A U-Net for Weak Earthquake Detection

Abstract:

Earthquake detection is the basic step involved in cataloguing and subsequent seismic processing and imaging. The explosion in the quantity of seismic data requires building accurate and stable automatic event detection methods. Classical techniques, routinely used at several observatories, rely on the analysis of seismograms with high signal-to-noise ratios (SNR). Therefore, traditional automated detections methods are only able to analyze larger events, commonly about Mw 3.0. Furthermore, stable methods, such as the short-term average over long-term average (STA/LTA) method, are sensitive to noise and may miss weak events even with an adaptive threshold. To solve the problems of high noise levels and to avoid manually setting thresholds, we present a novel U-Net model to pick earthquake phase arrival times. Although many approaches have been proposed to automatically pick seismic phases, little attention was paid to small earthquakes (M<3) in the literature. In this study, different types of ambiguous picking circumstances are considered. For instance, long-duration P wave coda and the long time lag between P and S phases hamper a robust determination of S-wave picks. In our training data sets, we intentionally include a large portion of small earthquakes in western North America. The raw seismograms contain weak seismic signals that are difficult or impossible to identify visually. The supplement of weak events allows our network to self-learn unseen signal features from intricate seismic patterns. Furthermore, unlike previous U-Net models that only use three-component seismic traces, our model uses additional channels that correspond to band-pass and low-pass filtered horizontal traces. The trained U-Net creates a mapping from five input channels (three-component seismograms plus two filtered traces) to three-channel probability functions describing the current data point as a P arrival, S arrival, or non-arrival. We compare our proposed model with PhaseNet and AR picker from Obspy. The comparison results show our model is highly robust to small magnitude, complex source wavelet patterns, and low signal-to-noise ratio cases. Our prediction results match those in calibrated earthquae catalogues and demonstrate its high precision and generalization performance.