{"id":"d92c6d38-4707-480c-b9f6-117ee41ef17a","arxiv_id":"2608.02387","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"An end-to-end workflow detects, characterizes, and localizes fin whale notes in DAS data with high precision and moderate recall.","lead":"Fin whale calls are automatically detected, characterized, and located from dense fiber-optic ocean-cable recordings using a kurtosis-based picker, clustering, and hyperbolic fitting. The pipeline turns raw distributed acoustic sensing (DAS) data into usable whale-song information, potentially enabling cheap, large-scale marine monitoring.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Localization and inferred movement rest on an unvalidated constant-speed, fixed-depth model; no ground-truth positions are provided, so the movement inference could be an artifact.","rationale":"The paper's central claim is not only the detection metrics; the title and abstract foreground localization, and the movement inference in Figs. 6 and 8 is presented as a headline capability. The detection metrics, though potentially in-sample, are at least internally quantified and backed by open-source code. In contrast, the localization stage has no external ground truth: the only checks are internally consistent score maps and biologically plausible speeds. The reader's weakest_assumption correctly identifies the hyperbolic moveout with constant apparent velocity and the fixed-depth constant-sound-speed grid as the load-bearing model assumption. I agree that this is the weakest point because if multipath, depth changes, or sound-speed variability break the hyperbola or the 1500 m/s grid, the localization score maps become uncalibrated and the inferred movement could be an artifact. A synthetic forward-model test is the clearest way to settle the concern: it tests the full pipeline under known conditions and quantifies the sensitivity to the assumptions. This does not change the verdict: the paper remains CONDITIONAL, with the condition being independent validation of the localization stage (synthetic or real ground truth).","tokens_in":13241,"tokens_out":6963,"duration_ms":78931,"concrete_test":"Run a fully synthetic end-to-end check: place a moving fin-whale-like source at known positions and depths along the actual Estepona cable geometry, generate DAS strain data with a range-dependent sound-speed profile and realistic noise/multipath, and run the full KVP+clustering+localization pipeline. Compare the estimated 99th-percentile score-map contours and inferred displacement/speed against the true trajectory. Repeat for source depths of 20 m and 50 m and for sound speeds of 1480 and 1520 m/s. If the inferred movement direction or speed changes by more than 20%, or if the true source position falls outside the estimated contour, the localization model is not robust enough to support the movement claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing gap is that the localization and apparent-movement claims (Sec. 4.3, Figs. 6 and 8) are not validated against any independent ground truth. The clustering and outlier rejection rely on the single-hyperbola moveout of Eq. 2 (constant apparent velocity), and the grid search in Sec. 3.4 assumes a fixed source depth and a constant 1500 m/s sound speed. These assumptions are plausible for an ideal homogeneous straight-cable geometry, but the cables have bathymetry and curvature; multipath, cable-coupling variations, and depth changes could distort relative arrival times. Since the score maps (Eqs. 7–8) are only internally consistent, the reported apparent swimming speeds (3.31 and 6.48 km/h) and movement directions could be artifacts of the assumed propagation model rather than true motion. The paper itself notes in Sec. 5.3 that score maps 'should not be interpreted as calibrated localization uncertainty,' yet the abstract still claims inference of apparent source movement, so the central claim depends on this unvalidated link.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an end-to-end, unsupervised workflow for detecting, characterizing, and localizing fin whale notes in distributed acoustic sensing (DAS) data from two submarine telecom cables. The pipeline adapts the kurtosis-value picker (KVP) with a Morlet wavelet for narrow-band note detection, groups channel-wise picks via DBSCAN and agglomeration, fits a hyperbolic moveout model to reject outliers, extracts temporal/spectral/energy features for note typing, and applies a grid-search relative-arrival-time method to produce matching-score maps for source localization. Detection is evaluated against manual annotations from six fin whale songs, reporting median pick-level precision 0.990 and recall 0.744, and cluster-level precision 0.880 and recall 