10:45 〜 12:15
[SCG45-P29] Automatic classification of tectonic tremors with an unsupervised machine learning algorithm
キーワード：スロー地震、テクトニック微動、機械学習、自動分類
Continuous seismic records include various earthquake signals such as ordinary fast earthquakes and slow earthquakes like tectonic tremors. A machine-learning-based automatic classification approach would allow us to process a large amount of waveform data and to understand geophysical phenomena around a target seismic network. In this study, we propose an unsupervised automatic classification algorithm for continuous records based on frequency characteristics. The unsupervised approach has an advantage in that the algorithm does not require prior knowledge (e.g., template) of what kind of signals are included in the record.
Our proposed algorithm first extracts frequency features by calculating running spectra and converts them into 10-dimensional vectors with a filter bank. Then the vector quantization is performed in the feature space (every data point is represented by any one of 2000 representative points). After that, the representative points are converted by the kernel principal component analysis (kPCA; the kernel is a transition matrix generated assuming the Markov chain) and are clustered by the Ward hierarchical clustering algorithm in the space mapped by the kPCA. Finally, classification results are obtained by cutting the dendrogram at 1/4 of the maximum height.
We tested the proposed algorithm by applying to one-week-long continuous waveforms recorded at five temporary ocean-bottom seismometers to observe aftershocks of the 2004 M7.4 off the Kii Peninsula earthquake (Yamazaki et al., 2008; the record includes many shallow tectonic tremors in addition to aftershocks; Tamaribuchi et al., 2019). The classification was done for each station individually (therefore five independent classification results were obtained). For every station, tectonic tremors with large amplitudes were assigned for unique class(es) different from those for background noises and fast earthquakes, indicating that the proposed algorithm successfully detected tectonic tremors with a good S/N ratio based on the unsupervised approach. We also evaluated the detection rate of tremors by comparing the tremor catalog of Tamaribuchi et al. (2019), provided by manual inspection. We assumed the algorithm detected a tremor in the catalog when a tremor class appeared in a record of at least one station within 60 s after the catalog origin time. The detection rate was 87% for the entire records, which suggested that the proposed algorithm could detect tremors in the catalog with a high detection rate although the algorithm did not use specific knowledge of tectonic tremor such as template. However, at the same time, the algorithm also had a high false positive rate for the tremor detection. In some cases, tremors and coda waves of fast earthquakes were not separated completely because of similar frequency characteristics. Therefore, we will improve the selection of features fed into the hierarchical clustering process.
Acknowledgements: This study was supported partially by JSPS KAKENHI grant number JP21H05205.
Our proposed algorithm first extracts frequency features by calculating running spectra and converts them into 10-dimensional vectors with a filter bank. Then the vector quantization is performed in the feature space (every data point is represented by any one of 2000 representative points). After that, the representative points are converted by the kernel principal component analysis (kPCA; the kernel is a transition matrix generated assuming the Markov chain) and are clustered by the Ward hierarchical clustering algorithm in the space mapped by the kPCA. Finally, classification results are obtained by cutting the dendrogram at 1/4 of the maximum height.
We tested the proposed algorithm by applying to one-week-long continuous waveforms recorded at five temporary ocean-bottom seismometers to observe aftershocks of the 2004 M7.4 off the Kii Peninsula earthquake (Yamazaki et al., 2008; the record includes many shallow tectonic tremors in addition to aftershocks; Tamaribuchi et al., 2019). The classification was done for each station individually (therefore five independent classification results were obtained). For every station, tectonic tremors with large amplitudes were assigned for unique class(es) different from those for background noises and fast earthquakes, indicating that the proposed algorithm successfully detected tectonic tremors with a good S/N ratio based on the unsupervised approach. We also evaluated the detection rate of tremors by comparing the tremor catalog of Tamaribuchi et al. (2019), provided by manual inspection. We assumed the algorithm detected a tremor in the catalog when a tremor class appeared in a record of at least one station within 60 s after the catalog origin time. The detection rate was 87% for the entire records, which suggested that the proposed algorithm could detect tremors in the catalog with a high detection rate although the algorithm did not use specific knowledge of tectonic tremor such as template. However, at the same time, the algorithm also had a high false positive rate for the tremor detection. In some cases, tremors and coda waves of fast earthquakes were not separated completely because of similar frequency characteristics. Therefore, we will improve the selection of features fed into the hierarchical clustering process.
Acknowledgements: This study was supported partially by JSPS KAKENHI grant number JP21H05205.