Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2020 Aug 7;11(1):3952.
doi: 10.1038/s41467-020-17591-w.

Earthquake transformer-an attentive deep-learning model for simultaneous earthquake detection and phase picking

S Mostafa Mousavi  1 , William L Ellsworth  2 , Weiqiang Zhu  2 , Lindsay Y Chuang  3 , Gregory C Beroza  2
Affiliations

Earthquake transformer-an attentive deep-learning model for simultaneous earthquake detection and phase picking

S Mostafa Mousavi et al. Nat Commun. .

Abstract

Earthquake signal detection and seismic phase picking are challenging tasks in the processing of noisy data and the monitoring of microearthquakes. Here we present a global deep-learning model for simultaneous earthquake detection and phase picking. Performing these two related tasks in tandem improves model performance in each individual task by combining information in phases and in the full waveform of earthquake signals by using a hierarchical attention mechanism. We show that our model outperforms previous deep-learning and traditional phase-picking and detection algorithms. Applying our model to 5 weeks of continuous data recorded during 2000 Tottori earthquakes in Japan, we were able to detect and locate two times more earthquakes using only a portion (less than 1/3) of seismic stations. Our model picks P and S phases with precision close to manual picks by human analysts; however, its high efficiency and higher sensitivity can result in detecting and characterizing more and smaller events.

PubMed Disclaimer

Conflict of interest statement

The authors declare no competing interests.

Figures

Fig. 1
Fig. 1. Network architecture.
Our network architecture. Full details of each block are provided in the method section. The convolutional layers read as (number of kernels) kr (kernel size).
Fig. 2
Fig. 2. The training and test dataset.
Geographic distribution of station locations recording 300 k noise and 1 M earthquake seismograms in STanford EArthquake Dataset (STEAD) used in this study.
Fig. 3
Fig. 3. Attention weights.
Input waveform overlain by contextual information - output of the attention layers for a transformer (I in Fig 1), b local attention for P-phase (II in Fig 1), and c the local attention for S-phase (III in Fig 1).
Fig. 4
Fig. 4. Test results.
Four representative waveforms (a–d) from the test set and four waveforms from applying the model to continuous data in Tottori, Japan (e) and Redgcrest, California (f–h) presenting performance of the model on different types of data. Each waveform is 60 seconds long with 100 samples per second band-bass filtered from 1-45 Hz. Each panel shows three channel waveforms on top and output predictions of the model for earthquake signal detection, P-arrival, and S-arrival at the bottom. In a to d, the vertical color-coded lines on first two channels are manual arrival time picks from catalogs. a is an event with the local magnitude of 2.2 recorded in 55 km distance from the epicenter, b is an event with the local magnitude of 4.3 recorded 173 km away from the epicenter, c is an event with the local magnitude of 0.1 recorded 38 km away from the epicenter, and d is an event with the local magnitude of 2.0 recorded 110 km away from the epicenter. The output probabilities are presented as distributions that can represent variations or model uncertainties. e to h are detected events after applying the model to continuous data, representing the performance of the model when more than one event exists in a 1-minute window (e and f), when data contains gaps filled by zeros (e), when an event occurs near the edge (e), when two channels are broken or noisy (g), or when only one component data exists (h).
Fig. 5
Fig. 5. Tottori Earthquakes.
Seismicity of Tottori region between 6 October and 17 November 2000. a relocated events in Fukuyama et al. using manual phase picks by JMA. b relocated events using the automatic phase picker (EQTransformer) of this study. c Distribution of 57 seismic stations used by JMA and Fukuyama et al.d distribution of 18 stations used in our study to detect and locate earthquakes in Tottori region.
Fig. 6
Fig. 6. Distributions of frequency magnitude of earthquakes and picking errors.
a frequency-magnitude distributions of located events in JMA catalog and relocated events in our catalog (EQTransformer). Magnitudes for all events have been estimated using a local magnitude scale. b distributions of arrival-time differences (in second) between P (left) and S (right) picks by JMAʼs analysts and EQTransformer.
Fig. 7
Fig. 7. Phase picking performance as a function of noise level.
P (a) and S (b) phase picks as a function of signal-to-noise ratio (SNR) for three deep-learning and three traditional pickers.
See this image and copyright information in PMC

References

    1. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521:436. doi: 10.1038/nature14539. - DOI - PubMed
    1. Wang J, Xiao Z, Liu C, Zhao D, Yao Z. Deep-learning for picking seismic arrival times. J. Geophys. Res. 2019;124:6612–6624. doi: 10.1029/2019JB017536. - DOI
    1. Dokht RM, Kao H, Visser R, Smith B. Seismic event and phase detection using time-frequency representation and convolutional neural networks. Seismol. Res. Lett. 2019;90:481–490. doi: 10.1785/0220180308. - DOI
    1. Zhu L, et al. Deep learning for seismic phase detection and picking in the aftershock zone of 2008 Mw7. 9 Wenchuan Earthquake. Phys. Earth Planet. Interiors. 2019;293:106261. doi: 10.1016/j.pepi.2019.05.004. - DOI
    1. Zhou Y, Yue H, Kong Q, Zhou S. Hybrid event detection and phase-picking algorithm using convolutional and recurrent neural networks. Seismol. Res. Lett. 2019;90:1079–1087. doi: 10.1785/0220180319. - DOI

Publication types