Active glacier processes from machine learning applied to seismic records

The great ice sheets of Antarctica evolve and respond to the changing global climate through a diverse set of active processes. Many of these deformational or hydrological processes are hidden from the view of satellite observations but give rise to a correspondingly diverse range of seismic signals. Seismology therefore provides a viable means of monitoring and studying remote, glaciated regions if the challenges of working with a heterogeneous population of signals can be addressed. A potential solution to the challenges of data-rich research or monitoring is semi-automated analysis, whereby manual time domain waveform appraisal is combined with unsupervised learning. Recent advances in the application of machine learning to seismic records suggest that machine learning applied to calculated waveform feature sets could be further developed for use in glaciology. In this thesis, I first assess how detection is performed in cryoseismology and diagnose the problems that need to be overcome. These are 1) the prevalence of weak amplitude signals that characterize a glacial environment, and 2) the variety of signals that result from distinct source types. As is typical in environmental seismology, the incoming motion from a cryogenic signal is expected at a lower signal-tonoise ratio than the P wave from a tectonic earthquake. Accordingly, it is difficult to apply algorithms that detect meaningful events over the background noise by using amplitude ratios or spectrograms. Other widely used algorithms are also unsuitable because they are based on a template that cannot generate the needed event detections for diverse waveforms. In response to the identified challenges, and for the purposes of systematizing analysis and building a database of events of various time scales and magnitudes, I develop and apply an algorithm that is tailored to the detection of cryoseismic events, the ‚ÄövÑv=multi-STA/LTA‚ÄövÑv¥ algorithm. I demonstrate the benefits of the new algorithm by applying it to continuous data from a seismic array deployed on the Whillans Ice Stream in West Antarctica over the period December 14, 2010 - January 31, 2011. Applying the multi-STA/LTA algorithm, I compile a catalogue of 1856 events that contains clearly discernible stick-slip impulses, signals that relate to teleseisms, and potentially diverse types of glacier processes. I compare the occurrence of events to large-scale influences such as tidal cycles impacting the Ross Ice Shelf and temperature variations during the seismic deployment that may affect surface and subsurface processes. I find that the occurrence of cryoseismic events correlates well with the tidal cycle indicating that the related motion of the Ross Ice Shelf is a probable causative influence. Temperature changes and the immediate seismic response of the Whillans Ice Stream and the surrounding region appear to be more weakly correlated, although cooling temperature through the months of available data may be a gradual influence. The source and variety of events in this catalogue are explored in a further study that employs a simple unsupervised machine learning method, k-means++ cluster analysis, to find patterns in events. The goal of such a method in cryoseismology is to improve understanding of the types of signals that emerge from a glacier. Event patterns typically include those that are evident from manual analysis and those that are not obvious in a standard dataset reconnaissance. I first compute a database of waveform, spectral, and polarization identifiers that decompose the events into numerical properties (features) that can characterize and separate one event from another. Then, using semi-automated clustering, I find commonalities among the catalogued events based on a division into ten clusters. The combination of manual review and clustering enables the following event groups to be identified: (1) Stick-slip events occurring near a central, seismogenic zone and (2) larger stick-slip events occurring near both the seismogenic zone and the grounding line, (3) event swarm potentially generated by a melt pulse from the Ross Ice Shelf, and (4) high energy events with a diurnal pattern, potentially generated as part of the external wavefield from fracture processes. Further noise signals that have distinct character likely correspond to other external wavefield processes outside the Whillans Ice Stream, such as ocean microseisms. The results suggest that future monitoring deployments could be partially automated to detect changes in movements related to slip, melt, and fracture. However, managing the complex seismic wavefield that originates from outside the ice stream or glacier itself will be a significant component in future analyses. The presented two-part workflow for cryoseismic studies, including new software components, is a response to the identified challenges in detection and subsequent data-driven analyses. Within the context of an exploratory work, I evaluate the needs that emerge when designing and implementing data-driven approaches, such as the tailoring of event detection parameters and finding a well-posed balance between manual and automated signal evaluation. I show systematically how to progress from a raw cryoseismic dataset to a robust and diverse event catalogue. Then, I evaluate an application of machine learning as a method for identifying patterns that can aid in detecting glacier processes. As a further result, this thesis illustrates reproducible, transparent analysis methods that could be used as benchmark methods for cryoseismic studies. These could be applied to the monitoring of a glacier from year to year, or for the comparative study of different glaciers.