Section outline

  • SUBMERSE (SUBMarine cablEs for ReSearch and Exploration) is an innovative EU-funded project that aims to turn existing submarine telecommunication cables into an international fiber optic sensing network, providing near real-time data flows to monitor the Earth and its systems.

    During the SUBMERSE Community Event in April 2026, this training session introduced the application of SeisBench, an open-source python toolbox for machine learning in seismology, to earthquake detection on Distributed Acoustic Sensing (DAS) records. This  practical tutorial demonstrated how the SeisBench library can help process the complex, high-volume DAS data produced by submarine cables, with the DAS-native model DeepSubDAS applied to an earthquake record used as example.

    • Picking DAS records with deep learning: DeepSubDAS and the SeisBench DAS interface File PDF
    • Notebook: SeisBench and DeepSubDAS URL

      Open this Notebook and follow more instructions on the video tutorial starting at 11:09 

      This notebook provides a comprehensive tutorial on using SeisBench for DAS (Distributed Acoustic Sensing) phase picking. It introduces how to install SeisBench with DAS support and import necessary libraries like seisbench.models and xdas. The tutorial then guides through loading and inspecting example DAS data, which is handled efficiently using xdas's virtual data arrays to manage large datasets.

      Two main approaches for phase picking are demonstrated:

      1. Wrapping 3C models: This section shows how to adapt models trained on traditional 3-component seismic data for DAS data using DASWaveformModelWrapper. It covers the process of moving models to preferred devices  for acceleration and introduces the concept of callbacks for handling large model outputs without memory overload.
      2. DeepSubDAS | A DAS-native model: This section highlights the advantages of using models specifically designed for DAS data, such as DeepSubDAS, which leverage the spatial coherence of DAS signals. It reiterates the process of loading the model, moving it to a GPU, and using callbacks for processing, demonstrating that DAS-native models yield more coherent and confident picks.

      Finally, the notebook addresses larger-than-memory processing capabilities of SeisBench. It explains how xdas handles virtual input data and how WriterCallback can be used to write large probability functions to disk, allowing the processing of datasets that exceed available memory.

      Documentation

      > helmholtz.software/software/seisbench

      > seisbench.readthedocs.io/en/stable/