Distributed Acoustic Sensing (DAS)

• Distributed Acoustic Sensing (DAS) is a fiber-optic technology that converts phase shifts in Rayleigh backscatter into distributed axial strain measurements.
• It provides long-range, high-resolution sensing for applications like seismic monitoring, structural health, marine acoustics, and urban infrastructure.
• Key parameters such as gauge length, channel spacing, and coupling conditions critically influence the system’s sensitivity and measurement accuracy.

Distributed Acoustic Sensing (DAS) is a fiber-optic sensing technology that turns an optical fiber into a long, quasi-continuous array of vibration sensors by interrogating the phase of Rayleigh backscatter and converting strain-induced optical path-length changes into distributed measurements of axial strain or strain rate. Across the literature, DAS is described as providing long-range coverage, fine spatial granularity, and a broad dynamic sensing range, with practical utility in seismic monitoring, structural health monitoring, transportation sensing, marine acoustics, sea-ice mechanics, and urban infrastructure observation. A recurring theme is that DAS is not a direct pressure sensor or a collection of isolated point sensors; rather, it is a distributed measurement system whose response depends on gauge length, channel spacing, cable geometry, fiber-ground or fiber-structure coupling, and the directional relation between incident wavefields and the fiber axis (Shi et al., 26 Mar 2025, Bölt et al., 21 Nov 2025, Xenaki et al., 25 Feb 2025).

1. Physical basis and measurement principle

DAS operates by launching coherent laser pulses into an optical fiber and measuring the backscattered light generated by microscopic refractive-index inhomogeneities. In the ocean-oriented formulation, the refractive index along the fiber axis is written as

$n(z)=n_f+\Delta n(z),$

and the emitted optical field for a pulse of duration $T_p$ is

$$E_p(t)=p\!\left(\frac{t}{T_p}\right)\cos(\omega t+\phi_0).$$

A scatterer at location $z_s$ returns a pulse after a two-way delay $2z_s/c_n$, so for a single scatterer the backscattered field is modeled as

$$E_s(\tau\mid z_s)=E_p\!\left(\tau-2\frac{z_s}{c_n}\right)s(z_s),$$

while the full received signal is the superposition

$$E_b(\tau)=\int_{-\infty}^{\infty}E_p\!\left(\tau-2\frac{z_s}{c_n}\right)s(z_s)\,dz_s.$$

The distributed measurement follows because optical time of flight maps to distance along the cable (Xenaki et al., 25 Feb 2025).

The sensing step is the interferometric comparison of backscatter from nearby fiber sections separated by a gauge length $L_g$. In one formulation, the differential phase is

$\Delta\Phi = k(2\Delta L),$

and also

$$\Delta\Phi = \frac{4\pi n \xi L_g}{\lambda}\frac{\Delta L}{L_g},$$

which yields the axial strain estimate

$T_p$0

The same literature makes explicit that DAS measures strain along the cable, not acoustic pressure itself. For a cable aligned with the $T_p$1-axis,

$T_p$2

and for an impinging acoustic plane wave the response contains both directional sensitivity and gauge-length spatial filtering: $T_p$3 This $T_p$4 dependence is central for ocean acoustics and more generally for interpreting DAS as an axial sensor with geometry-dependent sensitivity (Xenaki et al., 25 Feb 2025).

Several papers emphasize equivalent strain-rate formulations. For near-surface surface-wave imaging, DAS is written as

$T_p$5

with a geophone-to-DAS comparison

$T_p$6

For bridge monitoring, the relation between beam curvature and longitudinal strain is expressed as

$T_p$7

with

$T_p$8

These formulations underscore that DAS is most naturally interpreted as a distributed strain or strain-rate sensor whose physical observable depends on how deformation projects onto the fiber axis (Dalkhani et al., 2023, Mercerat et al., 28 Oct 2025).

A common misconception is that DAS is a drop-in replacement for microphones, hydrophones, or geophones. The literature instead describes it as a distinct sensing modality with predominantly axial sensitivity, finite gauge-length averaging, and strong dependence on coupling, cable construction, burial state, and deployment geometry. This is particularly explicit in ocean, bridge, and campus-scale studies, where identical physical sources can appear differently depending on cable orientation, structural coupling, or preprocessing choices (Mercerat et al., 28 Oct 2025, Morell-Monzó et al., 8 Apr 2026, Bölt et al., 21 Nov 2025).

