Real‐Time Earthquake Detection and Magnitude Estimation Using Vision Transformer. Issue 5 (5th May 2022)

• Record Type: Journal Article
• Authors: Saad, Omar M.; Chen, Yunfeng; Savvaidis, Alexandros; Fomel, Sergey; Chen, Yangkang
• Is Part Of: Journal of geophysical research. Volume 127:Issue 5(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-05-05
• Subjects: magnitude estimation -- vision transformer -- real‐time; Geomagnetism -- Periodicals; Geochemistry -- Periodicals; Geophysics -- Periodicals; Earth sciences -- Periodicals; 551.1
• DOI: 10.1029/2021JB023657
• Languages: English
• ISSNs: 2169-9313

Abstract

Abstract: We design a fully automated system for real‐time magnitude estimation based on a vision transformer (ViT) network. ViT is an attention mechanisms, which guides the proposed network to extract the significant features from the input seismic data, leading to robust magnitude estimation performance. We propose to design two separate ViT networks, that is, one for picking the P‐wave arrival time and the other for predicting the earthquake magnitude using a single station. For real‐time application, we pick the P‐wave arrival times and consider them as the reference, based on which the non‐normalized 30‐s (i.e., 1 s before and 29 s after the reference time) three‐component seismograms are used to predict the magnitudes of the corresponding earthquakes. The ViT picking network is first trained and tested using the STanford EArthquake Data set (STEAD) and shows robust picking performance, achieving an average picking error of less than 0.2 s compared to the manual picks. Then, the ViT magnitude estimation network is evaluated using several data sets, including those from California, STEAD repository, and Texas. The ViT demonstrates robust magnitude estimation performance in all these test cases as compared with the benchmark methods. For magnitude estimation, the mean absolute error (MAE) and the standard deviation error ( σ ) for the testing set of the STEAD data set are 0.112 and 0.164 (as compared with 0.141 and 0.219 for the state‐of‐the‐art MagNet method), respectively. The MAE and σ for the California testing set are 0.079 and 0.120 (as compared with 0.089 and 0.138 for the Magnet method), respectively. As a case study, the new ViT networks are applied to the 24‐hr continuous seismic data of the TexNet‐PB05 station recorded on September 20th. The network successfully picks all the events in the TexNet catalog with a small (<=0.42) magnitude error. The ViT network shows promising magnitude prediction results when tested with 4‐s long seismograms. This highlights its potential in the earthquake early warning (EEW) system for fast and reliable decisions. Plain Language Summary: Real‐time earthquake detection is fundamentally important to earthquake hazard mitigation and earthquake early warning (EEW). In addition to detecting a new earthquake, the EEW system also demands the real‐time magnitude estimation in order to decide whether a warning signal should be disseminated. In this work, we tackle the two tasks (detection and magnitude estimation) in one deep learning (DL) framework based on an advanced vision transformer architecture. The proposed DL framework obtains highly accurate results on a publicly accessible data set that is designed for benchmarking the performance of various seismological tasks. It is also applied in a real case study of a 24‐hr continuous waveform data set of a single station in Texas and shown to be robust in daily monitoring workflows. Key Points: An integrated deep learning framework for real‐time earthquake detection and magnitude estimation is developed The new framework leverages an advanced vision transformer that is better at feature extraction than the convolutional networks The proposed framework obtains accurate magnitude estimation results and detects 6 times more events than the analysts do for TexNet station