[S-TT38] Seismic Big Data Analysis Based on the State-of-the-Art of Bayesian Statistics

[STT38-P03] Creating slow earthquake template catalogs with You Only Search Once algorithm

*Gerardo Manuel Mendo Perez¹, Hiromichi Nagao¹ (1.Earthquake Research Institute, University of Tokyo)

The occurrence of slow earthquakes in subduction zones has gained interest in recent years due to its potential connection to large megathrust earthquakes. Although these occur in a specific bandwidth (2 – 8 Hz), their relatively low amplitude makes them challenging to detect in inland seismic station records. Hence, identifying low-amplitude waveforms associated with slow earthquakes (shallow and deep) with high accuracy is still ongoing research. Since the pioneering envelope correlation method to locate tremors proposed by Obara (2002), many detection methods have been developed to identify slow earthquakes. One of the most used is matched-filter, or template-matching, a correlation-based technique that searches through seismic databases for events similar to template event catalogs, relying its success on the input template catalog. If the template catalog does not represent the seismic events associated with the phenomenon, its success rate is affected. For this reason, we propose You Only Search Once (YOSO), an alternative algorithm designed to identify seismic waveforms in inland seismic station records and create template catalogs that can be applied in matched-filter techniques. This procedure performs a one-time search through the database waveform using an RMS amplitude-based peak detection and classifies the detections into families of events using hierarchical clustering. The template matching technique uses the resultant catalog in both time and frequency domains. We applied YOSO to observed data from April 2016 of Hi-net stations in eastern Shikoku, Japan, where we successfully retrieved waveforms associated with LFE and tectonic tremor with time lengths below 200 s. Comparing different metrics and linkage algorithms for classification, we found that well-defined families of events are obtained using average linkage with cosine distance as the dissimilarity metric and the 75th percentile of the distance matrix as a threshold. We present the main findings of using our method and compare the catalogs obtained using YOSO and template-matching. We believe that this method can obtain reliable template catalogs from seismic databases at reasonable times and, hence, can be used for similarity analysis.