Large-scale bioacoustic detection using semantic segmentation: a deep learning framework applied to fin whale calls in ocean-bottom seismometer recordings
Jocelyn Japnanto, Alex A. Saoulis, Miriam Romagosa, Rita Leitão, Gabrielle Arrieta, Mónica A. Silva, Matthew Graham, Ana M. G. Ferreira
physics.geo-ph · cs.LG
Quick verdict
Applies U-Net segmentation to detect 6.3M fin whale calls from 46 OBSs; strong cross-sensor precision (~97%) but UPFLOW recall 9–53% limits spatial claims.
Summary
The authors adapt a U-Net-style semantic segmentation model (ResNet-18 encoder, ImageNet initialisation) to detect fin whale 20-Hz notes in spectrograms from ocean-bottom seismometers. The model was trained on ~40,000 expert annotations (~350 hours) from a single hydrophone station (SJO03, São Jorge array) and applied without retraining to vertical-component seismometer recordings from the 40-station UPFLOW array across the Azores–Madeira–Canaries region. From 378,912 hours at 46 stations it produced 6.29 million detections, presented as the largest fin whale call catalogue to date. Precision remained high (~94% on the São Jorge test set; ~97%, range 93–100%, across 16 validated UPFLOW stations), but recall fell sharply on UPFLOW (33% overall, 9–53% per station), with a systematic onset bias of about -2.8 s; missed detections concentrate at low SNR. The catalogue shows calls concentrated October–February, appearing earlier at north-western stations, and three persistent inter-note-interval groups (~12, 16, 20 s) whose composition shifts seasonally. Main limitations: station-dependent detection completeness is uncorrected yet underpins the spatial-ecological interpretation, there is no baseline-detector comparison, no confidence intervals are reported, and INI validation is restricted to the training-domain station. Code and the catalogue are publicly released.
Review complete
Assessment at a glance
4
risk flags identified
● high 1
● medium 2
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single segmentation model trained on one hydrophone station (SJO03) transfers without retraining to seismometer data from a geographically distinct 40-station array, retaining at least 93% precision at all 16 validated UPFLOW stations.
- The pipeline produced 6,291,626 detections from 378,912 hours at 46 stations, which the authors present as the largest fin whale call catalogue assembled to date.
- The catalogue resolves call timing and spectral structure accurately enough for ecological analysis, evidenced by INI errors of 0.80 s mean / 0.045 s median / 2.56 s MAE at SJO03.
- Three persistent INI groups (~12, 16, 20 s) shift coherently in composition across the singing season, and calling concentrates October-February with earlier onset at north-western stations.
Strengths
Unprecedented scale with full artifact release
6,291,626 detections over 378,912 hours at 46 stations (Table 1); catalogue on Zenodo (DOI 10.5281/zenodo.21065182), code on GitHub (anonymous.4open.science/r/WhaleSeg-2B13), waveforms public via EIDA networks 4U/8J, fixed random seed for training.
Genuine cross-sensor transfer result
Trained only on SJO03 hydrophone data, applied without retraining to UPFLOW vertical seismometers; precision stayed at or above 0.93 at every one of the 16 validated stations (Table 2), corroborating Dréo et al. (2025) on hydrophone-seismometer correspondence.
Transparent failure analysis
Post-processing ablation with worked example
INI structure validated against annotations
Weaknesses
Selective headline framing of performance
Unexplained systematic onset bias on UPFLOW
Thin evaluation statistics
Spatial ecology confounded by station-dependent detectability
No baseline comparison and a training-data nuance
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This method-plus-application paper trains a U-Net-style semantic segmentation network (ResNet-18 encoder, ImageNet initialisation; Section 2.4) to detect fin whale 20-Hz notes in spectrograms from ocean-bottom seismometers (OBS). It makes four central claims: (1) a model trained on hydrophone data from one station (SJO03) transfers without retraining to vertical-component seismometer data from the 40-station UPFLOW array; (2) the resulting catalogue — 6,291,626 detections over 378,912 hours at 46 stations (Table 1) — is the largest fin whale call catalogue to date; (3) the catalogue resolves timing and spectral structure 'accurately enough' for ecological analysis; and (4) three persistent INI groups (~12, 16, 20 s) shift coherently across the season, with calling concentrated October-February and appearing earlier at north-western stations (Sections 3.2-3.3).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How many fin whale calls were detected in this study?
The detector produced 6,291,626 detections (about 6.3 million) from 378,912 hours of recordings at 46 ocean-bottom seismometer stations in the Azores-Madeira-Canaries region, which the authors present as the largest fin whale call catalogue to date. Precision was about 97% on the large UPFLOW array, but recall was 33% overall (9-53% by station), meaning many low-SNR calls were missed.
Can a model trained on hydrophone data work on seismometer data?
Partially, yes. Without any retraining, the model kept precision at or above 93% across 16 validated seismometer stations, showing the spectral representation transfers across sensor types. However, recall dropped sharply (33% overall) and detected call onsets carried a systematic bias of about 2.8 seconds early, so completeness and timing are weaker on seismometers than on the training-domain hydrophone station.
What fin whale song patterns did the catalogue reveal?
Calling was concentrated from October to February, appearing earlier at north-western (more offshore) stations. Three persistent inter-note-interval (INI) groups centred near 12, 16, and 20 seconds appeared at nearly every station, and their relative mix shifted coherently across the season: the 12 s group dominated in October, while the 16 s group dominated from January onward.
Is the code or data from this paper available?
Yes. The codebase is on GitHub via anonymous.4open.science/r/WhaleSeg-2B13, the full 6.3-million-call detection catalogue is on Zenodo (DOI 10.5281/zenodo.21065182), and the raw waveforms are public through the Earthscope and GFZ EIDA data centres under network codes 4U and 8J.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are raster images in the PDF extraction; assessments of Figs. 1, 4-10 and S1-S8 rely on their captions and in-text descriptions, not visual inspection of the figure content.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.