[S-TT43] 最先端ベイズ統計学が拓く地震ビッグデータ解析

[STT43-P04] PoViT-UQ : P-wave Polarity and Arrival Time Determination using Vision Transformer with Uncertainty Quantification

*加藤 慎也¹、長尾 大道¹、飯尾 能久² (1.東京大学地震研究所、2.阿武山地震・防災サイエンスミュージアム)

キーワード：P波極性、発震機構解、深層学習

Determining earthquake focal mechanisms is important. Estimating them using P-wave polarities is often not robust (Hardebeck and Shearer, 2002). Therefore, automating this process requires high-confidence data. In recent years, deep learning-based polarity determination models have been developed (Ross et al., 2018; Hara et al., 2019; Uchide et al., 2020; Zhang et al., 2023). However, previous deep learning models are point estimation models and cannot evaluate confidence levels. In this study, we propose a novel deep learning model, PoViT-UQ, which combines a Vision Transformer (ViT; Dosovitskiy et al., 2021) with Monte Carlo Dropout (MCD; Gal and Ghahramani, 2016) to estimate high-precision initial P-wave polarity classification and arrival time detection with uncertainty quantification. Using seismic waveform data sampled at 100 Hz and 250 Hz, the model classifies polarities into three classes (Up, Down, and Noise) and simultaneously estimates P-wave arrival times. The results show a classification accuracy exceeding 98% and a standard deviation of 0.027 seconds in arrival time estimation with the 250 Hz model. By integrating MCD, we evaluate prediction uncertainty and apply an interquartile range (IQR) threshold of 0.15 or less to improve the accuracy of focal mechanism estimates. Validation using aftershock data from the 2016 Central Tottori Earthquake confirms that our approach contributes to efficient and high-precision focal mechanism estimates. Our model advances automated initial P-wave polarity determination and enables reliable data selection based on uncertainty quantification, significantly improving upon the limitations of conventional models.