[S-TT38] Seismic Big Data Analysis Based on the State-of-the-Art of Bayesian Statistics

[STT38-P01] Two-stage approach for transfer learning of seismic-phase detection model to small sample size data

*Tokuda Tomoki¹, Hiromichi Nagao^1,2 (1.Earthquake Research Institute, The University of Tokyo, 2.Graduate School of Information Science and Technology, The University of Tokyo)

Keywords:Seismic detection, Deep learning, Transfer learning

Recently, a deep learning approach has gained much attention for seismic-phase detection. A deep learning model is trained using a large number of waveform data, which in turn enables to detect seismic-phase quite accurately. However, it is rare that such a large amount of data is readily available for a target region in question. To cope with this problem, Transfer Learning (TF) provides a useful framework to adapt the pre-trained model for the target region. Typically, a TF approach aims to re-train the deep learning model by reducing the number of parameters. Nonetheless, it still requires many data, which hinders its application where only dozens of seismic data is available. The objective of the present study is to address the problem of TF when available sample is rather small. To this end, we consider a two-stage approach. First, we apply the pre-trained model to waveform data in the target region. Second, for the estimated positive waveforms (e.g., P-phase), we classify them into the true and false positive. In this two-stage approach, the pre-trained (deep learning) model is directly applied to the waveform in question, and its (positive) phase detection result is further refined by a non-deep-learning classification method. The major contribution of the present study is to propose an effective classification method for the second stage. Our method is based on multiple clustering, which characterizes waveform focusing on specific features in a data-driven manner. This characterization achieves dimension reduction and non-linear classification of waveform, which contributes to effective classification. Using real data, we demonstrate effectiveness of our method compared with other supervised machine leaning method. In particular, when the true positive case is severely limited, it is shown that the proposed method excellently performs using the true positive case from other regions.