Research

Volcanic earthquakes are classified into types — such as A-type events resembling ordinary earthquakes and B-type events dominated by low-frequency energy — and this classification provides key information for assessing volcanic activity. Traditionally it has depended on visual inspection by human experts, which is subjective and labor-intensive.

Using the diverse seismic activity of Mount Asama, we developed a deep-learning classification model based on a Transformer encoder. The model achieved F1 scores of 0.930–0.980 for the three classes (A-type, B-type, and noise), outperforming a convolutional-neural-network-based approach.

A distinctive contribution is interpretability: by visualizing the attention weights inside the model, we showed where in the waveform the "black box" focuses when making decisions. The model attends to key waveform features much like human experts do, while ambiguous labels in the training data were found to affect accuracy — highlighting the importance of data quality management. The framework can be transferred to other volcanoes, paving the way for enhanced volcanic hazard assessment.

In volcanic regions, active earthquake swarms accompany volcanic activity, and their rapid detection and analysis are crucial for volcano disaster prevention. Reading P- and S-wave arrival times from seismograms, however, has ultimately relied on human judgment — a time-consuming process that makes detailed real-time monitoring impractical.

In this study we trained a deep-learning phase picker (based on a U-Net architecture) using approximately 220,000 P- and S-wave onset readings accumulated at the Hakone volcano between 1999 and 2020. Compared with a publicly available model trained mainly on ordinary crustal earthquakes, the models trained or fine-tuned with the Hakone data achieved substantially higher detection rates.

We also addressed the problem of small events being missed when multiple earthquakes occur within a single detection window: by creating training data containing two earthquakes and retraining the model, we successfully detected events that would previously have been overlooked — a foundation for detailed real-time monitoring of swarm activity.