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

Abstract

It is critical to understand the dynamic conditions of Earth's cryosphere, yet the subglacial and englacial environments that control many aspects of ice behavior are inherently difficult to observe. The study of seismicity in glaciers and ice sheets has provided valuable insights about the cryosphere for decades, more recently aided by tools from machine learning. Here, we present an unsupervised machine-learning approach to discovering and interpreting cryoseismic patterns using 5 weeks of seismic data recorded at Gornergletscher, Switzerland. Our algorithm utilizes non-negative matrix factorization and hidden Markov modeling to reduce spectrograms into characteristic, low-dimensional “fingerprints,” which we reduce further using principal component analysis, then cluster using k-means clustering. We investigate the timing, locations, and statistical properties of the clusters in relation to temperature, GPS and lake-level measurements, and find that signals associated with lake flooding tend to occupy one cluster, whereas signals associated with afternoon and evening melt-water flow reside in others. We suggest that the one cluster contains signals that include the true initiation of the flood's englacial and subglacial drainage components. This work demonstrates an unsupervised machine-learning approach to exploring both continuous and event-based glacial seismic data.