Best Practices for Machine Learning based Icequake Picking with Distributed Acoustic Sensing

Over the past years, a wide range of machine learning–based phase picking methods have been developed, primarily targeting three-component seismometer data from tectonic earthquakes. With the rapid growth of distributed acoustic sensing (DAS) applications, diversification of use cases, and availability of increasingly large DAS datasets, these methods are now being applied to single-component DAS recordings. However, their optimal use for DAS data and for alternative signal types such as cryoseismological events, remains rarely explored.
In this study, we present a systematic analysis of the performance of machine learning–based phase picking methods pretrained on tectonic earthquakes on one-component cryoseismological DAS data obtained on the Rhône Glacier in the Swiss Alps in July 2020. We evaluate multiple strategies for generating pseudo-three-component data from the intrinsically single-component DAS strain-rate data, including zero-padding of missing components, duplication of the single component, and the use of consecutive DAS channels as surrogate components. In addition, we assess the phase-picking performance across different preprocessing schemes, comparing conservatively band-pass filtered data with denoised data obtained using a J-invariant  autoencoder specifically trained on cryoseismological DAS data. Finally, we analyze the spatial and temporal distribution of located events over the full observation period and across the entire glacier. Event clusters are correlated with weather conditions, daily cycles, and the geometry of the glacier bed to explore potential patterns in cryoseismic activity.
Our results indicate that treating consecutive DAS channels as surrogate components yields the most reliable phase-picking performance, whereas extensive denoising can degrade picking accuracy. We further discuss spatial clusters of event locations and their correlations with glacier topography and meteorological conditions.