2. Acquisition parameters, array geometry, and signal representations

DAS data are inherently two-dimensional: fast time maps to distance along the fiber, while repeated interrogation pulses define the temporal evolution of the wavefield. The literature repeatedly identifies pulse width, pulse repetition frequency, gauge length, and channel spacing as the acquisition parameters that control spatial resolution, bandwidth, and signal-to-noise ratio. In the ocean overview, pulse width $T_p$9 and pulse length $$E_p(t)=p\!\left(\frac{t}{T_p}\right)\cos(\omega t+\phi_0).$$0 determine spatial averaging, $$E_p(t)=p\!\left(\frac{t}{T_p}\right)\cos(\omega t+\phi_0).$$1 must typically be longer than $$E_p(t)=p\!\left(\frac{t}{T_p}\right)\cos(\omega t+\phi_0).$$2, and the repetition frequency is limited by fiber length through

$$E_p(t)=p\!\left(\frac{t}{T_p}\right)\cos(\omega t+\phi_0).$$3

When $$E_p(t)=p\!\left(\frac{t}{T_p}\right)\cos(\omega t+\phi_0).$$4, the combined directivity becomes

$$E_p(t)=p\!\left(\frac{t}{T_p}\right)\cos(\omega t+\phi_0).$$5

These relations formalize the standard trade-off between spatial resolution, high-frequency sensitivity, long-range coverage, and optical SNR (Xenaki et al., 25 Feb 2025).

Field deployments show the diversity of operating regimes. The Albula River dataset used for machine-learning benchmarking spans about 10 km with a measurement length of 8 m, channel spacing of 4 m, and 512 channels; the data were downsampled to 20 Hz and transformed into 25.6-second intervals rendered as $$E_p(t)=p\!\left(\frac{t}{T_p}\right)\cos(\omega t+\phi_0).$$6 images (Shi et al., 26 Mar 2025). The Hamburg WAVE proto-network interrogated 12.132 km of heterogeneous fiber with a Silixa MK1 iDAS interrogator at 1 kHz and a 10 km subsegment with an OptaSense ODH4 at 10 kHz, using gauge lengths of 10 m and spatial channel spacing of 1 m (Bölt et al., 21 Nov 2025). The highway-bridge study used 0.8 m channel spacing, a gauge length of 4 m, and 8 kHz acquisition later downsampled to 500 Hz for operational modal analysis (Mercerat et al., 28 Oct 2025). Ocean-bottom vessel monitoring near Zeebrugge used a 28 km cable, gauge length $$E_p(t)=p\!\left(\frac{t}{T_p}\right)\cos(\omega t+\phi_0).$$7, spatial resolution of $$E_p(t)=p\!\left(\frac{t}{T_p}\right)\cos(\omega t+\phi_0).$$8, $$E_p(t)=p\!\left(\frac{t}{T_p}\right)\cos(\omega t+\phi_0).$$9 channels, and $z_s$0 (Ramirez-Torres et al., 15 Sep 2025). SeaFOAM in Monterey Bay recorded at 200 Hz over 10,245 channels with 5.1 m spacing along a 52 km array (Zhang et al., 16 Mar 2026).

Signal representation is correspondingly application-specific. Raw recordings are often visualized as waterfall images indexed by distance and time (Morell-Monzó et al., 8 Apr 2026). In the Albula benchmark, the chosen representation is the cross-spectral density matrix phase map,

$z_s$1

where the phase structure yields image-like textures used for classification (Shi et al., 26 Mar 2025). In multispectral DAS analysis, the signal $z_s$2 is decomposed into predefined frequency bands,

$z_s$3

and represented through band-limited energy images

$z_s$4

stacked into a spectral cube $z_s$5 for visualization and machine learning (Morell-Monzó et al., 8 Apr 2026).

