[S-CG50] Driving Solid Earth Science through Machine Learning

[SCG50-P08] Validation of Interpretability Enhancement in Volcanic Earthquake Classification Using Transformer Encoders

*Yugo Suzuki¹, Ahyi KIM¹, Masami Yamasaki¹, Yohei Yukutake² (1.Yokohama City University, 2.Earthquake Research Institute, University of Tokyo)

Keywords:Seismology, Volcanic Earthquake, Machine Learning, Transformer

In volcanic areas, active earthquake swarms frequently occur in association with volcanic activity. The number of these earthquakes increases as eruptions approach, and there are reported cases in which the number of low-frequency earthquakes in particular tends to increase, but the detailed mechanisms remain unknown. Accurate classification of volcanic earthquake types is crucial to clarify the relationship between these earthquakes and volcanic activities, but at present these processes are ultimately left to human judgement, which requires a great deal of time and human cost, and the classification criteria are also subjective. To address this, we developed a deep learning model using the encoder part of the transformer architecture to objectively and accurately classify various earthquake types. We tested the model on earthquakes from Mount Asama, where many low-frequency earthquakes of various types have been observed. The dataset included 6,634 A-Type (tectonic earthquakes) waveforms, 31,415 B-Type (low-frequency earthquakes) waveforms, and 11,205 noise waveforms. Then we allocated 80% of these data for training and the remaining 20% for validation. The results showed that precision of the A-type, B-type and noise was 0.88, 0.96 and 1.00 respectively, and recall was 0.87, 0.96 and 0.99 respectively, indicating that the performance was equal to or better than conventional machine learning models. Additionally, we visualized attention weights to enhance the interpretability of the model’s classification criteria, which are often opaque in deep learning. The analysis showed that the model, like human classifiers, focuses on the onset of P and S waves for A-Type, and the middle of the event signal following the P wave onset for B-Type. However, despite the high classification accuracy, some cases showed dispersed attention weights instead of concentration on P and S onsets. Further examination suggested that one of the main causes was the uncertainty in the labeling of the training data. Therefore, when we trained the model excluding waveforms that were labeled as B-Type but had clear onsets of P and S waves, the classification accuracy did not change significantly. However, the attention weights became more focused on the earthquake signal. Additionally, reducing the number of B-Type data randomly to match the number in the above experiment resulted in lower accuracy. This indicates that improving the model performance requires not only increasing the data but also scrutinizing it, such as by labeling according to uniform criteria.