Analyzing continuous infrasound from Stromboli volcano, Italy using unsupervised machine learning

Infrasound data are used by scientists and monitoring observatories to track shifts in eruptive behavior, identify signs of unrest, and ultimately help forecast major eruptions. However, infrasound analyses are often limited to a catalog of discrete or high-amplitude transient events, which can leave lower-amplitude emergent or continuous signals within the datastream unexplored. This study classifies continuous volcano infrasound data using unsupervised learning in order to better constrain eruptive behavior through time. Data were recorded from 9 through 12 September (2018) at Stromboli, Italy by three infrasound arrays sampling at 200 Hz and deployed within 400 m of the active vents. Recorded pressure amplitudes were synthesized into a set of characteristic features extracted from the time and frequency domains of five second overlapping windows. Features were then clustered via the k-means algorithm resulting in a time-series of discrete labels that track the evolutionary behavior during the three-day experiment. Waveforms associated with each cluster relate to commonly recorded volcanic signals including Strombolian events, puffing activity, and sustained degassing. Infrasound radiated predominantly from six vent regions, each of which exhibit temporal variability in their degassing behavior. The three-day history of activity reveals an exchange of function across multiple vents indicating potential linkages in the plumbing system.