{"id":"0b3ffff8-490e-4511-bb81-ea942ff73088","arxiv_id":"2509.11614","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A ten-day study shows machine learning on acoustic data from a submarine cable can detect vessels within 1 km with F1 above 90% and estimate distance with a mean error of 141 m.","lead":"This paper uses distributed acoustic sensing on a 28 km submarine fiber-optic cable to detect nearby ships and estimate how far they are, reporting above 90% detection F1 and a mean distance error of 141 m over ten days of North Sea data. It also releases a processed feature-label dataset so other groups can rerun the machine-learning experiments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AIS interpolation uncertainty (up to ~315 m/min) is larger than the reported 141 m localization MAE; the headline precision claim is not yet supported until error is quantified on non-interpolated AIS timestamps.","rationale":"The reader's weakest assumption identifies exactly the load-bearing concern I would raise: interpolated AIS positions are used as hard ground truth for a metric (141 m MAE) that is smaller than the stated interpolation uncertainty (~315 m/min). The paper has genuine strengths—ten days of continuous data, a released feature-level dataset, a beamforming baseline, and clear experimental structure—so this is not a rejection. But the localization claim is not independently verifiable from the current evaluation because the ground truth itself is smoothed to a coarser scale than the claimed error. I also considered the test-fold selection issue (final configuration and feature bands chosen using test-fold results), which is real and should be disclosed, but it is secondary: the AIS label problem affects the validity of the metric itself. The proposed check is directly runnable on the released dataset because the released archive contains timestamps and labels, so the authors or any reviewer can stratify by distance to the nearest raw AIS report. If the check shows that near-report MAE is similar to 141 m, the concern does not land and the headline is credible. If not, the paper should reframe the localization result as a relative, smoothed estimate rather than a physical precision claim. Therefore the reader's CONDITIONAL verdict is the right one, and my read does not change it.","tokens_in":876,"tokens_out":842,"duration_ms":76040,"concrete_test":"Using the released Marlinks-NS DAS dataset, recompute the localization MAE using only test frames whose labels are within 15 s of an original, non-interpolated AIS report, and separately for frames near the midpoint of an interpolation gap. If the near-report MAE remains close to 141 m (e.g., within ~20%), the interpolation concern is not decisive. If the near-report MAE is substantially larger, the headline MAE is dominated by interpolation smoothing and should be reported as an upper bound on model precision, not a validated physical localization error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is the 141 m MAE for vessel distance estimation. That number is computed against AIS positions linearly interpolated to 1-second resolution (Section IV-B.1). The paper itself states that 88% of vessels report only every 1–3 minutes and that this 'can lead to significant positional uncertainty (approximately 315 meters per minute in average).' A reported error of 141 m is smaller than the stated uncertainty of the labels against which it is computed. This does not mean DAS cannot sense vessels; it means the MAE conflates model error with label interpolation error. Because the interpolated distance-to-cable curve is smooth, and because the final system uses 50 s temporal context plus 5-window majority voting, the model may be learning to reproduce the smoothed AIS trajectory rather than estimating true physical distance. The same label noise also affects the detection F1-score near the 1000 m threshold, though the detection claim is more robust. This is not an external-consensus disagreement: it is an internal inconsistency between the reported label uncertainty and the precision of the metric derived from those labels.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a DAS-based vessel detection and distance-estimation system using 10 days of continuous recordings from a 28 km submarine fiber-optic cable in the Southern Bight of the North Sea, with AIS as ground truth. The pipeline computes 100 log-spaced spectral-band energies (4–98 Hz, excluding 49–51 Hz) from 10-second windows, optionally integrates 10–250 sensors and 10–50 s temporal contexts, and trains XGBoost or feed-forward NN classifiers/regressors. A 10-fold, day-level cross-validation is used. The authors report an overall F1-score above 90% for detecting a vessel within 1000 m of the cable and a mean absolute error of 141 m for distance estimation, and they compare the regressor with SRP beamforming. The processed feature dataset is publicly released. The central numerical claims are weakened by the fact that the AIS labels are linearly interpolated to 1-second resolution from updates that are 1–3 minutes apart for 88% of vessels, with a stated positional uncertainty of about 315 m per minute, which exceeds the reported 