Automatic Phase Picking by Deep Learning

Date: 4/24/2019

Time: 06:00 PM

Room: Grand Ballroom

The phase picking is one of the most fundamental process for seismic data. In this decade, the matched filter method has been often used. Then, template waveforms are required. This means that the phase pick data is needed for the additional phase picking in the case of the matched filter analyses. It is critical, especially in the case of campaign observation, where there is no pick data at first.

This study constructs the automatic picking system by the deep neural network with the inception module. I introduced the dropout and batch normalization techniques for preventing the overfit.

I use seismograms from Hi-net for earthquakes in the northern Ibaraki prefecture and the Fukushima Hamadori areas, where the seismicity rate is high since the 2011 Tohoku-oki earthquake. The P and S arrivals are the manually picked ones in the JMA Unified Earthquake Catalog. I introduce the picking data as one-hot vectors.

Generally, the deeper neural network gives better results, but the trend is weak. This may suggest that we seismologists should focus more on how to use the seismic data, rather than the design of the neural network. My neural network gives P arrivals within 0.02 s from the JMA data for more than 70 % of seismograms. We empirically relate the output to the probability, which will give the accuracy of the arrival times given by the neural network.

Acknowledgement: We used the seismic data from Hi-net of NIED and the phase data from JMA. We also used the PyTorch library, and ABCI and AAIC supercomputers of AIST. This work is supported by Research Grants in the Natural Sciences by the Mitsubishi Foundation and AIST EDGE Runners project.

Presenting Author: Takahiko Uchide

Authors