This diversity of representations reflects a general methodological shift: DAS records are commonly recast as structured spatiotemporal or multichannel images rather than treated as isolated 1D traces. A plausible implication is that representation choice is not a cosmetic step but a modeling decision that determines which physics—dispersion, coherence, modal structure, spectral occupancy, moveout, or texture—becomes legible to downstream algorithms (Shi et al., 26 Mar 2025, Morell-Monzó et al., 8 Apr 2026, Zhang et al., 16 Mar 2026).

3. Analytical methodologies and inversion strategies

A large fraction of DAS research uses array-processing and inverse methods that exploit dense spatial sampling. In near-surface surface-wave imaging, Multi Offset Phase Analysis (MOPA) is applied in moving windows to estimate local dispersion curves from the relation

$z_s$6

with lateral variability handled by allowing $z_s$7. The resulting local phase velocities are inverted jointly in a 2D transdimensional Bayesian tomography parameterized by Voronoi cells and explored with reversible-jump Markov chain Monte Carlo under

$z_s$8

The method is designed to preserve lateral correlation and reduce nonuniqueness relative to independent 1D inversions (Dalkhani et al., 2023).

Bridge monitoring uses operational modal analysis rather than travel-time or dispersion inversion. After preprocessing, filtering from 0.5 to 50 Hz, and windowing into 20.48 s segments, the bridge study constructs the cross-power spectral density matrix, applies singular value decomposition, identifies resonant peaks, and estimates modal shapes via Frequency Domain Decomposition (FDD). DAS data alone were sufficient to identify resonances at approximately 4.9 Hz, 7.6 Hz, 8.8 Hz, 12 Hz, 14.5 Hz, and higher modes above 20 Hz, while collocated three-component seismometers were useful for discriminating the main motion direction of each mode (Mercerat et al., 28 Oct 2025).

Sea-ice mechanics combines active-source and passive-wave analysis. Longitudinal-wave dispersion is used to estimate Young’s modulus through

$z_s$9

while hydroelastic flexural waves satisfy

$2z_s/c_n$0

Continuous Wavelet Transform is then applied to filtered strain-rate fields to track hydroelastic wavenumber evolution: $2z_s/c_n$1 with angle correction

$2z_s/c_n$2

This framework enabled mapping of effective sea-ice thickness, Young’s modulus, and flexural rigidity along a 600 m transect (Kuchly et al., 15 Jan 2026).

Several studies are organized around geometric inference and source localization. In edge traffic monitoring on the Åstfjord bridge, moving vehicles produce straight-line signatures in the spatiotemporal image, detected through the probabilistic Hough transform using

$2z_s/c_n$3

and consolidated with DBSCAN using the line-segment distance

$2z_s/c_n$4

In ship detection and distance estimation on submarine cables, 100 logarithmically spaced band energies over 4–98 Hz feed XGBoost or neural-network models (Ramirez-Torres et al., 15 Sep 2025). In the dual-bistatic forward-sensing proposal, source localization is formulated by time difference of arrival,

$2z_s/c_n$5

with the Cross Ambiguity Function used as a maximum-likelihood estimator: $2z_s/c_n$6 Although that system is explicitly described as DAS-inspired rather than conventional backscatter DAS, it illustrates how distributed optical sensing concepts have propagated into communication-and-sensing architectures (Truong et al., 2024, Ramirez-Torres et al., 15 Sep 2025, Grythe et al., 24 Sep 2025).

4. Application domains

DAS has developed into a multi-domain observational technology rather than a single-purpose seismic instrument. In transportation and built infrastructure, it has been used for railway- and roadway-adjacent monitoring, bridge modal identification, urban traffic detection, and bridge traffic plus structural-response monitoring. The highway-bridge study in southeastern France identified the first eight normal modes of a strut-frame overpass and reported strong seasonal effects: the first transverse mode shifted from about 4 Hz in winter to 4.9–5 Hz in summer, while the first longitudinal mode at 5.1 Hz in winter 2019 was almost absent in summer 2022 and reappeared at about 5.2 Hz in winter 2024 (Mercerat et al., 28 Oct 2025). In Tel Aviv University’s Klausner Street deployment, a DAS-only U-Net trained from video-derived labels achieved test results of 94.2% detection accuracy, 94.0% classification accuracy, Dice loss 0.581, and false alarm rate 1.28%, then supported about one week of traffic analytics over about 350 m of fiber (Cohen et al., 2024). On the Åstfjord bridge, DAS with edge computing detected vehicles in near real time and also observed S-waves in the bridge concrete with a speed of about $2z_s/c_n$7 (Truong et al., 2024).