141 m MAE.","tokens_in":20464,"tokens_out":6626,"duration_ms":83342,"significance":"If the quantitative claims hold, this would be a practically valuable result: it would show that an existing telecom cable, a commercial DAS interrogator, and gradient-boosted trees can serve as an all-weather, near-real-time cable-protection monitor in open-water conditions, and the released processed dataset would be a useful community resource. The paper has real strengths: the temporal split into day-level folds is a serious attempt to avoid leakage, the class-imbalance issues are acknowledged and handled in the metric reporting, bootstrap confidence intervals are provided, and the comparison with a classical SRP beamformer grounds the ML claims. The qualitative match between DAS feature energy and AIS-derived vessel proximity (Figs. 8d–e) is convincing evidence that the signal is physically present. However, two load-bearing issues—label uncertainty larger than the claimed precision, and model/feature selection performed on the same data used for testing—prevent the headline F1 and MAE from being accepted as unbiased estimates of operational performance.","major_comments":[{"comment":"The reported 141 m MAE is smaller than the uncertainty of the labels against which it is computed. Section IV-B.1 states that 88% of vessels report every 1–3 minutes and that linear interpolation introduces 'significant positional uncertainty (approximately 315 meters per minute in average).' A linearly interpolated AIS trajectory is a smooth curve, and the final system uses 50 s temporal context plus 5-window voting, so the model can learn to reproduce that smoothed curve rather than estimate true physical distance. The MAE in Eq. (1) therefore conflates model error with label-placement error. Please compute the MAE only on frames coinciding with raw AIS messages (or within a short, stated tolerance), stratify the error by time since last AIS update, and report an uncertainty bound on the labels. Without such an evaluation, the 141 m precision claim is not supported. The detection F1 is","section":"Section IV-B.1 and Eq. (1), Section VI-F"},{"comment":"The day-level 10-fold CV prevents temporal leakage for a fixed processing pipeline, but the pipeline is not fixed. The spectral feature bands (Section VI-C), distance threshold, number of sensors, temporal context, majority-voting length, and averaging strategy (Sections VI-E and VI-F) were all selected after inspecting results on the full dataset, including the test folds. This selection-on-test makes the reported F1 and MAE optimistic estimates of generalization. Please hold out at least one full day (or use nested cross-validation) for all configuration choices and report the performance of the frozen pipeline. At minimum, state clearly that the quoted numbers are post-model-selection and therefore not unbiased.","section":"Sections VI-C, VI-E, VI-F"},{"comment":"There is an apparent inconsistency in the headline localization numbers. The abstract and conclusions quote 141 m MAE, while Table II reports 171 m for the XGBoost regressor in the beamformer comparison, and Section VI-H says the 1000 m task with 10-second windows 'achieved a MAE of 171m.' If 141 m is obtained only with the final 50 s / 5-vote configuration, that configuration should be stated wherever the number appears, and the beamforming comparison should be run on the same configuration. As written, a reader cannot tell whether the methods are being compared on equal terms.","section":"Abstract/Conclusions vs. Table II and Section VI-H"}],"minor_comments":[{"comment":"'mean average error' should be 'mean absolute error' (also in the conclusions).","section":"Abstract"},{"comment":"'ACI' appears to be a typo for 'ATI' (spatial+temporal averaging), and the notation 'A[CH,TI' is incomplete.","section":"Figure 17 caption"},{"comment":"'V essel' has a spacing typo.","section":"Index Terms"},{"comment":"The Introduction motivates the method by the ability to detect non-cooperative 'dark' ships without AIS, but the supervised labels are entirely AIS-derived, so the evaluation cannot measure dark-ship detection. Please add an explicit limitation statement that the reported results apply to AIS-cooperative vessels.","section":"Introduction and Section IV-D"},{"comment":"The label is described as the distance from the cable to the closest vessel. For the 250-sensor (2.5 km) configuration, clarify whether this is the minimum distance over the whole segment and discuss how a vessel near one end of the segment is labeled for sensors far from it. This affects the interpretation of the regression target and of the spatial-averaging results.","section":"Section V-C / data labeling"},{"comment":"State explicitly whether the SRP comparison uses the same temporal context and majority voting as the final ML system. As written, Table II compares a 10-second-window regressor with a 10-second-window beamformer, while the final system uses 50 s context and voting.","section":"Section VI-G"}],"recommendation":"major_revision","confidential_remarks":"This is a promising field study with a useful public dataset and a physically convincing signal. The main risk is that the headline numbers are compared against labels whose interpolation uncertainty exceeds the claimed precision. If the authors can provide an evaluation on raw AIS timestamps and an unbiased model-selection protocol, the paper would be a strong contribution. The scope fits the journal well."