5:15 PM - 7:15 PM
[SCG60-P06] Detection of earthquakes around Kanagawa prefecture based on machine learning using waveform amplitude images of multiple stations
Keywords:machine learning, Hakone volcano, West Kanagawa earthquake
In recent years, machine learning technology has been progressed, and has been widely used in seismology (e.g., PhaseNet (Zhu and Beroza, 2019)). Many recent studies use machine learning in picking P-wave and S-wave arrivals such as PhaseNet and a phase association to match the picks across different observation stations (e.g., REAL (Zhang et al., 2019) and GaMMA (Zhu et al., 2022)). Furthermore, hypocenter relocation techniques such as HypoDD (Waldhauser and Ellsworth, 2002) are applied to construct high-quality earthquake catalogs (QuakeFlow, Zhu et al., 2023).
On the other hand, when small earthquakes such as slow earthquakes or volcanic earthquakes are manually detected, we generally check waveforms of multiple stations. The Hot Springs Research Institute of Kanagawa Prefecture has conducted manual detection of earthquakes using waveform images of multiple stations in addition to conventional earthquake detection using picks of P and S waves. They detected approximately four times of earthquakes determined by conventional earthquake detection (Itadera, 2023). Additionally, in the Owakudani area of Hakone volcano, earthquakes with unclear P and S wave arrival phases have been identified (Kurihara et al., 2025, in prep.), suggesting that a machine learning mthod using multiple stations as manual detection is effective.
On August 9, 2024, a magnitude 5.4 earthquake occurred in western Kanagawa Prefecture. Many aftershocks of the earthquake followed after the mainshock. In addition, some earthquakes also occurred in Hakone Volcano in the same day. It is important to distinguish the hypocenters in Hakone volcano or other region in order to understand tectonic or volcanic activity. Therefore, in this study I developed a machine learning model that classifies images of multiple stations waveform amplitudes into four categories: (1) earthquakes in Hakone volcano, (2) earthquakes around Kanagawa Prefecture except Hakone volcano, (3) teleseismic events, and (4) noise.
The input data consisted of RMS amplitude values every 0.5 seconds with a duration of 30 seconds. I prepared 4,000 data for each of the four categories. 12,000 data were used for training and 4,000 for validation. We developed two models: a simple convolutional neural network (CNN) and a Vision Transformer model.
As a result, both models record a high accuracy of approximately 96% on the validation data. Misclassifications mainly occurred due to mis-labeling of input data or earthquakes occurring near the boundary between the Hakone and other Kanagawa Prefecture area.
Next, we applied the Vision Transformer based model to continuous data, starting from January 1, 2024. During this period, a swarm of earthquakes was observed in Hakone volcano from April 29 to 30, and numerous aftershocks of the August 9 west Kanagawa Prefecture earthquake were also recorded. For continuous data analysis, I segmented the continuous data into 30-second time windows, shifting by 15 seconds. I classified each time window based on the probabilities assigned to categories (1) to (4) and defined detections as cases which the probability of an earthquake in Hakone or Kanagawa prefecture exceeded 90%.
The results of continuous detection showed that 1,195 earthquakes were detected in Hakone, while 5,344 earthquakes were detected in other areas of Kanagawa Prefecture. Few individual events were mis-detected when only later part of the coda wave was included in a time window. However, the overall detection performance was accurate. Notably, during the Hakone earthquake swarm and the aftershocks of the western Kanagawa earthquake, increases in detected earthquakes were observed, associated with the trend of manually detected earthquake catalogs. In the future, I plan to improve detection accuracy and add hypocenter classification by enhancing the model and training data.
On the other hand, when small earthquakes such as slow earthquakes or volcanic earthquakes are manually detected, we generally check waveforms of multiple stations. The Hot Springs Research Institute of Kanagawa Prefecture has conducted manual detection of earthquakes using waveform images of multiple stations in addition to conventional earthquake detection using picks of P and S waves. They detected approximately four times of earthquakes determined by conventional earthquake detection (Itadera, 2023). Additionally, in the Owakudani area of Hakone volcano, earthquakes with unclear P and S wave arrival phases have been identified (Kurihara et al., 2025, in prep.), suggesting that a machine learning mthod using multiple stations as manual detection is effective.
On August 9, 2024, a magnitude 5.4 earthquake occurred in western Kanagawa Prefecture. Many aftershocks of the earthquake followed after the mainshock. In addition, some earthquakes also occurred in Hakone Volcano in the same day. It is important to distinguish the hypocenters in Hakone volcano or other region in order to understand tectonic or volcanic activity. Therefore, in this study I developed a machine learning model that classifies images of multiple stations waveform amplitudes into four categories: (1) earthquakes in Hakone volcano, (2) earthquakes around Kanagawa Prefecture except Hakone volcano, (3) teleseismic events, and (4) noise.
The input data consisted of RMS amplitude values every 0.5 seconds with a duration of 30 seconds. I prepared 4,000 data for each of the four categories. 12,000 data were used for training and 4,000 for validation. We developed two models: a simple convolutional neural network (CNN) and a Vision Transformer model.
As a result, both models record a high accuracy of approximately 96% on the validation data. Misclassifications mainly occurred due to mis-labeling of input data or earthquakes occurring near the boundary between the Hakone and other Kanagawa Prefecture area.
Next, we applied the Vision Transformer based model to continuous data, starting from January 1, 2024. During this period, a swarm of earthquakes was observed in Hakone volcano from April 29 to 30, and numerous aftershocks of the August 9 west Kanagawa Prefecture earthquake were also recorded. For continuous data analysis, I segmented the continuous data into 30-second time windows, shifting by 15 seconds. I classified each time window based on the probabilities assigned to categories (1) to (4) and defined detections as cases which the probability of an earthquake in Hakone or Kanagawa prefecture exceeded 90%.
The results of continuous detection showed that 1,195 earthquakes were detected in Hakone, while 5,344 earthquakes were detected in other areas of Kanagawa Prefecture. Few individual events were mis-detected when only later part of the coda wave was included in a time window. However, the overall detection performance was accurate. Notably, during the Hakone earthquake swarm and the aftershocks of the western Kanagawa earthquake, increases in detected earthquakes were observed, associated with the trend of manually detected earthquake catalogs. In the future, I plan to improve detection accuracy and add hypocenter classification by enhancing the model and training data.