0.806. Representative examples illustrate overlapping-song separation, type-A/B characterization, and inference of apparent source movement at speeds of 3.31 and 6.48 km/h.","tokens_in":13572,"tokens_out":2670,"duration_ms":25974,"significance":"If the results hold, the workflow is a useful contribution to DAS-based marine mammal monitoring: it is fully unsupervised, interpretable, open-source, and demonstrated on two real-world cable deployments with contrasting geometry and noise conditions. A notable strength is that the detection stage does not require labeled training data, and the code is publicly available. The paper also correctly limits some claims, noting that score maps are not calibrated uncertainty. However, the two most consequential claims—transferable detection performance and localization/movement inference—currently rest on thinner evidence: the metrics are computed on the same data used to tune the pipeline, and the localization stage has no independent validation. These gaps need to be addressed before the paper's conclusions can be taken at face value.","major_comments":[{"comment":"The detection metrics (median pick precision 0.990, recall 0.744; cluster precision 0.880, recall 0.806) are reported on the same six songs that were presumably used to develop and tune the KVP settings and clustering parameters. There is no held-out set or cross-validation, so the reported numbers likely overestimate performance on new recordings. This is load-bearing for the paper's 'transferability' claim over two datasets. Please provide an explicit train/test split or cross-validation, or clearly state which parameters were fixed a priori and which were tuned on these six songs.","section":"Sec. 4.1 and Table 1"},{"comment":"The spatio-temporal clustering stage depends on several parameters that are not reported: DBSCAN neighborhood radius eps, min_samples, temporal/spatial agglomeration margins, and the hyperbolic-fit outlier residual threshold. Table 1 lists only the KVP/Morlet settings. Without these values (or a pointer to a configuration file in the GitHub repository), the workflow cannot be reproduced or adapted to other DAS arrays. Please include a complete parameter table or cite the exact configuration in the open-source code.","section":"Sec. 3.2 (DBSCAN and agglomeration parameters)"},{"comment":"The abstract claims 'inference of apparent source movement,' but the localization results are not validated against any independent ground truth, synthetic test, or known source position. The method assumes a constant sound speed of 1500 m/s, a fixed source depth, and a horizontal propagation path (Sec. 3.4), and the clustering/outlier rejection already assumes a single-hyperbola moveout with constant v_app (Eq. 2). The paper's own Sec. 5.3 states that score maps 'should not be interpreted as calibrated localization uncertainty.' Without validation, the reported swimming speeds of 3.31 and 6.48 km/h could be artifacts of the assumed propagation model rather than true whale motion. Please add a quantitative validation (e.g., synthetic waveforms, a known source, or a comparison with independent tracking) or explicitly reframe the movement claims as model-dependent 'apparent' displacements","section":"Sec. 4.3 and Sec. 5.3 (localization and movement inference)"},{"comment":"The hyperbolic moveout model in Eq. (2) is central to both clustering and localization, but the cables are not straight: Fig. 1 shows bathymetric variations and curvature. If the along-fiber distance d differs from the straight-line distance to the source, the arrival-time pattern may deviate from a pure hyperbola, affecting both outlier rejection and the grid-search localization. The paper does not quantify this error or test the sensitivity of clustering/localization to cable geometry. Please add a sensitivity analysis or at least a clear statement of the geometric assumptions and their expected impact on the reported results.","section":"Sec. 3.2, Eq. (2) and Sec. 2 (bathymetry)"}],"minor_comments":[{"comment":"Typo: 'KPV' should be 'KVP' in the second paragraph ('In the present study, KPV was adapted').","section":"Sec. 3.1"},{"comment":"Typo: 'spectrogram chacracteristics' -> 'spectrogram characteristics'; also 'descrptors' in Sec. 4.2 and 'reatining' in Sec. 5.1.","section":"Sec. 3.3"},{"comment":"The phrase 'adirectly' in Sec. 5.3 is a typo, and Fig. 8 caption contains '99 th percentile os each' — should be 'of each'.","section":"Sec. 5.3 / Fig. 8 caption"},{"comment":"The dimensionless factor σ is set to 1.5, which is described as defining the signal support. It would help to state explicitly that this value was chosen by visual inspection and whether the features are sensitive to it.","section":"Sec. 