Caution: The prediction inconsistency of neural phase pickers should not be overlooked

Yongsoo Park, Gregory C. Beroza, & William L. Ellsworth

Submitted September 11, 2022, SCEC Contribution #11939, 2022 SCEC Annual Meeting Poster #004 (PDF)

Neural phase pickers – neural networks that are designed and trained to pick seismic phase arrivals – have shown to be a powerful tool for developing earthquake catalogs. However, these pickers suffer from what we refer to as prediction inconsistency where even a small perturbation in the input waveform data can change the output, sometimes substantially. As a result, a picker model can produce different sets of phase picks from the same data rather arbitrarily, depending on how we preprocess or window the continuous waveform data. This is not the case for conventional pickers such as STA/LTA. The problem has not been addressed in the literature and developers and users of these phase pickers have either been unaware of, or have overlooked it so far. In this presentation, we demonstrate how prediction inconsistency can affect the number of true positive and false positive phase picks that a picker model can produce and how that impacts the completeness and reliability of earthquake catalogs. We use the initial stages of the 2019 Ridgecrest, California earthquake sequence for demonstration and compare with published template matching catalogs. We also discuss potential strategies for mitigating the issue such as using a vote threshold and/or using multiple models for phase picking.

Key Words
Earthquake Monitoring, Machine Learning, Ridgecrest

Citation
Park, Y., Beroza, G. C., & Ellsworth, W. L. (2022, 09). Caution: The prediction inconsistency of neural phase pickers should not be overlooked. Poster Presentation at 2022 SCEC Annual Meeting.

Related Projects & Working Groups
Seismology