In seismology and geophysics, DAS has been used for earthquake monitoring, arrival-time picking, surface-wave imaging, and rapid transient-event observation. The semi-supervised PhaseNet-DAS framework treated dense fiber data as a 2D spatial-temporal problem, generated pseudo-labels with PhaseNet plus GaMMA, and on Long Valley data produced about 36 million P picks and about 53 million S picks, associated into 9,588 earthquakes with more than 2,000 associated P and S picks (Zhu et al., 2023). The OSIRIS-REx re-entry deployment near Eureka, Nevada, recorded the first reported DAS observation of a sample return capsule entry, including an impulsive arrival and extended coda, though with lower signal-to-noise than collocated seismometer-infrasound pairs and unexpectedly reduced low-frequency content (Carr et al., 2024).

Marine and ocean-bottom uses form another major branch. The ocean overview describes whale calls, ship noise, T waves from earthquakes, ocean surface gravity waves, microseisms, and Scholte waves as accessible to submarine DAS, particularly at low frequencies (Xenaki et al., 25 Feb 2025). The SeaFOAM DASNet study extended this to three years of continuous data, identifying more than 500,000 events including over 68,000 T-wave events, 320,000 fin whale calls, and 80,000 blue whale calls (Zhang et al., 16 Mar 2026). In the Southern Bight of the North Sea, submarine-cable DAS combined with AIS labels achieved over 90% $2z_s/c_n$8-score for vessel detection and a mean average error of 141 m for vessel distance estimation at the operationally emphasized 1000 m threshold (Ramirez-Torres et al., 15 Sep 2025). In controlled wave-tank experiments, optical fibers embedded in submarine power cables were used for sea-state monitoring, with wave period estimated with about 0.7–0.8% error, wave-height estimation showing RMSPE $2z_s/c_n$9, and direction-of-arrival estimated with about 1.5° error when at least two cable laying angles were available (Yajima et al., 27 Apr 2026).

Cryospheric and environmental mechanics provide further examples. On sea ice in the St. Lawrence Estuary, a 600 m fiber crossing three morphological regions recovered ice thicknesses from 25 cm to 68 cm and Young’s modulus values between 4.5 GPa and 5.7 GPa, in good agreement with collocated geophone arrays and drill hole thickness measurements (Kuchly et al., 15 Jan 2026). On research campuses and large scientific infrastructures, the Hamburg WAVE proto-network showed that DAS can observe natural, anthropogenic, and infrastructural vibrations across DESY, European XFEL, PETRA III, and the University of Hamburg, including a magnitude 7.4 Qinghai earthquake, persistent ocean-generated microseisms, facility noise sources, and previously ambiguous tones traced to helium compressors (Bölt et al., 21 Nov 2025).

At the city scale, DAS has also been framed as an urban sensing backbone. A graph-theoretic study of urban DAS argued that low coverage below 10% can, with optimal design, resolve earthquake early warning, groundwater monitoring, geological mapping, and urban activity tracking; that a percolation transition occurs at 51.6% coverage; and that only effectively complete coverage enables infrastructure monitoring, individual vehicle tracking, and pedestrian movement analysis (Cohen et al., 15 Jun 2026). This suggests that the application envelope of DAS is constrained as much by network topology and coverage geometry as by interrogator physics.

5. Machine learning, automation, and representation learning

Machine learning has become a central mechanism for making DAS operational at scale because raw DAS streams are high-dimensional, noisy, and often weakly labeled. The Albula River benchmark explicitly organizes ML contributions into preprocessing and representation learning, feature extraction, and event recognition or classification. Using augmented cross-spectral-density phase maps and a split of 8:1:1 into training, validation, and test data, it compares logistic regression, SVM, KNN, XGBoost, MLP, ResNet, LSTM, Transformer, and Mamba. Among classical methods, SVM performed best with 0.950 accuracy and 0.913 F1-score; among deep models, ResNet was strongest with 0.958 accuracy and 0.932 F1-score, followed by MLP at 0.953 accuracy and 0.924 F1-score and Transformer at 0.952 accuracy and 0.914 F1-score (Shi et al., 26 Mar 2025).

