Efficacy of automatic detection and classification methods for volcano seismicity

At volcanic islands, seismometers are the main tool for monitoring various active processes including seismicity that could forewarn impending eruptions. However, the sparse seismic networks, high background noise levels, large diversity of seismic signals, and high event rates during unrest episodes make automatic detection and classification of volcano seismicity a difficult challenge. In this paper, we use the Alaska Volcano Observatory’s catalogue of ~120,000 long-period (LP) and volcano-tectonic (VT) earthquakes at 34 volcanoes from 1989-2018 to evaluate the efficacy of automatic detection and classification methods. For each event, we calculate the frequency index (FI) based on the ratio of mean spectral amplitudes in the higher and lower frequency bands of the recorded waveforms. Using the local minima in the FI distribution at each volcano as the classification boundary, we find that our labels generally agree with the catalogue’s manual labels. The classification boundaries separating LP and VT earthquakes are also relatively consistent (FI = -1) between volcanoes. Therefore, the FI method is an effective method for automatic classification of volcano seismicity. We then evaluate the performance of two machine-learning-based models (PhaseNet and EQTransformer) and the cross-correlation-based template-matching method for automatic detection. While the template-matching method is computationally more expensive, we find that most of the catalogue events can be detected by using another event as a template, with relatively low false positive rates. In comparison, both machine-learning-based models’ performances are worse than previously reported results and deteriorate systematically with decreasing FI index values. The bias might have resulted from the models having been trained using earthquake catalogues from non-volcanic regions that lack LP events. Therefore, these models should be retrained with a dataset of volcano seismicity before being applied for automatic earthquake detection at volcanic regions.