2:30 PM - 2:45 PM
[SVC28-10] Automatic Detection of First Arrival Time of Seismic Waves with the Fine-tuning: Applying to Observed Datas in Hachijojima
Keywords:deep learning, fine-tuning, first arrival time of seismic wave, automatic detection
It was reported that volcanic activity has been getting active near Hachijojima recently (Kimura et al.,2002). In order to evaluate a risk of eruption, it is important to analyze underground structure. Among various techniques for analyzing underground structure, we aim to find understructure by technique using natural earthquake, such as natural earthquake tomography. Hence, we need many and accurate first arrivals of seismic waves generated by natural earthquakes to improve the accuracy of underground structure analysis. However, the number of earthquakes near Hachijojima is small due to low the volcanic activity. Therefore, we try to increase the number of earthquakes to improve the accuracy underground structure analysis.
we focus on deep learning. The recent development of the deep learning is leading innovations in various fields. However, it is necessary for deep learning to prepare a large-scale dataset. It takes vast time to prepare the dataset. In order to solve the problem, we use the method called fine-tuning. The fine-tuning is a method to retrain some parameters of a trained model with a small amount of data and enables to extract the features of that small amount of data efficiently. We aim for the increase of the number of earthquakes by the applications of the method to detection of first arrivals of seismic waves. Moreover, we also aim to make this detection automatically for labor saving.
We installed seismometers to observe seismic waves in Hachijojima. Then, we applied Generalized Seismic Phase Detection with Deep learning (GPD) (Ross et al., 2018) to these waves. GPD is a Convolutional Neural Network (CNN) model that was developed by group of California Institute of Technology. This model can classify waveform datas as P-wave or S-wave, noise.
When we applied GPD to observed waveform datas in Hachijojima, there were a lot of false detections. We thought two factors caused the false detections that GPD classified the waveform data which was not seismic wave as P-wave or S-wave. One is that noise caused by wind and sea wave interferes with observed waveform datas. The other is that the compatibility between the datas used for training of GPD and observed waveform datas.
To solve these problems, we applied fine-tuning for two intentions. One is learning noise waveforms. It could make the number of false detections of GPD small. The other is solving the compatibility between the datas used training for GPD and observed waveform datas in Hachijojima.
We evaluated the effects of fine-tuning by comparing the results of fine-tuning with manual detections in terms of the number of detections and the accuracy of detection time. As a result, sometimes we could reduce false detections and succeed in highly accurate detections, sometimes we couldn’t.
we focus on deep learning. The recent development of the deep learning is leading innovations in various fields. However, it is necessary for deep learning to prepare a large-scale dataset. It takes vast time to prepare the dataset. In order to solve the problem, we use the method called fine-tuning. The fine-tuning is a method to retrain some parameters of a trained model with a small amount of data and enables to extract the features of that small amount of data efficiently. We aim for the increase of the number of earthquakes by the applications of the method to detection of first arrivals of seismic waves. Moreover, we also aim to make this detection automatically for labor saving.
We installed seismometers to observe seismic waves in Hachijojima. Then, we applied Generalized Seismic Phase Detection with Deep learning (GPD) (Ross et al., 2018) to these waves. GPD is a Convolutional Neural Network (CNN) model that was developed by group of California Institute of Technology. This model can classify waveform datas as P-wave or S-wave, noise.
When we applied GPD to observed waveform datas in Hachijojima, there were a lot of false detections. We thought two factors caused the false detections that GPD classified the waveform data which was not seismic wave as P-wave or S-wave. One is that noise caused by wind and sea wave interferes with observed waveform datas. The other is that the compatibility between the datas used for training of GPD and observed waveform datas.
To solve these problems, we applied fine-tuning for two intentions. One is learning noise waveforms. It could make the number of false detections of GPD small. The other is solving the compatibility between the datas used training for GPD and observed waveform datas in Hachijojima.
We evaluated the effects of fine-tuning by comparing the results of fine-tuning with manual detections in terms of the number of detections and the accuracy of detection time. As a result, sometimes we could reduce false detections and succeed in highly accurate detections, sometimes we couldn’t.