Research ArticleSeismology

Convolutional neural network for earthquake detection and location

See allHide authors and affiliations

Science Advances  14 Feb 2018:
Vol. 4, no. 2, e1700578
DOI: 10.1126/sciadv.1700578
  • Fig. 1 Earthquakes and seismic station in the region of interest (near Guthrie, OK) from 14 February 2014 to 16 November 2016.

    GS.OK029 and GS.OK027 are the two stations that record continuously the ground motion velocity. The colored circles are the events in the training data set. Each event is labeled with its corresponding area. The thick black lines delimit the six areas. The black squares are the events in the test data set. Two events from the test set are highlighted because they do not belong to the same earthquake sequences and are nonrepeating events.

  • Fig. 2 ConvNetQuake architecture.

    The input is a waveform of 1000 samples on three channels. Each convolutional layer consists in 32 filters that downsample the data by a factor of 2 (see Eq. 1). After the eighth convolution, the features are flattened into a 1D vector of 128 features. A fully connected layer outputs the class scores (see Eq. 2).

  • Fig. 3 Probabilistic location map of two events.

    (A) The event is correctly located, and the maximum of the probability distribution corresponds to the area in which the event is located. (B) The event is not located correctly, and the maximum of the probability distribution corresponds to an area different from the true location of the event.

  • Fig. 4 Detection accuracy between ConvNetQuake and template matching.

    Percentage of the inserted events detected by the two methods in the synthetic data constructed by inserting an event template T2 unseen during training as a function of the SNR. The training set consists of 4-day-long seismograms (SNR ranging from −5 to 1 dB) containing 45 inserted event templates T1. The test set consists of 10-day-long seismograms (SNR ranging from −1 to 8 dB) containing 45 event templates T2.

  • Fig. 5 Event waveforms detected by ConvNetQuake that are similar to an event that occurred on 7 July 2014 at 16:29:11.

    (A) North component and (B) vertical component. Top: 479 waveforms organized by increasing absolute correlation coefficient and aligned to the S-wave arrival. Waveforms are flipped when anticorrelated with the reference event window. Bottom: Stack of the 479 events.

  • Fig. 6 Scaling properties of ConvNetQuake and other detection methods as a function of continuous data duration.

    (A) Runtime of the three methods where 1.5-hour one-time training is excluded for ConvNetQuake and where FAST’s runtimes include the feature extraction (38%) and database generation phases (11%). (B) Memory usage of FAST and ConvNetQuake.

  • Table 1 Performances of three detection methods, excluding the overhead runtimes (1.5 hours of offline training for ConvNetQuake and 47 min of feature extraction and database generation for FAST).

    Autocorrelation and FAST results are as reported by Yoon et al. (11). The computational runtimes are for the analysis of 1 week of continuous waveform data. NA, not applicable.

    AutocorrelationFASTConvNetQuake
    (this study)
    Precision100%88.1%94.8%
    Recall77.5%80.1%100%
    Event location
    accuracy
    NANA74.6%
    Reported runtime9 days, 13 hours48 min1 min, 1 s

Supplementary Materials

  • Supplementary material for this article is available at http://advances.sciencemag.org/cgi/content/full/4/2/e1700578/DC1

    section S1. Generalization ability of ConvNetQuake

    section S2. Autocorrelation for detection of new events during July 2014

    fig. S1. Waveform of an event recorded with station GS.OK029 and GS.OK027.

    fig. S2. Earthquakes in the region of interest (near Guthrie, OK, USA) from 14 February 2014 to 16 November 2016 partitioned into 50 clusters.

    fig. S3. Day-long synthetic seismic record constructed by inserting the waveform template shown in fig. S4A at random times into a time series of Gaussian noise.

    fig. S4. Templates used to generate synthetic data.

    fig. S5. Distribution of the correlation coefficients after autocorrelation of the windows classified as events by ConvNetQuake.

    fig. S6. Waveforms of the events detected in cluster 3 using a correlation coefficient threshold of 0.1.

    fig. S7. Waveforms of the events detected in cluster 3 using a correlation coefficient threshold of 0.2.

    fig. S8. Waveforms of the events detected in cluster 3 using a correlation coefficient threshold of 0.3.

  • Supplementary Materials

    This PDF file includes:

    • section S1. Generalization ability of ConvNetQuake
    • section S2. Autocorrelation for detection of new events during July 2014
    • fig. S1. Waveform of an event recorded with station GS.OK029 and GS.OK027.
    • fig. S2. Earthquakes in the region of interest (near Guthrie, OK, USA) from 14 February 2014 to 16 November 2016 partitioned into 50 clusters.
    • fig. S3. Day-long synthetic seismic record constructed by inserting the waveform template shown in fig. S4A at random times into a time series of Gaussian noise.
    • fig. S4. Templates used to generate synthetic data.
    • fig. S5. Distribution of the correlation coefficients after autocorrelation of the windows classified as events by ConvNetQuake.
    • fig. S6. Waveforms of the events detected in cluster 3 using a correlation coefficient threshold of 0.1.
    • fig. S7. Waveforms of the events detected in cluster 3 using a correlation coefficient threshold of 0.2.
    • fig. S8. Waveforms of the events detected in cluster 3 using a correlation coefficient threshold of 0.3.

    Download PDF

    Files in this Data Supplement:

Navigate This Article