[3L1-OS-3a] OS-3

[3L1-OS-3a-02] Towards Addressing Challenges in Seismic Wave Arrival-Time Picking Models Using Deep Learning

〇Shinya KATOH¹, Hiromichi NAGAO¹, Masaaki Imaizumi² (1. Earthquake Research Institute,The University of Tokyo, 2. Department of Basic Science, Graduate School of Arts and Sciences, The University of Tokyo)

Keywords:seismic wave arrival time picking, deep learning

Deep learning models for seismic wave arrival time picking have been recognized for their high accuracy. However, a problem arises when multiple seismic waves exist within a single trace, making the detection of all seismic wave arrival times difficult (Park et al., 2023). This issue presents a significant barrier when adapting deep learning models to data characterized by frequent earthquakes in a short period, such as aftershock observations.

This study posits that the cause of this problem is label imbalance. The model commonly used (Zhu and Beroza, 2019) approaches the travel time picking task as a semantic segmentation task. The inputs are seismic waveforms, and the ground truth labels are shaped as Gaussian distributions with peaks for the arrival times of P and S waves and a height of 1, while all other points are labeled with a noise label of 1. Consequently, most points in the ground truth labels are noise. This results in a label imbalance, potentially hindering effective model training.

To address label imbalance, modifications to the loss function were implemented. As a result, the model could detect travel times for multiple seismic waves contained within the data.