Understanding the uncertainties in phase pick labels using neural networks as analyst analogs

Yongsoo Park, Ting Chen, & Brent G. Delbridge

Published September 8, 2024, SCEC Contribution #13591, 2024 SCEC Annual Meeting Poster #013

Human labeled seismic phase arrival times have been directly used for training phase-picking neural networks. However, human errors and biases are naturally baked into these labels due to the subjectivity of the labeling task, and neural networks will inherit them when trained. Here, we exploit this logic and use neural networks as analogs of human analysts to understand the uncertainties in phase picks. We compile the waveforms and analyst labels provided by the Nevada Seismological Laboratory to build a parent dataset containing over a million P and S arrivals, each. We train 301 neural networks, each for P and S arrivals, using 100,000 samples sub-sampled from the parent dataset. With each trained neural network, we re-pick the phase arrivals from the remaining samples in the parent dataset that were not used for training. This gives 300 predictions per sample in the parent dataset. Our premise is that each trained neural network will inherit a different set of errors and biases and thus we will be simulating a scenario of hiring 300 human analysts to re-pick the phase arrivals. We present our findings on the distribution of the predicted phase picks, and the correlation between their uncertainties and attributes such as signal-to-noise ratio and event to station distance.

Key Words
Phase picking; Maching learning; Uncertainty

Citation
Park, Y., Chen, T., & Delbridge, B. G. (2024, 09). Understanding the uncertainties in phase pick labels using neural networks as analyst analogs. Poster Presentation at 2024 SCEC Annual Meeting.

Related Projects & Working Groups
Seismology