Figure 2 Projections of all ray paths used in this study onto vertical planes showing the depth distribution of the ray paths. (a) Projection onto the north-south plane. (b) Projection onto the east-west plane.
Picking the first-arrival onset times from more than 300,000 seismic waveforms manually is a tedious task, which is both inefficient and prone to human error. Effective algorithms have been developed in recent years based on machine learning to pick the arrival times of seismic phases from a large quantity of seismic records (e.g., Ross et al., 2018; Zhu and Beroza, 2019). In this study, we use the deep learning model PickNet of Wang et al. (2019) for our first-arrival onset time picking purpose. The PickNet model has been trained with ~500,000 manually checked first-arrival times and its validity has been tested for various datasets.
For each record, we cut a 20-s-long P-wave window centered at the theoretical first-arrival time in the AK135 velocity model (Kennett et al., 1995). The windowed waveforms are processed to have zero mean and normalized by maximum values before they are fed to the PickNet algorithm. An example is shown in Figure 3 for the 10 February 2019 MW5.0 earthquake in the border region between Myanmar and India. Although all seismograms are processed by PickNet for automatic picking, we still visually checked all the picks to ensure their reliability. Some of the onsets in waveforms with lower SNRs tend to be unclear, but PickNet still yields the picks. Those picks are removed from further analysis, and we end up with a total of 43,540 reliable first-arrival time picks for the subsequent P-wave tomography inversion.