[S-CG55] Driving Solid Earth Science through Machine Learning

[SCG55-P01] Towards Constructing a High-precision Seismic Phase picker for Volcanic Earthquakes Using Deep Learning: Performance Evaluation with Various Models and Examination of Transfer Learning

*Yuji Nakamura¹, Ahyi KIM¹, Yohei Yukutake², Yuki Abe³ (1.Yokohama City University, 2.Earthquake Research Institute, 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, active earthquake swarms often occur associated with volcanic activity, and their rapid detection and measurement are extremely important for volcanic disaster prevention. Currently, however, these processes are ultimately left to human judgment and require a great deal of time and money, and it makes detailed verification in real time impossible. To overcome this problem, we employed machine learning, which has been studied in many earthquake observation fields in recent years. Currently, there are already several state-of-the-art models available that have been trained using a vast amount of training data. Although there are some cases where these models can be applied as they are, regional dependence of the training data has also been reported (e.g. Münchmeyer et al., 2022). Since the target of this study is earthquakes that occur in a special region such as a volcanic area, existing trained models learned mainly from crustal earthquake waveforms may be difficult to apply.In this study, to develop a model that can more accurately detect earthquakes occurring in the Hakone volcanic area, as the first step, we employed the U-Net architecture (Ronneberger et al., 2018) to train and evaluate its performance using approximately 30,000 earthquake events from 1999 to 2020 recorded in the area. As a result, more than 1,000 more earthquakes were detected compared to PhaseNet (Zhu and Beroza, 2018). As the next step, we constructed models using R2U-Net, to which recurrent residual units were added, Attention U-Net, to which an attention mechanism was added, and R2AU-Net, to which both recurrent residual units and an attention mechanism were added to U-Net. The performance of the trained models was evaluated using the Hakone data used in the previous PhaseNet training. In addition, we applied transfer learning to the trained model by fixing the weights of the encoder of the first half of the model and initializing the weights of the decoder of the last half of the model. The model with the transfer learning was applied to earthquakes at Kirishima volcano, which has about 3,000 events, and the performance of seismic phase detection at Kirishima volcano, which has less data than Hakone volcano, was improved.