Efficient neural network-based detection of seismicity in fibre optic data from Store Glacier, West Greenland

Seismic surveys are widely used to characterise the properties of glaciers, their basal material and conditions, and ice dynamics. The emerging technology of Distributed Acoustic Sensing (DAS) uses fibre optic cables as seismic sensors, allowing observations to be made at higher spatial resolution than possible using traditional geophone deployments. Passive DAS surveys generate large data volumes from which the rate of occurrence and failure mechanism of ice quakes can be constrained, but such large datasets are computationally expensive and time consuming to analyse. Machine learning tools can provide an effective means of automatically identifying seismic events within the data set, avoiding a bottleneck in the data analysis process.

Here, we present a novel approach to machine learning for a borehole-deployed DAS system on Store Glacier, West Greenland. Data were acquired in July 2019, as part of the RESPONDER project, using a Silixa iDAS interrogator and a BRUsens fibre optic cable installed in a 1043 m-deep borehole. The data set includes controlled-source vertical seismic profiles (VSPs) and a 3-day passive record of cryoseismicity.  To identify seismic events in this record, we used a convolutional neural network (CNN). A CNN is a deep learning algorithm and a powerful classification tool, widely applied to the analysis of images and time series data, i.e. to recognise seismic phases for long-range earthquake detection.

For the Store Glacier data set, a CNN was trained on hand-labelled, uniformly-sized time-windows of data, focusing initially on the high-signal-to-noise-ratio seismic arrivals in the VSPs. The trained CNN achieved an accuracy of 90% in recognising seismic energy in new windows. However, the computational time taken for training proved impractical. Training a CNN instead to identify events in the frequency-wavenumber (f-k) domain both reduced the size of each data sample by a factor of 340, yet still provided accurate classification. This decrease in input data volume yields a dramatic decrease in the time required for detection. The CNN required only 1.2 s, with an additional 5.6 s to implement the f-k transform, to process 30 s of data, compared with 129 s to process the same data in the time domain. This suggests that f-k approaches have potential for real-time DAS applications.

Continuing analysis will assess the temporal distribution of passively recorded seismicity over the 3 days of data. Beyond this current phase of work, estimated source locations and focal mechanisms of detected events could be used to provide information on basal conditions, internal deformation and crevasse formation. These new seismic observations will help further constrain the ice dynamics and hydrological properties of Store Glacier that have been observed in previous studies of the area.

The efficiency of training a CNN for event identification in the f-k domain allows detailed insight to be made into the origins and style of glacier seismicity, facilitating further development to passive DAS instrumentation and its applications.