17:15 〜 19:15
[MTT36-P06] Preliminary study on event detection by machine learning and seismic activity around East Ongul Island, East Antarctica
キーワード：地震計アレイ観測、地震活動、南極、ディープラーニング
Seismic observations have been carried out around the coast of Lützow-Holm Bay, East ANtarctica, recording earthquakes that occurred in the Antarctic regeion, teleseismic earthquakes, and local events (Kanao and Kaminuma, 2006). In addition, icequakes and tremors also occur in the cryosphere; several studies reported that tides and moving icebergs produce such vibration phenomena in the area (Murayama et al., 2017; Tanaka et al., 2019). Monitoring icequake activity can be one of the indicators of environmental change because icequakes are also affected by environmental changes such as glacier flow and temperature changes on ice. However, the existing seismic network was insufficient for detecting small seimic events such as local icequakes and tremors due to its location sparsity. Hence, we carried out a special experiment using two seismic arrays in East Ongul Island during the summer of 2017-2018.
The two arrays consisted of seven three-component seismometers and six vertical seismometers, respectively, with a spacing of about 100m. First, we picked short-duration seismic events based on the short-term-average to long-term-average (STA/LTA) ratio from continuous waveforms with a high signal-to-noise ratio (SNR). After noise data were namually filtered out, we estimated the direction of arrival of these signals by applying semblance analysis (Neidel, N. S. and M. T. Taner, 1971). As a result, we found many events that propagated with high slowness throughout the observation period. Meanwhile, some of the low-slowness events were concentrated in the latter half. According to the satellite images of MODIS, the sea-ice condition around the Lützow-Holm Bay had been changing dramatically during the observation period. This seismic activity may reflect the difference in the condition.
In order to monitor seismic activity in the area where several types of events can occur, e.g., earthqakes, icequakes, and tremors, it is not enough to use the conventional detection technique described above. The method is vulnerable to being affected by the selection of parameters such as the length of the time window, threshold of signal-to-noise ratio, and number of stations for associating phases. Characteristics of triggered events also depend on these parameters. Relaxing the conditions can reduce missing events and event tyeps, but there is a tradeoff with false detection. Hence, we applied deep learning techniques to detect more events in the continuous waveform data. Using an unsupervised clustering approach over the wavelet power spectrum, we first categorized all the waveforms triggered by the STA/LTA method. Then, clusters were manually annotated and used to train an ordinal image recognition neural network. Finally, we used the trained network to classify the whole waveform again, aiming to detect more events. Some preliminary results will be presented.
Acknowledgements:
This work was supported (, inpart,) by ROIS-DS-JOINT (049RP2023, 040RP2024) to M. Hashimoto.
The two arrays consisted of seven three-component seismometers and six vertical seismometers, respectively, with a spacing of about 100m. First, we picked short-duration seismic events based on the short-term-average to long-term-average (STA/LTA) ratio from continuous waveforms with a high signal-to-noise ratio (SNR). After noise data were namually filtered out, we estimated the direction of arrival of these signals by applying semblance analysis (Neidel, N. S. and M. T. Taner, 1971). As a result, we found many events that propagated with high slowness throughout the observation period. Meanwhile, some of the low-slowness events were concentrated in the latter half. According to the satellite images of MODIS, the sea-ice condition around the Lützow-Holm Bay had been changing dramatically during the observation period. This seismic activity may reflect the difference in the condition.
In order to monitor seismic activity in the area where several types of events can occur, e.g., earthqakes, icequakes, and tremors, it is not enough to use the conventional detection technique described above. The method is vulnerable to being affected by the selection of parameters such as the length of the time window, threshold of signal-to-noise ratio, and number of stations for associating phases. Characteristics of triggered events also depend on these parameters. Relaxing the conditions can reduce missing events and event tyeps, but there is a tradeoff with false detection. Hence, we applied deep learning techniques to detect more events in the continuous waveform data. Using an unsupervised clustering approach over the wavelet power spectrum, we first categorized all the waveforms triggered by the STA/LTA method. Then, clusters were manually annotated and used to train an ordinal image recognition neural network. Finally, we used the trained network to classify the whole waveform again, aiming to detect more events. Some preliminary results will be presented.
Acknowledgements:
This work was supported (, inpart,) by ROIS-DS-JOINT (049RP2023, 040RP2024) to M. Hashimoto.