← Back to Gallery

[Un]supervised Clustering of [Non-]Earthquake Signals Commonly Recorded on Regional Seismic Networks

It is well known that surficial mass movements (SMMs), such as landslides and rock falls, have seismic signatures distinct from other routinely-recorded seismic sources like earthquakes and explosions. However, overlaps between the characteristics of these signals can still make it difficult to discriminate between source types during operational seismic monitoring. This ambiguity motivates the development of automated techniques for seismic signal classification. Furthermore, seismic differentiation within the broad class of SMMs has additional scientific and rapid response value. Examination of SMM seismic waveforms highlights a particularly strong contrast between the signals produced by primarily vertical processes — i.e., "falls" — versus processes that have a non-negligible horizontal component — e.g., landslides and avalanches, or "slides" in our abbreviated terminology.

Here, we present a machine learning (ML) classification scheme for differentiating between seismic signals generated by falls versus slides. We additionally include shallow earthquakes and blasts because these are most similar to SMM signals and are commonly recorded on regional seismic networks. These classes, therefore, are the most useful for automated classification. We neglect debris flows in this study because of their much longer durations and low amplitudes. Our signals derive from waveform picks in the Exotic Seismic Events Catalog, a diverse collection of non-earthquake seismogenic surface events, and the ANSS Comprehensive Earthquake Catalog. We use a shallow, feature-based approach to classification using statistical metrics extracted from waveforms. Feature importance metrics provide insight into the ML methods we use, which leverage both unsupervised techniques (to assess the diversity of seismic signals in our training dataset via clustering) and supervised techniques (to train and evaluate a classifier). We picture our classification workflow as one modular element of a larger non-earthquake assessment workflow that includes detection and location as preliminary (or concurrent) steps.

Session: Detecting, Characterizing and Monitoring Mass Movements [Poster Session]

Type: Poster

Room: Exhibit Hall

Date: 5/2/2024

Presentation Time: 08:00 AM (local time)

Presenting Author: Liam Toney

Student Presenter: No

Additional Authors