2025 Volume 78 Pages sp47-sp61
The accurate picking of seismic phase arrivals is crucial for hypocenter determination and subsurface imaging. However, the growing demand for seismic monitoring in CCS projects and the increasing adoption of DAS for high-density seismic observations have revealed the limitations of manual picking methods. This study aims to achieve accurate first arrival picking under limited training data conditions by evaluating the effectiveness of data augmentation, self-training, and fine-tuning techniques applied to deep learning models (PhaseNet and EQTransformer). Experiments using Hi-net data demonstrate that increasing the amount of training data generally improves accuracy. However, even when the training data was limited to one month, combining the proposed methods enabled the models to achieve F1 scores comparable to those obtained with one year of training. Among the methods, fine-tuning using pretrained models on the large-scale STEAD dataset yielded the most significant improvement, with additional gains from self-training and data augmentation. Furthermore, we validated the models using DAS data, which were acquired under observational conditions distinct from those of the seismic records used for pretraining. The results demonstrated that EQTransformer exhibited high adaptability through fine-tuning, whereas PhaseNet showed no improvement. This difference is attributable to the architectural characteristics of each model, highlighting the importance of selecting appropriate model architectures and training strategies based on the characteristics of the target data and intended application. The findings of this study provide guidelines for effectively leveraging deep learning under limited data conditions and are expected to contribute to the advancement of seismic monitoring and disaster mitigation efforts.
