← Back to Sessions

Making Phase-Picking Neural Networks More Consistent and Interpretable

Description: 

Improving the reliability and interpretability of phase-picking neural networks (herein referred to as NN-pickers) remain as important tasks to facilitate their deployment to seismic monitoring programs. Existing NN-pickers seek to estimate sample points in a given windowed waveform that is most likely to represent the arrival of a seismic phase by identifying the peaks in the output prediction score. Ideally, this score would scale with the confidence of the prediction result, so that applying a higher threshold would lead to fewer false detections. However, it has been reported that predictions from existing NN-pickers do not necessarily correlate with diagnostic waveform characteristics (e.g. the signal-to-noise ratio) and fluctuate rather arbitrarily depending on how the continuous waveform data is windowed. This lack of interpretability makes it challenging to control the quality of seismic monitoring products. Here, we present two approaches to increase the consistency of the existing NN-pickers and to improve their interpretability. First, we apply an anti-aliasing filter at down- and up-sampling layers inside neural networks, a popular technique applied within in the computer vision community to make convolutional neural networks more consistent. Second, we alter the training procedure such that shifted versions of the waveforms are systematically included in the training dataset. We demonstrate the improvements by applying the approaches to one of the most widely used NN-pickers and test on the waveform data that recorded the 2019 Ridgecrest earthquake sequence from July 4 to July 16. We show that using a batch containing sequentially shifted versions of a single waveform at each training iteration can significantly improve the consistency of the NN-picker we tested. We also show that with the help from an anti-aliasing filter, this training strategy can make the output prediction score scale with the signal-to-noise ratios of waveforms.

Session: Network Seismology: Recent Developments, Challenges and Lessons Learned - III

Type: Oral

Date: 5/2/2024

Presentation Time: 08:00 AM (local time)

Presenting Author: Yongsoo

Student Presenter: No

Invited Presentation: 

Authors