[S-TT37] Seismic Big Data Analysis Based on the State-of-the-Art of Bayesian Statistics

[STT37-P05] Detection of deep low-frequency earthquakes based on matched filter technique using mutual information

*Ryo Kurihara¹, Aitaro Kato¹, Hiromichi Nagao^1,2, Sumito Kurata² (1.Earthquake Research Institute, The University of Tokyo, 2.Graduate School of Information Science and Technology, The University of Tokyo)

Keywords:Matched filter technique, Deep low-frequency earthquake, Mutual information

Matched filter technique is often used to detect microearthquakes such as deep low-frequency(DLF) earthquakes occurred in plate subduction zones or volcanic regions. Matched filter technique is usually based on the summed correlation coefficients between the template waveform and the observed data at multiple observation stations. However, the detection accuracy decreases when there are not enough observation stations or enough templates. Then, matched filter technique is able to be applied to only limited cases. In this study, we developed a new method of the matched filter technique using only one station data to evaluate the similarity of waveforms by calculating not only the correlation coefficient but also mutual information.

In the method, waveforms in a wide frequency band were used and the product of mutual information and correlation coefficient was used as an index for detection. Mutual information is an index which show relationships between two variables including non-linear relationship. Mutual information was calculated by normalizing the amplitude of the template waveform and the amplitude of the continuous waveform for all time steps in the time window. The normalized amplitudes were divided into a cell out of 5×5 cells.

First, we tested the new method on a dataset of artificial noise waveforms in which a waveform of DLF earthquakes is added to. As a result, it was confirmed that for both cases of white noise and sinusoidal noise of 1.25 Hz which corresponds to the dominant frequency of DLF earthquakes, the product of mutual information and correlation coefficient has a clear peak corresponding to the signal of the earthquake, compared to the case which only correlation coefficient is used.

Next, we used 200 template waveforms and applied the method to the data of Kirishima volcano in December 2010 before the eruptions of 2011. We detected 354 DLF earthquakes using the new matched filter technique using the product of mutual information and correlation coefficient at one station, while only two earthquakes were detected in the catalog of Japan Meteorological Agency in this term. Compared to the results of the matched filter technique using summed correlation coefficient of six stations, we missed only 8 events out of 88 events that had a high probability of DLF earthquakes, and 302 out of 354 events were events to be regarded as events with low probability of false detection.

In the case of using six stations, only small peaks which is not over a threshold value were obtained for some templates because the correlation coefficient changes due to slight differences in the source location and waveform characteristics of the templates. Therefore, it is necessary to use a large number of templates for the matched filter technique of multiple stations. On the other hand, in the case of one-station matched filter technique, even such template had a clear peak over the threshold value of detection.
In conclusion, it is considered that even when using data of only one station, it is possible to create an accurate catalog of DLF earthquakes by using the matched filter technique using the product of the correlation coefficient and the mutual information as the index of detection.