Unsupervised learning applied to cryoseismic signals: identification of glacier processes from the Whillans Ice Stream

• Main Authors: Latto, R, Turner, R, Reading, A, Winberry, JP, Kulessa, B, Cook, S
• Format: Conference Object
• Language: English
• Published: . 2021
• Subjects: Earth Sciences; Geophysics; Seismology and seismic exploration; Antarctic; Ross Ice Shelf; West Antarctica; Whillans; Whillans Ice Stream; Antarc*; Antarctica; Ice Shelf
• Online Access: http://ecite.utas.edu.au/150132

Seismic deployments in Antarctica are contributing further datasets to the growing field of cryoseismology. Such deployments are motivated by the opportunity to improve understanding of active glacier processes, especially those that are hidden from satellite reconnaissance. However, as a barrier to analysis, the inherently complex seismic wavefield of Antarctic glaciers imposes challenges to standard methods of detection and research. A potential solution is a data-driven, semi-automated approach that combats the difficulties associated with identifying and understanding the characteristically weak amplitude, diverse signals that are present in the continuous seismic data streams from glaciers. For these purposes, we present a methodology to extract patterns of information from a robust and heterogeneous event catalogue, using the machine learning algorithm k-means++. As a case study, we apply our methods to seismic data collected from an array deployed to the Whillans Ice Stream, West Antarctica from December 14, 2010 -January 31, 2011. Then, we use the event catalogue to form a database of characteristic features of the waveforms for each event, such as duration, spectral content, polarity, and aspects of network geometry. Following a manual appraisal of the catalogue and database, we group the event features with the k-means++ algorithm to find event types that we expect to represent glacier deformation mechanisms. The advantage of the semi-automated process is that we can employ prior knowledge of some of the potential clusters to guide evaluation of the yielded clusters. For example, we validate the machine learning outputs by corroborating the times and features of certain clustered events with identified (i.e. labelled) stick-slip events. Other clusters of deformation we expect include melt-related processes, teleseisms, and microseisms related to ocean and other processes impacting the nearby Ross Ice Shelf. Our approach could be used as a standard workflow to allow comparisons to be made over time and ...