10:00 AM - 10:15 AM
[SVC26-05] Classification of continuous seismograms during the eruption off Ioto in 2023 based on the wavelet scattering transform
Keywords:Ioto, isolated tremor, wavelet scattering transform
Ioto is a caldera volcano that continues to be active in seismic, tectonic, and geothermal activities and small eruptions have occurred repeatedly on the island and offshore. The analysis of continuous seismograms is important for the continuous monitoring of volcanic activities at Ioto. In this presentation, we report the results of our classification of continuous seismograms during the eruptive activity period that began on October 21, 2023, off the coast of Okinahama.
We used continuous seismograms on the vertical component at the Meganeiwa seismic station of NIED. The analysis period was from October 10, 2023 (before the eruption) to November 30, 2023 (during the eruption period). To reduce the computational cost, continuous seismograms were decimated from 100 Hz to 50 Hz, and then applied a 0.5 Hz high-pass filter. The continuous seismograms were then divided into 60-second-long segments, and a wavelet scattering transform was applied based on the method of Seydoux et al. (2020). The wavelet scattering transform is a method for computing wavelet scattering coefficients, which represent the time-frequency characteristics of time series data. This transform is achieved by performing a convolution and pooling process using multiple wavelet transforms. In this study, we used wavelets with center frequencies in the range of 0.78-25 Hz to calculate first- and second-order wavelet scattering coefficients. The wavelet scattering coefficient vectors obtained by concatenating all the wavelet scattering coefficients were used to cluster the 60-second waveforms. First, ICA was applied to the wavelet scattering coefficient vectors every 60 seconds to reduce the number of dimensions. In this study, the number of dimensions was fixed at 10. Hierarchical clustering was then used to perform unsupervised classification of the 60-second continuous seismograms.
We divided 60-second continuous seismograms into 15 clusters by hierarchical clustering. Clusters consisting of isolated tremors, volcanic earthquakes, signals caused by human activities, and other seismic ambient noises were identified, respectively. Moreover, the cluster with an increase in the number of detections from October 12 to 15 and November 19 was also obtained. Hierarchical clustering was then performed again within these 15 clusters, and clusters with an increase in the number of detections after October 21, when the eruption began, were selected. The number of detections of isolated tremors increased during the periods of increased eruptive activity (October 25-November 4, November 11-16, and November 20-22). The number of detections of volcanic earthquakes significantly increased on October 25 and November 14-25, but the relationship between eruptive activities and this volcanic earthquake activity is difficult to interpret. There were also clusters of increased detections during October 27-28 and November 21-23. The waveforms of these clusters were relatively similar to those of isolated tremors, however, higher frequency (~10 Hz) signals were also superimposed. The formation of land and the emergence of the vent above the sea surface may have excited these high-frequency signals. For the cluster with increased detections on October 12-15 and November 19, the relationship with eruptive activities needs further study because the timings of the increase in the number of detections are before the onset of eruptive activity/pause.
Acknowledgments: The author would like to thank NIED for providing continuous seismograms.
We used continuous seismograms on the vertical component at the Meganeiwa seismic station of NIED. The analysis period was from October 10, 2023 (before the eruption) to November 30, 2023 (during the eruption period). To reduce the computational cost, continuous seismograms were decimated from 100 Hz to 50 Hz, and then applied a 0.5 Hz high-pass filter. The continuous seismograms were then divided into 60-second-long segments, and a wavelet scattering transform was applied based on the method of Seydoux et al. (2020). The wavelet scattering transform is a method for computing wavelet scattering coefficients, which represent the time-frequency characteristics of time series data. This transform is achieved by performing a convolution and pooling process using multiple wavelet transforms. In this study, we used wavelets with center frequencies in the range of 0.78-25 Hz to calculate first- and second-order wavelet scattering coefficients. The wavelet scattering coefficient vectors obtained by concatenating all the wavelet scattering coefficients were used to cluster the 60-second waveforms. First, ICA was applied to the wavelet scattering coefficient vectors every 60 seconds to reduce the number of dimensions. In this study, the number of dimensions was fixed at 10. Hierarchical clustering was then used to perform unsupervised classification of the 60-second continuous seismograms.
We divided 60-second continuous seismograms into 15 clusters by hierarchical clustering. Clusters consisting of isolated tremors, volcanic earthquakes, signals caused by human activities, and other seismic ambient noises were identified, respectively. Moreover, the cluster with an increase in the number of detections from October 12 to 15 and November 19 was also obtained. Hierarchical clustering was then performed again within these 15 clusters, and clusters with an increase in the number of detections after October 21, when the eruption began, were selected. The number of detections of isolated tremors increased during the periods of increased eruptive activity (October 25-November 4, November 11-16, and November 20-22). The number of detections of volcanic earthquakes significantly increased on October 25 and November 14-25, but the relationship between eruptive activities and this volcanic earthquake activity is difficult to interpret. There were also clusters of increased detections during October 27-28 and November 21-23. The waveforms of these clusters were relatively similar to those of isolated tremors, however, higher frequency (~10 Hz) signals were also superimposed. The formation of land and the emergence of the vent above the sea surface may have excited these high-frequency signals. For the cluster with increased detections on October 12-15 and November 19, the relationship with eruptive activities needs further study because the timings of the increase in the number of detections are before the onset of eruptive activity/pause.
Acknowledgments: The author would like to thank NIED for providing continuous seismograms.