DeepDAS: An Earthquake Phase Picker for Submarine Distributed Acoustic Sensing Data

Given the scarcity of seismometers in marine environments, traditional seismology has limited effectiveness for early earthquake warning in oceanic regions. Submarine Distributed Acoustic Sensing (DAS) systems offer a promising alternative for seismic monitoring in these areas. The EU-INFRATECH funded SUBMERSE project will establish continuous monitoring of several oceanic telecom cables for landing sites in Portugal, Greece, and Svalbard.  However, the existing machine learning models trained on land-based DAS data do not perform well with submarine DAS due to differences in noise characteristics, deployment conditions, and environmental factors. 

This study presents a machine learning approach tailored specifically for submarine DAS data to enable automated seismic event detection and P and S wave identification. Leveraging DeepLab v3, a neural network architecture optimized for semantic segmentation, we developed a specialized model to handle the unique challenges of submarine DAS data. Our model was trained and validated on a dataset comprising nearly 92 million manually and semi-automatically labeled seismic records from multiple international submarine sites, providing a robust basis for accurate seismic detection. We compared the performance of DeepDAS and PhaseNet DAS in picking seismic P and S waves from submarine DAS data. Our findings suggest that DeepDAS (F1 score 0.89) outperforms PhaseNet DAS (F1 score 0.53,) in those datasets. This result is understandable, as PhaseNet DAS was originally trained on DAS seismic data from land-based DAS.

Beyond developing the model, we generated a comprehensive submarine DAS earthquake dataset with manually picked P and S arrivals. This dataset includes 6,326 submarine seismic events (magnitudes ranging from -2 to 5, depths from 0 to 200 km) and spans diverse deployment scenarios with varying cable lengths, configurations, and channel spacings. Recognizing the importance of open collaboration and reproducibility, we plan to open-source this dataset. We aim to establish it as a benchmark dataset for submarine DAS research, enabling broader adoption and facilitating advancements in the field.