A deep learning framework for marine acoustic and seismic monitoring with distributed acoustic sensing

Abstract

Distributed Acoustic Sensing (DAS) enables high-resolution and long-duration monitoring of marine acoustic and seismic activity by turning existing fiber-optic cables into dense sensor arrays. However, extracting diverse signals from continuous DAS data remains challenging due to the massive data volumes and signal complexity. Here, we present DASNet, a deep learning framework for automated detection, classification, and arrival-time picking of diverse marine signals in DAS data. The model is trained using a semi-supervised pipeline on continuous recordings and jointly predicts spatiotemporal bounding boxes and segmentation masks for each detected event. Applied to three years of data from the Seafloor Fiber-Optic Array in Monterey Bay (SeaFOAM), DASNet identified over 500,000 events spanning multiple signal categories. For seismic monitoring, the model detects the majority of cataloged local earthquakes within 100 km and identifies distant earthquake-generated T-waves, with beamforming analysis revealing source azimuths clustered toward the southwestern Pacific and along mid-ocean ridge systems. For bioacoustic monitoring, DASNet detects and tracks more than 400,000 blue and fin whale calls, revealing seasonal and interannual variability consistent with independent hydrophone records. For anthropogenic activity, DASNet detects and localizes vessel traffic near the cable, with estimated positions validated against Automatic Identification System (AIS) tracks. These results demonstrate that combining DAS with deep learning provides a scalable, high-resolution monitoring approach for marine environmental observation. As submarine DAS deployments expand, this framework could substantially enhance seismic, bioacoustic, and anthropogenic observations in regions where conventional instrumentation remains sparse.

Figures (7)

• Figure 1: DASNet framework overview. (a) Model architecture with the feature extraction backbone, region proposal network, and output branches for classification, bounding box regression, and mask prediction. (b) Semi-supervised data engine workflow for iterative dataset expansion and model refinement. Only a subset of whale call categories is shown here; the complete set is presented in \ref{['fig:example_and_eval']}.
• Figure 2: SeaFOAM DAS array configuration and data characteristics. (a) Location of the SeaFOAM DAS array. The red line denotes the section used in this study, whereas the blue line indicates the unused portion. Black dots mark every 1,000 channels along the cable from the interrogator unit. The inset in the lower right shows the location of the cable in Monterey Bay, California. (b) Depth profile of the underwater DAS channels, overlaid with the daily median power spectral density (PSD) for each channel on 25 November 2023 (solid line). Shaded areas represent the 10th - 90th percentiles of PSDs, illustrating channel-to-channel signal variation.
• Figure 3: Model performance evaluation and signal examples. (a) Representative examples of each signal type detected in the SeaFOAM dataset. Earthquake and T-wave segments are bandpass filtered between 2–10 Hz, while the remaining signals are highpass filtered above 10 Hz to enhance relevant features. The corresponding spectrograms are shown in Figure S1. (b) Precision-recall (PR) curves for all signal categories evaluated on the test set at an intersection-over-union (IoU) threshold of 0.5. Stars on each curve mark the point corresponding to the highest F1 score. (c) Distribution of arrival-time residuals on the test set, computed as the difference between predicted and manually labeled arrival times. Predicted arrivals are extracted as the peak of the model’s probability mask using a threshold of 0.5. Boxplots display the median (bold horizontal line), interquartile range (IQR; box boundaries), and whiskers extending to 1.5 $\times$ IQR from the quartiles. Outliers beyond this range are omitted for clarity.
• Figure 4: Earthquake detection and localization results. (a) Comparison of the earthquake detection performance between DASNet and PhaseNet-DAS as a function of event magnitude and epicentral distance, where the distance is computed relative to the average location of the DAS cable. (b) Earthquake epicenters determined by combining P- and S-phase arrival times picked by DASNet on SeaFOAM DAS and by PhaseNet on the conventional seismic station BK.BUCR.
• Figure 5: T-wave detection and multi-array source localization. (a) T-wave waveforms recorded by the SeaFOAM cables, beginning at 2024-05-08 13:35:22 UTC. The red line indicates the theoretical arrival times computed from the estimated T-wave azimuth shown in (b) for the Vanuatu event (assuming a propagation velocity of 1.46 km$/$s, which provides the best fit to the observed signal onset), showing close agreement with the observed waveforms. (b) Heat maps derived via beamforming for azimuthal estimation. (c) Distribution of estimated T-wave source azimuths. Color shading within each azimuth bin indicates the distribution of signal amplitudes, with darker colors representing stronger arrivals. Three prominent azimuthal clusters are observed. (d) Map showing the locations of the SeaFOAM, OOI, and RESIF F1 DAS installations romanowicz2023seafoamshi2024multiplexedhttps://doi.org/10.15778/resif.f1. Arrows indicate the dominant azimuths of T-wave arrivals detected over the 3-year period. Yellow regions and dashed lines illustrate two examples of multi-array localization through azimuthal triangulation. Green star mark the epicenters of two earthquake events (Mw 5.2, 2024-05-08 11:51:39 UTC, Vanuatu; and Mw 5.2, 2025-05-14 16:33:05 UTC, Pacific-Antarctic Ridge) corresponding to the observed T-waves.