37]

    On the overlap between grad-cam saliency maps and explainable visual features in skin cancer images,

    F. Nunnari, M. A. Kadir, and D. Sonntag, “On the overlap between grad-cam saliency maps and explainable visual features in skin cancer images,” inMachine Learning and Knowledge Extraction (CD-MAKE 2021), ser. Lecture Notes in Computer Science, vol. 12844. Springer, 2021, pp. 241–253. [Online]. Available: https: //link.springer.com/chapter/10.1007/978-3-03...

    work page doi:10.1007/978-3-030-84060-0 2021

  38. [38]

    Ai assistance for grammar, structural edit- 12 ing, results extrapolation, and figure preparation with human oversight,

    Microsoft Copilot, “Ai assistance for grammar, structural edit- 12 ing, results extrapolation, and figure preparation with human oversight,” https://copilot.microsoft.com, 2025, assisted Nicholas Rasmussen in refining manuscript text, contextualizing results, and preparing figures. All core ideas and analyses by the author

    work page 2025

  39. [2022]

    Available: https://pmc.ncbi.nlm.nih.gov/articles/ PMC9589621/

    [Online]. Available: https://pmc.ncbi.nlm.nih.gov/articles/ PMC9589621/

    work page

  40. [2023]

    Available: https://doi.org/10.1093/iob/obad039

    [Online]. Available: https://doi.org/10.1093/iob/obad039

    work page doi:10.1093/iob/obad039

  41. [2024]

    Available: https://arxiv.org/abs/2404.16436

    [Online]. Available: https://arxiv.org/abs/2404.16436

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.