3.3, Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the open-source code is a plus. The central detection/characterization workflow is plausible, but the detection evaluation lacks a held-out set and the localization/movement claims are currently unsupported by independent validation. These are load-bearing for the abstract, so I recommend major revision rather than rejection; the issues are addressable with additional experiments or by substantially tempering the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a genuine integration job: KVP adapted to Morlet wavelets for fin whale notes, followed by DBSCAN clustering, agglomeration, hyperbolic fitting, feature extraction, and a grid-search localizer. None of the pieces are brand new, but putting them together into a working end-to-end pipeline that runs on two real submarine cables and quantifying detection performance is a real contribution. The detection evaluation is the strongest part: six manually annotated songs, median pick precision 0.990 and recall 0.744, cluster precision 0.880 and recall 0.806. Those are credible numbers, and the open-source code helps. I also appreciate that the authors discuss limitations in the discussion rather than hiding them.\n\nThe soft spots are real but not fatal. Most importantly, the evaluation is likely in-sample: parameters appear to be tuned on the same six songs, with no held-out set, so the reported precision/recall probably overstate field performance. Several DBSCAN thresholds and other parameters are not given in the text, which limits reproducibility even with code. The localization section is the weakest: source positions are never compared with any ground truth. The constant 1500 m/s, fixed-depth, single-hyperbola model is a reasonable starting point, but the paper itself cautions that the score maps should not be interpreted as calibrated localization uncertainty. That makes the abstract's claim about inferring apparent source movement a bit stronger than the evidence supports. The movement speeds in Figs. 6 and 8 could be sensitive to the assumed propagation model, and the bilateral ambiguity is acknowledged but not resolved.\n\nThe note-type characterization and inter-note-interval estimates are shown only on examples, not validated quantitatively, so they should be seen as illustrative. Overall, the detection and characterization parts are defensible; the localization is promising but needs either independent ground truth or a much softer conclusion.\n\nRecommendation: send it to peer review. The paper deserves referee time. Ask the authors for a held-out evaluation or at least a parameter-sensitivity analysis, disclose the missing thresholds, and either validate the positions or clearly mark the movement inference as tentative. With those changes it will be a useful contribution to the DAS bioacoustics community.","headline":"Useful integrated DAS workflow for fin whale notes with credible detection numbers, but the localization and movement claims outrun the validation.","tokens_in":679,"tokens_out":713,"would_cite":true,"duration_ms":27381,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An unsupervised workflow turns submarine fiber-optic DAS recordings into fin whale note detections, characterizations, and locations, with median pick-level precision of 0.990.","keywords":["distributed acoustic sensing","fin whale","passive acoustic monitoring","KVP picker","hyperbolic localization","note characterization","submarine fiber-optic cable"],"falsifier":"A controlled test would be to record a fin whale with a collocated hydrophone array and a DAS cable, localize the caller with the hydrophones, and check whether the DAS-derived matching-score peak coincides with the true location; if the peak is systematically offset or the hyperbola residuals are large in a sizable fraction of notes, the central claim is weakened.","tokens_in":13171,"feed_emoji":"🐋","tokens_out":3950,"duration_ms":33416,"temperature":0.7,"pith_summary":"The paper claims that raw distributed acoustic sensing (DAS) data from submarine telecom cables can be converted, without any supervised training, into a compact set of fin whale note detections, each labeled with temporal, spectral, and energy features and assigned a candidate source location. On six manually annotated songs, the method reaches median pick-level precision 0.990 and recall 0.744, and cluster-level precision 0.880 and recall 0.806. It also distinguishes type-A (20-Hz downsweep) from type-B (backbeat) notes and infers apparent swim direction from the sequence of locations. A sympathetic reader would care because DAS arrays are dense and already deployed, so this offers a scalable route to long-term monitoring and song-structure analysis.","feed_headline":"Fin whale notes auto-detected