Other studies show that the form of supervision is highly varied. PhaseNet-DAS uses semi-supervised learning: a pre-trained PhaseNet teacher generates noisy pseudo-labels channel by channel, GaMMA refines those picks with a 1 s time window, and a DAS-specific 2D U-Net-style model with 7 × 7 kernels and 4 × 4 stride is trained on $$E_s(\tau\mid z_s)=E_p\!\left(\tau-2\frac{z_s}{c_n}\right)s(z_s),$$0 patches. Under a differential-time benchmark using 2,539 event pairs and about 9 million differential-time measurements, the resulting residuals were mean 0.001 s and std 0.06 s for P waves, and mean 0.005 s and std 0.25 s for S waves (Zhu et al., 2023). DASNet in Monterey Bay instead adapts Mask R-CNN, treating each event as an object in the channel-time plane with a bounding box, class label, and probability mask; on a held-out SeaFOAM test set, F1 scores were typically $$E_s(\tau\mid z_s)=E_p\!\left(\tau-2\frac{z_s}{c_n}\right)s(z_s),$$1 for earthquakes, T-waves, and fin whale calls (Zhang et al., 16 Mar 2026).

A separate line of work focuses on physically interpretable representations. The multispectral framework decomposes $$E_s(\tau\mid z_s)=E_p\!\left(\tau-2\frac{z_s}{c_n}\right)s(z_s),$$2 into predefined frequency bands, computes $$E_s(\tau\mid z_s)=E_p\!\left(\tau-2\frac{z_s}{c_n}\right)s(z_s),$$3, and uses stacked band-energy images as both a visualization and feature-extraction basis. On whale-vocalization data, a ResNet-18 classifier trained on three-band composites using 16–28 Hz, 30–40 Hz, and 40–60 Hz achieved 97.3% accuracy, 97.2% precision, and 97.1% recall (Morell-Monzó et al., 8 Apr 2026). This complements the Albula benchmark’s image-like CSDM phase maps and suggests that frequency-structured encodings can retain physical meaning while remaining compatible with off-the-shelf vision backbones (Shi et al., 26 Mar 2025, Morell-Monzó et al., 8 Apr 2026).

DAS research also includes data-reduction-oriented ML. ORION addresses near real-time seismic channel selection through a geometry-aware clustering stage followed by waveform-based selection using SNR, local coherence, and pre-event RMS amplitude. On the August 2024 Kefalonia sequence, ORION reduced data volume by about two orders of magnitude and improved the median distance to the reference catalog from 10.69 km for uniform selection to 3.98 km (Bozzi et al., 9 Dec 2025). Compressed-domain vibration detection and classification goes further upstream by avoiding reconstruction. Using compressed sensing with a 30% measurement ratio, compressed-domain frequency band energy features reduced transmitted data size by 70% while achieving 99.4% true positive rate and 0.04% false positive rate along 5 km sensing fiber, together with 95.05% classification accuracy on a 5-class task (Shen et al., 2022).

The literature also includes hybrid-labeling strategies. In urban traffic monitoring, YOLOv11 detections from video are mapped onto the DAS time-distance plane as Gaussian probability blobs, and a U-Net then learns a DAS-only detector used at deployment time (Cohen et al., 2024). In large-scale surface-wave identification, a physics-informed labeling rule based on spectral content and amplitude statistics produced a 14-layer ResNet that reached final accuracy 94.67% and enabled inference on an additional 170 GB of data in less than 30 minutes using parallel computing (Dumont et al., 2020). Taken together, these studies show that DAS machine learning is not a single paradigm but a set of solutions spanning weak supervision, semi-supervision, physics-informed labels, compressed sensing, and geometry-aware selection.

