Clustering of records of short-period ocean-bottom seismometers

Recently, low-frequency tremors (LFT) off the Tohoku region have been observed by Seafloor observation network for earthquake and tsunamis along the Japan Trench (S-net). However, it is necessary to analyze the data of temporary observation using short-period ocean bottom seismometer (SP-OBS) to clarify the activities of LFTs before and after past great earthquakes including the 2011 Tohoku-Oki earthquake since the available S-net data periods is from 2016. In this study, we try to detect LFTs by applying a cluster analysis to the OBS waveforms to realize detection with no specific master records characterizing target LFT so that we can detect LFTs more comprehensively than previous studies.

In this study, we applied one of the unsupervised learning, K-means ++ (David and Vassilvitskii., 2007), to five SP-OBS records used in Takahashi et al. (2019), deployed in the northern part of the Japan Trench after the Tohoku-Oki earthquake. They indicated the LFTs have different shape of the power spectral density (PSD) between 1-8 Hz from those of ordinary earthquakes. Based on this knowledge, we used the PSDs at all the stations as the input data vector to the cluster analysis. Input data ensemble is composed of 570042 PSDs taken from continuous records for a half a year (Apr., 2011 to Oct. 2011).

As a result of clustering, the input PSDs were classified into four clusters. The optimum number of clusters was estimated by an elbow method based on the relation between the amount of scatter of data around the cluster center and the number of assumed clusters. It is confirmed that the six LFTs with high S/N ratio were all classified into the same cluster, indicating that the clustering of PSDs of SP-OBS seismograms can be applicable to identify unknow LFT events. However, we have to examine characteristics of the clusters more in detail. 35% of the input PSDs were classified into the cluster including the known LFTs but we suspect that there could be significant events different from LFTs, but having similarities to the known LFTs. Coda parts of ordinary earthquakes or teleseismic signals, which may have similar PSD shapes to LFTs, could be included in the cluster. For reliable identification of LFTs, it would be effective to apply additional cluster analysis to the data belonging to the same cluster with different characterization of input seismograms.