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe main thing to know: this is the largest real-world DAS vessel monitoring study to date, and it comes with a publicly released dataset (feature vectors, not raw DAS). The detection story holds up; the 141 m localization MAE does not, because the AIS labels against which it is computed are themselves uncertain by several hundred meters per minute.\n\nWhat's new: scale (ten days, 74,771 frames, 28 km cable), the explicit cable-protection framing with distance-to-cable regression, a published dataset, and a fair SRP/beamforming comparison. The experimental design is also more careful than most work in this area: day-level 10-fold CV avoids temporal leakage, class imbalance is discussed and handled, and bootstrap CIs are reported. The qualitative figures show clear vessel-related energy along the fiber, so the basic physical effect is real.\n\nSoft spots, in order of severity:\n\n1. The AIS interpolation problem. The paper states that 88% of vessels report every 1–3 minutes, giving roughly 315 m/min positional uncertainty, yet the 141 m MAE is computed against linearly interpolated 1-second positions. That is a label-noise problem larger than the reported error, so the MAE conflates model error with interpolation error (and with 50 s context plus 5-window voting, the model may partly be tracking the smoothed AIS curve). The detection F1 at the 1000 m threshold is less affected, since a mislabeled distance near the boundary doesn't change the qualitative result. The authors acknowledge the interpolation issue, but they don't quantify its impact on the reported metric.\n\n2. Configuration selection. The spectral bands (4–98 Hz, excluding 49–51 Hz) and the final system (250 sensors, 50 s context, 5 voting windows) were chosen after looking at the full dataset, so the evaluation is not fully blind. Day-level folds prevent direct temporal leakage, but not this broader kind of selection bias. The reported numbers are likely optimistic.\n\n3. Smaller: the released dataset contains features and labels, not raw DAS data, so independent re-analysis of the preprocessing is limited.\n\nWho this is for: anyone working on DAS-based maritime surveillance will want the dataset and the pipeline details. The central claim that DAS plus standard ML can detect nearby vessels is credible. The localization claim should be treated as conditional until it is re-evaluated at original (non-interpolated) AIS timestamps or with a proper uncertainty model.\n\nRecommendation: send it to peer review. A serious referee should ask for a re-analysis of the localization metric, but the study is substantial and worth the referees' time.","headline":"Large-scale DAS vessel detection with a released dataset and credible detection results, but the headline 141 m localization MAE is undermined by AIS interpolation uncertainty larger than the reported error.","tokens_in":21002,"tokens_out":2461,"would_cite":true,"duration_ms":27788,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A submarine telecom cable, repurposed as a distributed acoustic sensor, can detect vessels within 1 km at over 90% F1-score and estimate their distance to about 141 meters.","keywords":["vessel detection","distributed acoustic sensing","submarine cable protection","vessel distance estimation","XGBoost","AIS ground truth","maritime surveillance","optical fiber sensing"],"falsifier":"Run a controlled vessel with a high-rate GPS receiver (e.g. 1 Hz or better) past a DAS-instrumented cable for a few hours, compute the model's distance errors against those GPS positions instead of interpolated AIS, and compare with 141 m. If the error against high-rate GPS is substantially above 141 m—or if the model misses a vessel whose AIS is off—the paper's central deployment claim does not survive.","tokens_in":20095,"feed_emoji":"🚢","tokens_out":3993,"duration_ms":45418,"temperature":0.7,"pith_summary":"This paper argues that existing submarine fiber-optic cables can serve as practical, all-weather monitors for cable protection: a data-driven pipeline that turns DAS strain recordings into spectral features and feeds them to machine-learning models can tell when a vessel is within 1000 m of the cable and how far away it is. The claim is grounded in a continuous ten-day open-water deployment on a 28 km cable, covering many vessel types and conditions, with about 74,000 frames and a released dataset. The best detection system reaches an overall F1-score above 90% with balanced class performance, and the best regression system reaches a mean absolute error of 141 m for distance estimation, beating a beamforming baseline. If true, this bridges experimental DAS work and real deployment by showing that standard cable infrastructure plus ML can provide continuous, weather-independent surveillance without relying on AIS.","feed_headline":"DAS turns submarine cable into vessel detector at 90% F1","feed_subtitle":"Ten