6. Limitations, unresolved issues, and research directions

The limitations reported across DAS studies are consistent and technically consequential. First, coupling and deployment conditions strongly shape sensitivity. Research-campus monitoring emphasizes uneven coupling to the ground and variable recording quality in dark-fiber networks (Bölt et al., 21 Nov 2025). Sea-ice deployments report uncoupled intervals in rough zones and energy-intensive interrogators that limited observations to one hour per day (Kuchly et al., 15 Jan 2026). The OSIRIS-REx campaign shows that fiber type, placement, vegetation proximity, and surface deployment can dominate transient-event data quality (Carr et al., 2024). Ocean and maritime studies likewise note sensitivity to burial depth, seabed coupling, cable construction, and precise channel location uncertainty (Xenaki et al., 25 Feb 2025, Ramirez-Torres et al., 15 Sep 2025).

Second, DAS remains a directional and spatially filtered sensor. This is explicit in the $$E_s(\tau\mid z_s)=E_p\!\left(\tau-2\frac{z_s}{c_n}\right)s(z_s),$$4 angular dependence for ocean acoustics and in the fact that bridge, sea-ice, and campus studies frequently combine DAS with seismometers or geophones to recover motion components that DAS alone does not directly observe (Xenaki et al., 25 Feb 2025, Mercerat et al., 28 Oct 2025, Kuchly et al., 15 Jan 2026). A common misconception is that dense channel spacing removes this issue; the literature instead indicates that channel density does not eliminate axial sensitivity, gauge-length averaging, or differential coupling between nearby channels.

Third, noise and self-noise remain active issues. The Hamburg WAVE double-redundant-loop study provided an upper bound on DAS self-noise by subtracting common-mode signals from co-located fibers, finding that residual differential spectra often followed an approximately $$E_s(\tau\mid z_s)=E_p\!\left(\tau-2\frac{z_s}{c_n}\right)s(z_s),$$5 trend and that the residual noise floor rose by about a factor of ten as channel separation increased from 250 m to 2914 m (Bölt et al., 21 Nov 2025). The same paper notes that common-mode rejection below 0.1 Hz was limited to roughly a factor of 10 even for the strongest Vibrotruck peaks, constraining noise cancellation and channel stacking. The OSIRIS-REx study adds a separate but related observation: unexpectedly weak sub-15 Hz DAS content relative to seismometer and infrasound records (Carr et al., 2024).

Fourth, generalization and operational deployment remain difficult. The Albula benchmark notes sensitivity to environmental and seasonal variation, limited labeled data in specific geographic regions or operating conditions, and real-time deployment challenges because of computational cost (Shi et al., 26 Mar 2025). The bridge case documents strong seasonal shifts in modal frequencies and shapes (Mercerat et al., 28 Oct 2025). The North Sea vessel study highlights class imbalance, imperfect AIS ground truth, and finite detection range (Ramirez-Torres et al., 15 Sep 2025). Urban graph-theory work adds an infrastructural constraint: current city fiber topologies are fragmented, motivating short-range, on-chip DAS rather than assuming long uninterrupted optical paths (Cohen et al., 15 Jun 2026).

The principal research directions identified in the literature are correspondingly pragmatic. The Albula benchmark highlights transfer learning, domain adaptation, class-imbalance handling, federated learning, multimodal fusion, and improved interpretability (Shi et al., 26 Mar 2025). PhaseNet-DAS points to more advanced semantic-segmentation backbones, iterative self-training, and combination with conventional seismic networks (Zhu et al., 2023). Sea-ice work suggests longer recordings for passive interferometry and ambient-noise inversion (Kuchly et al., 15 Jan 2026). The Hamburg WAVE initiative emphasizes low-latency analytics, robust geo-referencing, and better understanding of fiber-ground coupling and system noise (Bölt et al., 21 Nov 2025). A plausible synthesis is that the next phase of DAS development depends less on demonstrating detectability and more on controlling coupling, representation, calibration, and cross-domain generalization under operational constraints.

In aggregate, the contemporary literature presents DAS as a distributed strain-sensing platform whose strength lies in continuous spatial coverage, meter-scale sampling, and compatibility with existing fiber infrastructure. Its most robust advances arise where optical physics, array processing, inverse methods, and machine learning are treated as a single integrated system rather than as separable stages (Xenaki et al., 25 Feb 2025, Shi et al., 26 Mar 2025, Zhang et al., 16 Mar 2026).