5:15 PM - 7:15 PM
[SCG60-P03] Automatic detection of T-phase using deep learning
Keywords:T-phase, deep learning, Hi-net, submarine volcano
Automatic T-phase detection using deep learning
The T-phase consists of seismic waves generated by earthquakes and eruptions of submarine volcanos that propagate through a low-velocity layer called the SOFAR channel as underwater sound waves and are converted back into seismic waves on the coastal seabed. Therefore, the T-phase is often observed not only by ocean bottom seismometers but also by seismometers installed on land. In the October 2023 earthquake near Sofu Seamount, Izu Islands, P and S waves were indistinct, while T-phase was clearly recorded at several land-based stations. T-phase is not usually used for analysis when an earthquake occurs, but when P and S waves are unclear, as in the October 2023 earthquake, T-phase observation is important for estimating the epicenter and early tsunami warning. However, a system for immediate detection of T-phase has not yet been developed. In this study, as the first stage of system development, we tried to detect T-phase automatically by deep learning model using continuous waveform data of Hi-net. As a result, very high precision detection of T-phase became possible.
In this study, 250 Hi-net stations located along the Pacific coast of the Japanese archipelago were selected, and continuous waveform records of vertical components were used for analysis. To create the deep learning training data and test data, 44 events in which the T-phase was observed at selected Hi-net stations for earthquakes of Mw7 or higher occurring from 2006 to 2023, which are listed in the GCMT catalog, were extracted.
In deep learning, spectrogram images are trained. For 30 events that occurred from 2011 to 2023, spectrogram images of 10 minutes in length including the time when T-phase was observed at each station were created, and the pixel size was resized to 256px×256px. The spectrograms were created at a frequency range of 0.5 Hz to 20 Hz with 100 Hz sampling frequency, a window length of 5 seconds, and overlapped by 50%. Next, resized spectrogram images at three stations, i.e., (1) an arbitrary station, (2) a station closest to the station, and (3) a second closest to the station, were uniformly composited by taking the average value of RGB to create teacher data for deep learning. Similarly, by creating a composite spectrogram image including body waves and seismic noises, the teacher data in the case where the T-phase is not included was created. Finally, a total of 8,400 images were created, 2,800 images each for T-phase, body waves, and seismic noises. In this study, we created a convolutional neural network (CNN) model using Python's deep learning library Keras, and distributed 8,400 image data at the ratio of 75% for training and 25% for verification. In addition, as a test for unknown data, composite spectrogram images were prepared for 14 events that occurred from 2006 to 2010, and the accuracy of detection of T-phase was evaluated by applying the trained model.
As a result of training the model, it became clear that the accuracy was 98.6%, the precision was 98.3%, and the recall was 97.6% for the verification data, which was very high accuracy. In addition, when the trained model was applied to unknown data, the accuracy was 92.4%, the precision was 87.0%, and the recall was 93.2%. In the presentation, we will discuss automatic detection and source estimation of T-phase occurrence events based on automatic detection of T-phase.
Acknowledgments: This work was supported by JST SPRING, Grant Number JPMJSP2132. Also, continuous seismic records of Hi-net and GCMT catalogs were used for analysis, and Keras was used to build CNN models. We would like to thank all concerned.
The T-phase consists of seismic waves generated by earthquakes and eruptions of submarine volcanos that propagate through a low-velocity layer called the SOFAR channel as underwater sound waves and are converted back into seismic waves on the coastal seabed. Therefore, the T-phase is often observed not only by ocean bottom seismometers but also by seismometers installed on land. In the October 2023 earthquake near Sofu Seamount, Izu Islands, P and S waves were indistinct, while T-phase was clearly recorded at several land-based stations. T-phase is not usually used for analysis when an earthquake occurs, but when P and S waves are unclear, as in the October 2023 earthquake, T-phase observation is important for estimating the epicenter and early tsunami warning. However, a system for immediate detection of T-phase has not yet been developed. In this study, as the first stage of system development, we tried to detect T-phase automatically by deep learning model using continuous waveform data of Hi-net. As a result, very high precision detection of T-phase became possible.
In this study, 250 Hi-net stations located along the Pacific coast of the Japanese archipelago were selected, and continuous waveform records of vertical components were used for analysis. To create the deep learning training data and test data, 44 events in which the T-phase was observed at selected Hi-net stations for earthquakes of Mw7 or higher occurring from 2006 to 2023, which are listed in the GCMT catalog, were extracted.
In deep learning, spectrogram images are trained. For 30 events that occurred from 2011 to 2023, spectrogram images of 10 minutes in length including the time when T-phase was observed at each station were created, and the pixel size was resized to 256px×256px. The spectrograms were created at a frequency range of 0.5 Hz to 20 Hz with 100 Hz sampling frequency, a window length of 5 seconds, and overlapped by 50%. Next, resized spectrogram images at three stations, i.e., (1) an arbitrary station, (2) a station closest to the station, and (3) a second closest to the station, were uniformly composited by taking the average value of RGB to create teacher data for deep learning. Similarly, by creating a composite spectrogram image including body waves and seismic noises, the teacher data in the case where the T-phase is not included was created. Finally, a total of 8,400 images were created, 2,800 images each for T-phase, body waves, and seismic noises. In this study, we created a convolutional neural network (CNN) model using Python's deep learning library Keras, and distributed 8,400 image data at the ratio of 75% for training and 25% for verification. In addition, as a test for unknown data, composite spectrogram images were prepared for 14 events that occurred from 2006 to 2010, and the accuracy of detection of T-phase was evaluated by applying the trained model.
As a result of training the model, it became clear that the accuracy was 98.6%, the precision was 98.3%, and the recall was 97.6% for the verification data, which was very high accuracy. In addition, when the trained model was applied to unknown data, the accuracy was 92.4%, the precision was 87.0%, and the recall was 93.2%. In the presentation, we will discuss automatic detection and source estimation of T-phase occurrence events based on automatic detection of T-phase.
Acknowledgments: This work was supported by JST SPRING, Grant Number JPMJSP2132. Also, continuous seismic records of Hi-net and GCMT catalogs were used for analysis, and Keras was used to build CNN models. We would like to thank all concerned.