Non-stationarity of volcanic tremor signals revealed by blind source separation and manifold learning

Abstract

Volcanic tremors are one of many seismic signals recorded on volcanoes and are associated with different pre- and co-eruptive processes. Therefore, they are widely used in volcano monitoring. The properties of the tremor signals such as duration, spectral content, or intermittency are very variable, reflecting the possible different tremor source mechanisms. In many cases, several tremor-generating processes can act simultaneously resulting in overlapping signals in the seismogram. Despite their complex signal characteristics and different source mechanisms, volcanic tremors are either treated as one seismic signal class or as a set of seismic signal classes. With a scattering network, we can access the information conveyed by volcanic tremors, even in the presence of short-term impulsive signals. We apply blind source separation methods and manifold learning techniques to continuous seismograms recorded at the Klyuchevskoy Volcanic Group (Kamchatka, Russia) and reveal the underlying patterns in the time series data dominated by volcanic tremors. The data-driven descriptors of the year-long seismogram reveal an ever-changing tremor signal, challenging the division of the observed volcanic tremors into a few distinct classes. The results highlight the complexity and non-stationarity of the volcanic tremors, suggesting a non-stationary volcanic system. Relating the data-driven patterns to the different underlying processes is the next step to understanding better the inner workings of a volcano.