講演情報
[S22-04]Stochastic determination of arrival time and initial polarity of seismic waveform
*庄 建倉1、裴 瑋来3、周 仕勇2 (1. 統計数理研究所、2. 北京大学、3. 中国石化石油勘探開発研究院)
We present a training-data–free probabilistic framework for automatic seismic phase picking and first-motion polarity determination that requires no prior information about the signal, station, or site. The method poses an eigen-equation linked to the probability distribution of the background-noise threshold; solving this yields posterior probability density functions (PDFs) for both arrival time and polarity, providing principled separation of signal from noise and quantified uncertainties for all decisions.
In addition to robust P-wave analysis, the approach can, under suitable signal-to-noise and path conditions, identify S-wave onsets. We evaluate performance on synthetic waveforms and on records from established seismic databases. In a local catalog test, automatically picked arrivals and polarities agree with analyst picks at 98% accuracy, demonstrating stability across diverse noise environments. Compared with leading neural-network–based pickers, our approach attains comparable reliability while offering superior portability and interpretability, as it avoids model training and site-specific calibration.
This work contributes (i) the introduction of a minimum-entropy criterion for data segmentation that explicitly accounts for multiple presumptions about background noise levels; (ii) joint inference of arrival time and polarity with full uncertainty quantification; (iii) a lightweight implementation suitable for deployment across heterogeneous networks; and (iv) a principled eigen-equation formulation for noise-aware phase picking. Collectively, these advances enable unified, reproducible, and easily maintainable catalog-scale phase picking and polarity assessment.
In addition to robust P-wave analysis, the approach can, under suitable signal-to-noise and path conditions, identify S-wave onsets. We evaluate performance on synthetic waveforms and on records from established seismic databases. In a local catalog test, automatically picked arrivals and polarities agree with analyst picks at 98% accuracy, demonstrating stability across diverse noise environments. Compared with leading neural-network–based pickers, our approach attains comparable reliability while offering superior portability and interpretability, as it avoids model training and site-specific calibration.
This work contributes (i) the introduction of a minimum-entropy criterion for data segmentation that explicitly accounts for multiple presumptions about background noise levels; (ii) joint inference of arrival time and polarity with full uncertainty quantification; (iii) a lightweight implementation suitable for deployment across heterogeneous networks; and (iv) a principled eigen-equation formulation for noise-aware phase picking. Collectively, these advances enable unified, reproducible, and easily maintainable catalog-scale phase picking and polarity assessment.