from undersea fiber cables at 99% precision","feed_subtitle":"An unsupervised DAS workflow detects, characterizes, and localizes whale notes without training data.","key_machinery":"The key machinery is the hyperbolic arrival-time model t(d)^2 = t0^2 + (d-d0)^2/v_app^2, which links each note to a localized source and drives both the rejection of incoherent picks and the grid-search localization. Around it sit the kurtosis-value picker (KVP) with a Morlet wavelet for narrow-band detection, DBSCAN plus hierarchical agglomeration to group channel-wise picks, and a spectrogram-based feature extractor producing centroids, spreads, slopes, and SNR. The matching score M = 1/(1+epsilon^2) converts arrival-time residuals into a spatial map of candidate source positions.","core_discovery":"The central discovery is that a multi-stage signal-processing chain—a kurtosis-based picker adapted with a Morlet wavelet, density-based spatio-temporal clustering, cluster agglomeration, hyperbolic arrival-time fitting, and grid-search localization—can extract biologically interpretable note-level information from raw DAS strain records. The hyperbolic fitting rejects incoherent picks, yielding validated clusters whose features support note-type discrimination, and whose relative arrival times produce matching-score maps of the source. The evaluation against manual annotations shows that the method is both precise and transferable across two different cable geometries.","pith_inferences":["If the hyperbolic assumption holds, the same pipeline could track individual whales over longer timescales by linking cluster apices, a step the paper leaves for future work.","The matching-score maps, though not calibrated uncertainties, could be used to design better cable geometries (curved or multi-cable) to remove the bilateral ambiguity—an extension hinted but not pursued.","The feature set might generalize to other pulsed low-frequency tonal calls (e.g., sei or Bryde's whales) if the wavelet scale and frequency band are adjusted; this is an untested extrapolation.","A natural stress test is to apply the workflow to a dataset with known ground-truth positions (e.g., tagged whales) to calibrate localization accuracy, which the current evaluation does not do."],"forward_implications":["DAS-equipped submarine cables can serve as continuous passive acoustic monitors for fin whales without labelled training data.","Note-level features enable estimation of inter-note intervals for AA, AB, BA, BB sequences, allowing song-structure and population studies.","Sequential localization of notes yields apparent swimming speed and direction, with estimates in this paper around 3.3–6.5 km/h, consistent with known behavior.","The workflow is transferable to hydrophone arrays, cabled observatories, or OBS networks with minor adaptation.","Overlapping vocalizations from two individuals can be separated when their hyperbola apices differ, as demonstrated on a 15 January 2024 recording."],"fun_headline_variants":["Undersea cables listen for fin whales with 99% precision","Automated fin whale detection from telecom cables","Fiber-optic cables reveal fin whale calls and locations","Whale songs decoded from undersea fiber optics","Unsupervised fin whale detection on undersea fiber cables"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that every fin whale note generates a single hyperbolic arrival-time pattern on the DAS array with a constant apparent velocity; if multipath propagation, depth changes, or variable cable coupling break that hyperbola, the clustering validation and localization results become unreliable.","fun_headline_variants_meta":{"raw":{"variants":["Undersea cables listen for fin whales with 99% precision","Automated fin whale detection from telecom cables","Fiber-optic cables reveal fin whale calls and locations","Whale songs decoded from undersea fiber optics","Unsupervised fin whale detection on undersea fiber cables"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000925,"raw_usage":{"total_tokens":3808,"prompt_tokens":755,"completion_tokens":3053,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":2983}},"tokens_in":499,"tokens_out":3053,"duration_ms":20729,"temperature":1.0,"reasoning_tokens":2983,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T08:08:53.716779+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled test would be to record a fin whale with a collocated hydrophone array and a DAS cable, localize the caller with the hydrophones, and check whether the DAS-derived matching-score peak coincides with the true location; if the peak is systematically offset or the hyperbola residuals are large in a sizable fraction of notes, the central claim is weakened.","supporting_citations":[],"review_version":1}