An Unsupervised-Machine-Learning Approach to Understanding Seismicity at an Alpine Glacier

Most machine-learning applications in seismology have focused on supervised-classification approaches. Here, we use an unsupervised-feature-extraction algorithm to reveal subtle spectral differences between microseismic icequakes at Gorner Glacier, Switzerland. Previous studies at Gorner Glacier have concluded that surface-crevasse formation is the main source of local seismicity, resulting in hundreds to thousands of icequakes per day. The highest seismicity rates occur in the afternoon, due to faster ice flow and deformation mediated by subglacial hydrology. We apply a machine-learning feature-extraction and clustering algorithm developed by Holtzman et al. (2018) to ~8000 icequake spectrograms recorded during June-July 2007. Based on methods from audio and speech analysis, the algorithm first applies nonnegative matrix factorization (NMF) to the spectrograms, followed by a hidden Markov model (HMM). From this, fingerprints are generated that emphasize the spectral differences between events rather than their shared features. The fingerprints are then clustered via a K-means clustering algorithm; similar results are attained for experiments allowing for 3-6 clusters. We examine independent geophysical data to identify possible physical causes for the clustering. No clear differences in event locations or simple signal statistics (e.g., peak amplitude, energy) between clusters are observed. However, the daily timing of events in different clusters varies: some display the expected afternoon peak in activity, while others show approximately uniform activity throughout the day. We hypothesize that the glacial hydraulic system may influence the spectral content of an icequake, and therefore how it is clustered. Additional auxiliary geophysical data (e.g., GPS, subglacial water pressure) will aid in understanding the causes of the newly identified subtle spectral differences between icequakes.