days of open-sea data show a telecom fiber plus machine learning can estimate ship distance to 141 meters.","key_machinery":"The key machinery is a DAS-plus-ML pipeline built on logarithmically spaced frequency-band energy features. Each 10-second frame from each sensing position is reduced to 100 band energies over 4–98 Hz (with the 49–51 Hz band removed to suppress interrogator harmonics), then averaged across up to 250 channels and concatenated in 50-second temporal contexts; XGBoost classifiers/regressors consume these features, with majority voting over five windows. AIS trajectories linearly interpolated to 1-second resolution provide the labels. This feature design follows from a spectral analysis showing a clear vessel-vs-noise contrast below 100 Hz and from the gauge-length Nyquist argument that caps reso","core_discovery":"The central discovery claim is that vessel proximity to a buried submarine cable can be learned directly from DAS spectral energy. The authors select a 1000 m protection threshold, integrate spectral features from 250 channels spanning ~2.5 km, use 50-second contexts and majority voting, and train XGBoost models under a day-based 10-fold cross-validation that avoids temporal leakage. On unseen days, detection achieves >90% overall F1 with balanced class-wise scores, and vessel distance regression achieves 141 m MAE; the same ML regression beats an SRP beamforming implementation (225 m MAE) on vessels within 1 km. The paper also identifies a physical limitation: with a 10 m gauge length and ~","pith_inferences":["The reported 141 m MAE is measured against AIS positions interpolated at 1-second resolution from updates that arrive every 1–3 minutes for 88% of vessels. Since the authors quote ~315 m of positional uncertainty per minute, the true physical localization error could be lower or higher than 141 m; a high-rate GPS validation would settle it.","Because useful energy concentrates below 100 Hz, the same features may carry information about vessel speed, size, and type, so extending the pipeline to those tasks is a natural next step that the authors flag as future work.","A model trained on one cable segment with a particular burial depth, bathymetry, and traffic mix may not transfer directly to other deployments; testing on diverse public DAS datasets would show how much retraining is needed.","If the method generalizes, fused DAS-AIS monitoring could shift from reaction to prediction: cable-protection systems could issue proximity alerts minutes before crossing rather than after radar/AIS contact."],"forward_implications":["Cable operators can deploy a protection alarm on an existing fiber with no new underwater hardware, only an interrogator and trained models.","Because detection does not rely on cooperative AIS transmissions, the method could flag vessels that have switched off or spoofed AIS—dark ships.","The ML distance estimator is both more accurate and computationally lighter than grid-search beamforming, making continuous real-time surveillance feasible along long cables.","The released dataset enables other groups to reproduce the 141 m/90% figures and benchmark alternative models on the same real-world conditions.","The 10-minute early response seen in examples implies warnings can be raised before a vessel actually crosses the cable, leaving time for intervention."],"fun_headline_variants":["Submarine cable listens for ships: 90% detection, 141m range","Telecom fiber becomes ship radar: 10 days, 90% F1, 141m error","Ship tracking via existing submarine cables: ML beats beamforming","10-day trial turns undersea fiber into ship locator","Existing internet cables can pinpoint ships to 141m"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is that AIS positions linearly interpolated to 1-second intervals are an accurate ground truth for vessel-to-cable distance; if those labels are off by hundreds of meters, the detection and localization numbers are measuring label noise as much as physical accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Submarine cable listens for ships: 90% detection, 141m range","Telecom fiber becomes ship radar: 10 days, 90% F1, 141m error","Ship tracking via existing submarine cables: ML beats beamforming","10-day trial turns undersea fiber into ship locator","Existing internet cables can pinpoint ships to 141m"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000488,"raw_usage":{"total_tokens":2275,"prompt_tokens":811,"completion_tokens":1464,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":1368}},"tokens_in":555,"tokens_out":1464,"duration_ms":14251,"temperature":1.0,"reasoning_tokens":1368,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:44:26.474550+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled vessel with a high-rate GPS receiver (e.g. 1 Hz or better) past a DAS-instrumented cable for a few hours, compute the model's distance errors against those GPS positions instead of interpolated AIS, and compare with 141 m. If the error against high-rate GPS is substantially above 141 m—or if the model misses a vessel whose AIS is off—the paper's central deployment claim does not survive.","supporting_citations":[],"review_version":2}