Characterization of seismicity from different glacial bed types: machine learning classification of laboratory stick-slip acoustic emissions

Abstract

Abstract Subglacial seismicity presents the opportunity to monitor inaccessible glacial beds at the epicentral location and time. Glaciers can be underlain by rock or till, a first order control on bed mechanics. Velocity-weakening, necessary for unstable slip, has been shown for each bed type, but is much stronger and evolves over more than an order of magnitude longer distances for till beds. Utilizing a de-stiffened double direct shear apparatus, we found conditions for instability at freezing temperatures and high slip rates for both bed types. During stick–slip stress-drops, we recorded acoustic emissions with piezoelectric transducers frozen into the ice. The two populations of event waveforms appear visually similar and overlap in their statistical features. We implemented a suite of supervised machine learning algorithms to classify the bed type of recorded waveforms and spectra, with prediction accuracy between 65–80%. The Random Forest Classifier is interpretable, showing the importance of initial oscillation peaks and higher frequency energy. Till beds have generally higher friction and resulting stress-drops, with more impulsive first arrivals and more high frequency content compared to rock emissions, but rock beds can produce many till-like events. Seismic signatures could enhance interpretation of bed conditions and mechanics from subglacial seismicity.