Precisely classifying earthquake types is crucial for elucidating the relationship between volcanic earthquakes and volcanic activity. However, traditional methods rely on subjective human judgment, which requires considerable time and effort. To improve this, we developed a deep learning model using a transformer encoder for a more objective and efficient classification. Tested on Mount Asama’s diverse seismic activity, our model achieved high F1 scores (0.876 for tectonic, 0.964 for low-frequency earthquakes, and 0.995 for noise), equivalent to or better than other methods. According to the attention weight visualization, our model focuses on critical seismic signal features for classification, similar to expert analysis. However, it has been demonstrated that removing subjective elements and employing standardized labeling of the training data based on waveform features are necessary to enhance the interpretability of the model. Additionally, the analyses suggest that stations near the volcanic crater are essential for a highly interpretative and accurate classification.

Volcanic earthquakes provide essential information for evaluating volcanic activity. As volcanic earthquakes are often characterized by swarm-like features, conventional methods using manual picking require much time in constructing seismic catalogs. In this study, using a machine learning framework and a trained model from a volcanic earthquake catalog, we obtained a detailed picture of volcanic earthquakes during the past 12 years at Kirishima volcano, southwestern Japan. We could detect earthquakes about 7.5 times larger than those in a conventional seismic catalog and obtain a high-resolution hypocenter distribution through waveform correlation analysis. Hypocenter clusters were estimated below the craters where magmatic or phreatic eruptions occurred in recent years. Increases in seismic activities, b-values, and low-frequency earthquakes were detected before the eruptions. The process can be carried out in real time, and monitoring volcanic earthquakes through machine learning contributes to understanding the changes in volcanic activity and improving eruption predictions.

An intense earthquake swarm is occurring in the crust of the northeastern Noto Peninsula, Japan. Fluid movement related to volcanic activity is often involved in earthquake swarms in the crust, but the last volcanic activity in this area occurred in the middle Miocene (15.6 Ma), and no volcanic activity has occurred since then. In this study, we investigated the cause of this earthquake swarm using spatiotemporal variation of earthquake hypocenters and seismic reflectors. Hypocenter relocation revealed that earthquakes moved from deep to shallow areas via many planes, similar to earthquake swarms in volcanic regions. The strongest M5.4 earthquake initiated near the migration front of the hypocenters. Moreover, it ruptured the seismic gap between the two different clusters. The initiation of this earthquake swarm occurred at a locally deep depth (z = ~17 km), and we found a distinctive S-wave reflector, suggesting a fluid source in the immediate vicinity. The local hypocenter distribution revealed a characteristic ring-like structure similar to the ring dike that forms just above the magma reservoir and is associated with caldera collapse and/or magma intrusion. These observations suggest that the current seismic activity was impacted by fluids related to ancient or present hidden magmatic activity, although no volcanic activity was reported. Significant crustal deformation was observed during this earthquake swarm, which may also be related to fluid movement and contribute to earthquake occurrences. A seismic gap zone in the center of the swarm region may represent an area with aseismic deformation.

We give a preliminary report on results of detecting low-frequency earthquakes (LFs) occurring at Mt. Fuji, Japan, using the matched filter method (MF method: e.g., Peng & Zhao, 2009). LFs have been observed in the depth 10-25km beneath Mt. Fuji (Hamada, 1981; Ukawa et al, 2005). These LFs seem to occur at an almost constant rate at all times, but it may become remarkably active as in the fall of 2000 (Yoshida et al., 2006). Because it is considered that the activity of LFs is associated with behavior of magmatic fluid at depth (e.g., Nakamichi et al., 2003), an investigation into the relationship between LFs and volcanic activity (e.g., Harada et al., 2010) is important. Understanding of the details of LFs activity is the first step for this investigation. Here, a system using the MF method for detecting LFs at Hakone volcano, Japan (Yukutake, 2017; Yukutake et al., 2019), was modified to be applicable to the detection of LFs at Mt. Fuji. Then, this was applied to continuous seismic record at seismic stations around Mt. Fuji during the period of 2012-2020. Next, the template waveforms of LFs were prepared on the basis of the earthquake catalog maintained by the Japan Meteorological Agency (JMA). Then, the cross-correlation analysis was conducted between the template waveforms and the seismic records. Finally, a catalog of LFs, obtained by using the MF method, was created. Using this catalog, we confirmed that LFs in 2012-2020 occurred at an almost constant rate, and that this is also true for LFs included in the JMA catalog. However, our case shows that LFs occurred at a rate of about 1,250 per year, which is about 10 times higher than that shown for the JMA case (a rate of about 125 per year). It was also confirmed that the larger LFs tend to have fewer numbers and smaller LFs tend to have more numbers, again a feature found by using the JMA catalog. Our research is underway, and tackling challenges such as selection of appropriate template waveforms of LFs, correction of magnitude estimate, and extension of analysis period will improve our results, which will be reported in the presentation.