[S-CG45] Science of slow-to-fast earthquakes

[SCG45-03] Locating tectonic tremors with uncertainty estimates: optimizing time- and amplitude-differences, wave propagation-based quality control, and Bayesian inversion

*悪原 岳¹、山下 裕亮²、杉岡 裕子³、篠原 雅尚¹ (1.東京大学地震研究所、2.京都大学防災研究所、3.神戸大学)

キーワード：テクトニック微動、ベイジアンインバージョン、南海トラフ沈み込み帯

Precise locations of tectonic tremors have posed implications for undergoing slow-slip events on the subducting plate interface and help clarify their source processes. Despite its importance, their location errors have rarely been evaluated in a statistically satisfactory manner. Also, locating methods must account for possible biases from unmodeled subsurface structures. This study presents three novel steps for locating tectonic tremors while addressing the above issues. In the first step, we optimize time- and amplitude difference measures between station pairs and obtain station-specific relative time and amplitude measures with uncertainty estimates. Then, using the optimized data, we roughly estimate propagation speed (i.e., shear wave velocity) and attenuation strength from time- and amplitude-distance relationships through linear regression. We apply this regression to each event and use the resulting velocity and attenuation strength as quality control factors to retain events with good-quality data. Finally, we formulate the tremor location problem within a Bayesian framework, where model parameters include source locations, local site delay/amplification factors, shear wave velocity, and attenuation strength. The Markov-chain Monte Carlo algorithm is used to sample the posterior probability, augmented by a parallel tempering scheme for an efficient global search. The proposed method is tested with ocean-bottom data that observed an intense episode of tectonic tremors in Kumano-nada, the Nankai Trough subduction zone. The results show that typical location errors (1σ) are ~1–2 km horizontally and <5 km vertically. A series of experiments with different inversion settings reveals the advantages of adopting amplitude data and site correction factors, which significantly reduce random error and systematic bias, respectively. Probabilistic sampling allows us to spatially map the probability that any tremor occurs at the place. From this probability map, lineaments of tremor sources, the second-order feature, are identified, as well as the first-order along-trench migration.