← Back to Sessions

Medium Changes and Source Diversity Revealed by Unsupervised Machine Learning.

Description: 

The description of the evolution of seismic signals is most often relying on discontinuous catalogues of events, and by temporal evolution of elastic parameters deduced from extensive signal processing. We will present two examples where machine learning, namely a scattering network combined with independent component analysis, is used to monitor medium changes or the dynamic evolution of seismic activity from continuous records.

First, we investigated the signature of a thin centimetric layer of ground freezing at the surface, inducing known subtle changes in the medium. With ICA we reduce the dimension of the scattering network output and use a hierarchical clustering to identify clusters associated with freezing despite despite the extreme variability of urban-induced noise sources surrounding the station. We found that actually one the components of the ICA is precisely describing the temperature, meaning that the surface freezing is directly encoded in the seismograms and can be extracted automatically with our approach. This suggests that slight changes could be detected without relying on difficult measurements of physical parameters such as seismic speed temporal change.

Second, we studied volcanic tremors that cover a wide range of different signal characteristics. Despite their complex signal characteristics and their different source mechanism, volcanic tremors are either treated as one seismic signal class or as a set of seismic signal classes. We apply blind source separation methods and manifold learning techniques to continuous seismograms and reveal the underlying patterns in the time series data dominated by volcanic tremors. During a year period, the data-driven descriptors of the seismogram recorded at the active Klyuchevkoy volcano 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, revealing a non-stationary volcanic system.

Session: Opportunities and Challenges for Machine Learning Applications in Seismology

Type: Oral

Date: 4/19/2023

Presentation Time: 02:45 PM (local time)

Presenting Author: Michel Campillo

Student Presenter: No

Invited Presentation: No

Authors