DeepDAS: An Earthquake Phase Identification Tool Using Submarine Distributed Acoustic Sensing Data

The EU-INFRATECH funded SUBMERSE project will establish continuous monitoring of several oceanic telecom cables for landing sites in Portugal, Greece, and Svalbard. We develop tools for the automated analysis of these upcoming data sets as well as other submarine DAS data. We try to leverage DeepLab v3, a Deep Neural Network (DNN) architecture for semantic segmentation, to train a machine learning model for earthquake detection and P and S wave picking using submarine Distributed Acoustic Sensing (DAS) data. The input data being two-dimensional necessitated the adoption of the DeepLab v3 model, known for its superior performance in image segmentation. The complexity of submarine DAS data, characterized by diverse ocean noise environment and levels, as well as varying parameters such as cable length, shape, channel spacing, deployment environment, and location, led us to employ a larger model for effective earthquake detection.   Given the scarcity of submarine DAS seismic records, we adopted a strategy to pre-train our model using terrestrial DAS seismic records before fine-tuning it with submarine DAS records including Madeira Island, Svalbard Island, Chile and Greece coast. This approach aimed to leverage the abundant and diverse seismic event data available from land-based DAS system to establish a robust base model. Subsequently, the model was fine-tuned using the scarcer, yet critically important, submarine DAS data to adapt its earthquake detection capabilities to the unique characteristics and challenges presented by the submarine environment. This two-step training process allowed us to efficiently exploit the available data resources, ensuring that our model benefited from a broad learning base while achieving specialized performance for submarine earthquake detection.   Our results demonstrate the model's robust ability to identify seismic events and label P and S waves accurately. For three-component seismometer data it is generally assumed, that distinction of P and S waves relies primarily on the polarization of the arrivals.  Our model's capacity to recognize P and S waves in single-component DAS data is therefore intriguing. We conducted tests under various scenarios to understand how the model discriminates between P and S waves: inverting the sequence of P and S waves did not affect identification performance, suggesting that order does not play a role. Even when the input consisted solely of P or S waves, the model could still identify them, indicating the identification is not based on their simultaneous or nonsimultaneous appearance. Amplifying the P wave's amplitude by five or ten times—surpassing the S wave's amplitude—still allowed for discrimination of P and S waves, albeit with diminished accuracy, highlighting that amplitude information is significant. However, our model also learned additional aspects we have yet to understand, suggesting it captures more complex patterns