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

[SCG50-P10] Toward a High-Performance Volcanic Earthquake Phase Detection: R2AU-Net Transfer Learning and Hyperparameter Analysis

*Yuji Nakamura¹, Yukino Yazaki¹, Yohei Yukutake², Yuki Abe³, Ahyi KIM¹ (1.Yokohama City University, 2.Earthquake Research Institute, the University of Tokyo, 3.Hot Springs Research Institute of Kanagawa Prefecture)

Keywords:Phase Picker, Volcanic Earthquakes, Swarms, Machine Learning, Deep Learning, Transfer Learning

In volcanic regions, earthquakes swarm related to volcanic activity frequently occur, and their rapid detection and measurement are crucial for volcanic disaster mitigation. However, currently, these processes ultimately rely on human judgment, requiring a vast amount of time and cost, making real-time detailed verification impossible. To solve this problem, in this study we construct a model that can more accurately detect earthquakes occurring in the Hakone volcano, using the R2AU-Net, which adds both recurrent residual units and attention mechanisms to the U-Net architecture. Performance evaluation and parameter tuning were conducted on seismic waveforms of about 30,000 events that occurred in the Hakone volcano from 1999 to 2020. As a result, the precision and recall of P-wave detection of the model trained with Hakone’s earthquake data were 95.7% and 87.6%, respectively, significantly improving compared to the existing model using the same Hakone data in this study. Next, recognizing the challenge of limited data in many volcanic areas, we further explored transfer learning for broader applicability. The model, initially trained with Hakone’s data (model HKN), was adapted for Kirishima Volcano, which has about 2,300 recorded events, 13 times fewer than Hakone. This adaptation involved fixing the encoder weights of model HKN and initializing the decoder weights for Kirishima’s data (model KSM), enabling effective transfer learning. The results of applying this model to the observation data of Kirishima, which was not used for training, showed that the performance of the model KSM significantly improved the performance of seismic phase detection compared to the case where the model HKN was applied as is, indicating that it is possible to construct a phase pick model by transfer learning even for volcanoes where training data is insufficient. Additionally, it is generally considered that in seismic motion detection for disaster prevention, models with higher recall scores are often more valued than those with higher precision scores. Upon conducting parameter tuning for both constructed models, it was found that under conditions with a higher learning rate for the model, the precision score for seismic wave detection slightly decreased while the recall significantly improved. This result indicates the possibility of constructing models with higher recall scores, more suited for volcanic disaster prevention.