11:00 AM - 1:00 PM
[SVC34-P08] Classification of volcanic earthquakes using a convolutional neural network
Keywords:Volcanic Earthquakes, Machine Learning, Convolutional Neural Network
Various types of earthquakes are observed around active volcanoes. These volcanic earthquakes are usually classified into several types based on waveform characteristics, and their respective activity often change with activity level (e.g., McNutt, 1986). Therefore, the development of a classification method based on objective and consistent criteria is important for understanding volcanic activity and evaluating the imminence of eruption. Recently, Nakano et al. (2019) developed a method to categorize seismic waveforms using a CNN (convolutional neural network) and running spectra of the waveforms. In this study, we use a similar CNN-based method to the seismic records at Miyakejima and Zao and evaluate its performance in the classification of volcanic earthquakes.
Miyakejima has erupted at intervals of about 20 years since 1940. It is thus important to investigate the recent volcanic activity to evaluate the imminence of the eruption. For this purpose, a temporal seismic observation network surrounding the summit crater was deployed in 2018, and it became easier to classify volcanic earthquakes using the detailed characteristics of their waveforms. Using the data of the temporal observation, Morita and Ohminato (2020) classified and located volcanic earthquakes and clarified the segregation of the source region of each type. They also suggested the relationship between hypocenter distribution and the resistivity structure. In this study, we use the data of the temporal observation as the training data for a CNN-based method and aimed to understand the recent activity of Miyakejima.
In this study, we select 241 A-type, 218 BH-type, 218 BL-type earthquakes, and train the network using these events recorded at two stations V.MJON (JMA) and N.MKAV (NIED). The CNN consists of two convolutional layers, a pooling layer, and a fully connected layer, and outputs the probability of each signal category. The final classification is evaluated using the product of the output probabilities for the two stations. As the input to the CNN, we use the running spectral images of 3-component data which cover 5.12 s in time and 0-25 Hz in the frequency direction. After training the CNN, we apply it to 2160 volcanic earthquakes from Jan. 2013 to Aug. 2016 and classify them into the types with the highest probability. The earthquakes classified into A-type (523 events) are characterized with distinct P and S waves, while those classified into BH-type (126) and BL-type (1481) tend to be classified according to the fraction of intermediate-frequency (5-15 Hz) components. To verify the accuracy and consistency of the classification, we train a CNN model using the data from Jan. 2013 to Jun. 2015, and apply the model to the data from Jul. 2015 to Aug. 2016 to derive three evaluation metrics (precision, recall, and F1-score). We then compute these evaluation metrics based on the classification results of this study (CNN classification) and the JMA catalog (JMA classification), respectively. As a result, we obtain 0.808, 0.828, and 0.817 for the CNN classification, and 0.649, 0.646, and 0.640 for the JMA classification. This result suggests that CNN is appropriate for evaluating the long-term activity based on more accurate and consistent classification.
In addition, it also becomes clear the fraction of BL-type events showing more high-frequency components near the initial motion have been increasing since around Apr.-May 2016. During this period, a sudden decrease in gas emission and inflation of a magma chamber were observed.
It is promising that further analyses based on a systematic classification using consistent criteria over a longer period contribute to clarifying the details of temporal change in volcanic activity and the conditions inside the volcano.
Acknowledgements: We used data from JMA, NIED, and ERI. This study was supported by the MEXT Integrated Program for Next Generation Volcano Research and Human Resource Development.
Miyakejima has erupted at intervals of about 20 years since 1940. It is thus important to investigate the recent volcanic activity to evaluate the imminence of the eruption. For this purpose, a temporal seismic observation network surrounding the summit crater was deployed in 2018, and it became easier to classify volcanic earthquakes using the detailed characteristics of their waveforms. Using the data of the temporal observation, Morita and Ohminato (2020) classified and located volcanic earthquakes and clarified the segregation of the source region of each type. They also suggested the relationship between hypocenter distribution and the resistivity structure. In this study, we use the data of the temporal observation as the training data for a CNN-based method and aimed to understand the recent activity of Miyakejima.
In this study, we select 241 A-type, 218 BH-type, 218 BL-type earthquakes, and train the network using these events recorded at two stations V.MJON (JMA) and N.MKAV (NIED). The CNN consists of two convolutional layers, a pooling layer, and a fully connected layer, and outputs the probability of each signal category. The final classification is evaluated using the product of the output probabilities for the two stations. As the input to the CNN, we use the running spectral images of 3-component data which cover 5.12 s in time and 0-25 Hz in the frequency direction. After training the CNN, we apply it to 2160 volcanic earthquakes from Jan. 2013 to Aug. 2016 and classify them into the types with the highest probability. The earthquakes classified into A-type (523 events) are characterized with distinct P and S waves, while those classified into BH-type (126) and BL-type (1481) tend to be classified according to the fraction of intermediate-frequency (5-15 Hz) components. To verify the accuracy and consistency of the classification, we train a CNN model using the data from Jan. 2013 to Jun. 2015, and apply the model to the data from Jul. 2015 to Aug. 2016 to derive three evaluation metrics (precision, recall, and F1-score). We then compute these evaluation metrics based on the classification results of this study (CNN classification) and the JMA catalog (JMA classification), respectively. As a result, we obtain 0.808, 0.828, and 0.817 for the CNN classification, and 0.649, 0.646, and 0.640 for the JMA classification. This result suggests that CNN is appropriate for evaluating the long-term activity based on more accurate and consistent classification.
In addition, it also becomes clear the fraction of BL-type events showing more high-frequency components near the initial motion have been increasing since around Apr.-May 2016. During this period, a sudden decrease in gas emission and inflation of a magma chamber were observed.
It is promising that further analyses based on a systematic classification using consistent criteria over a longer period contribute to clarifying the details of temporal change in volcanic activity and the conditions inside the volcano.
Acknowledgements: We used data from JMA, NIED, and ERI. This study was supported by the MEXT Integrated Program for Next Generation Volcano Research and Human Resource Development.