Classifying small earthquakes, explosions and collapses in the western United States using physics-based features and machine learning

• Published in: Geophysical journal international 2024-09, Vol.239 (2), p.1257-1270
• Main Authors: Koper, Keith D, Burlacu, Relu, Armstrong, Alysha D, Herrmann, Robert B
• Format: Article
• Language: English
• ISSN: 0956-540X; 1365-246X
• DOI: 10.1093/gji/ggae316

Classifying the source type of small seismic events is a key task in seismology. A common goal is distinguishing tectonic earthquakes from explosions and human induced seismicity. To this end, we applied a spectral modelling workflow to Pg and Sg waves from ∼10 000 seismic events that occurred in or near Utah and were recorded by broadband seismometers in the western United States at distances of 10–300 km. The events were a mixture of tectonic earthquakes (EQ), industrial explosions (EX) and mining-induced seismicity (MIS, primarily collapses) and were mostly small (median magnitude of 1.34 MC). Our spectral modelling was successful for 54 per cent of the events, resulting in a new catalogue of M0 and fc values. We evaluated 13 physics-based features—including differential magnitudes, Pg/Sg spectral amplitude ratios, long-period/short-period spectral amplitude ratios and spectral misfit—as source classifiers. We found that Φ ≡ log10(M0) + 3log10(fc) was the most effective individual feature for distinguishing EQ from EX and MIS sources because EQ spectra are relatively enriched in high frequencies. We selected five less correlated features that spanned the feature space and used a naïve Bayes approach to create a three-way classification model. The model had 97.5 per cent accuracy when applied to an independent test data set. Model performance deteriorated when more than six features were combined. We conclude that models developed with a few physics-based waveform features can classify small seismic events with performance comparable to high-dimensional deep-learning models. Simple models that rely on physics-based features require less training data and make more interpretable decisions than deep-learning models, though they may require higher signal-to-noise ratios.