17:15 〜 19:15
[SCG60-P04] 通常地震と低周波地震の位相の分類に基づく気象庁一元化地震カタログの改善検討
キーワード：深層学習、低周波地震、PhaseNet
Japan Meteorological Agency is responsible for emergency tasks such as issuing earthquake information and tsunami warnings, and also constructs the unified earthquake catalogue by the Regional Earthquake Data Center system (REDC). The waveforms used to construct the earthquake catalogue are collected from relevant institutions. Using these waveforms, P- and S-phases are automatically detected, and the determination methods are applied such as the PF method (Tamatibuchi et al., 2016), the grid search method (Kiyomoto et al., 2013), and for deep low-frequency earthquakes and tremors (LFEs) that occur along the Nankai Trough, the Matched-Filter technique (Moriwaki et al., 2017). After careful inspection and manual correction by operators, the results are listed in the unified earthquake catalogue.
On the other hand, since the late 2010s, many technologies using the deep-learning have been developed worldwide in the field of observational seismology (Mousavi and Beroza, 2022). PhaseNet (Zhu and Beroza, 2018), one of the CNN models for picking the phase and the phase discrimination, output probability values of P-, S-phase and noise every 0.01 seconds by inputting seismic waveforms. Furthermore, by using a model in which PhaseNet is retrained with observed waveforms in Japan (Naoi et al., 2024) instead of the conventional picker of the PF method, we can detect more earthquakes compared to the result by the conventional picker although there is no way to distinguish between normal earthquakes and LFEs. In this study, we combine one or more machine learning models for phase detection that contribute to the automtic creation of the unified earthquake catalog, including the differentiation between normal earthquakes and LFEs, and verify its performance when used as an auto-picker for the PF method.
On the other hand, since the late 2010s, many technologies using the deep-learning have been developed worldwide in the field of observational seismology (Mousavi and Beroza, 2022). PhaseNet (Zhu and Beroza, 2018), one of the CNN models for picking the phase and the phase discrimination, output probability values of P-, S-phase and noise every 0.01 seconds by inputting seismic waveforms. Furthermore, by using a model in which PhaseNet is retrained with observed waveforms in Japan (Naoi et al., 2024) instead of the conventional picker of the PF method, we can detect more earthquakes compared to the result by the conventional picker although there is no way to distinguish between normal earthquakes and LFEs. In this study, we combine one or more machine learning models for phase detection that contribute to the automtic creation of the unified earthquake catalog, including the differentiation between normal earthquakes and LFEs, and verify its performance when used as an auto-picker for the PF method.