Development of a P-Wave Polarity Estimation Model Using Deep Learning with Consideration of Uncertainty

Shinya Katoh, Hiromichi Nagao, & Yoshihisa Iio

Published September 8, 2024, SCEC Contribution #13803, 2024 SCEC Annual Meeting Poster #035

Recent advancements in the application of deep learning in seismology have led to the development of deep learning models that can determine P-wave first-motion polarity with accuracy comparable to that of human experts (Ross et al., 2018; Hara et al., 2019; Uchide, 2020). The first-motion polarity of P-waves is a critical factor in estimating the earthquake focal mechanism. However, focal mechanism estimation using P-wave first-motion polarity is susceptible to the influence of misclassification. Particularly near the nodal planes of the fault, incorrect polarity determinations are known to significantly affect the accuracy of focal mechanism solutions. Therefore, improving the accuracy of polarity determination is essential for automatically estimating focal mechanisms.
In this study, we developed a method to not only improve the accuracy of polarity determination by deep learning models but also to evaluate the uncertainty of the model output. Traditional point estimation models have relied solely on threshold settings based on output probability values, providing insufficient evaluation of output uncertainty. In contrast, our approach introduces a method for quantitatively evaluating uncertainty in polarity determination models, demonstrating that focal mechanism accuracy can be improved even in situations where misclassification is problematic by utilizing only highly reliable data. This presentation will report on the details of this novel polarity determination model and the impact of uncertainty evaluation on focal mechanism estimation.

Key Words
Polarity, Deep Learning, Focal Mechanism

Citation
Katoh, S., Nagao, H., & Iio, Y. (2024, 09). Development of a P-Wave Polarity Estimation Model Using Deep Learning with Consideration of Uncertainty. Poster Presentation at 2024 SCEC Annual Meeting.


Related Projects & Working Groups
Seismology