Research ArticleEARTH SCIENCES

MyShake: A smartphone seismic network for earthquake early warning and beyond

+ See all authors and affiliations

Science Advances  12 Feb 2016:
Vol. 2, no. 2, e1501055
DOI: 10.1126/sciadv.1501055
  • Fig. 1 Noise floor of the phones.

    Noise floors of the smartphones color coded by the phone release date (also shown in the legend as MM/YY). Dashed black lines are typical ground motion amplitudes of earthquakes 10 km from the epicenter for various magnitudes. Noise floor for high-quality MEMS sensor (HP MEMS, blue) and a typical force-balance accelerometer from a regional network (BKS in northern California, purple) are also shown.

  • Fig. 2 3D shake table test.

    The input seismogram is from a real earthquake that has been modified for IEEE-693-2005 tests. (A) Waveform comparison between phone (blue) and reference accelerometer (red) recordings from an input signal that has peak acceleration of 0.5g. (B) Spectrum comparison of Y components. The X and Y components are in the plane of the phone, which is lying flat on the horizontal shake table and is not attached. The Z component is perpendicular to the plane of the phone and is vertical for this test.

  • Fig. 3 Shake table test with an input sweep signal (0.5 to 7 Hz).

    (A) Waveform comparison between a phone fixed on the table (blue), a phone placed freely on the table (black), and the reference accelerometer attached to the table (red). (B) Frequency domain comparison of the signals in (A). (C) Calculated correlation coefficient and RMS (root mean square) ratio between the signals recorded by the phone placed freely on the shake table and the reference accelerometer. The correlation coefficient is a measure of the phase match, and RMS is a measure for amplitudes match. We used a 1-Hz frequency band to filter the record and calculate the coefficient with a step frequency of 0.1 Hz. The x axis is the center frequency of the frequency band. The correlation coefficient shows how well the phase is recorded by the phone, and the RMS ratio shows the amplitude recovery. Above 2 to 3 Hz, the phone starts to slide, so the full amplitude is not recovered; however, the phase is recovered up to 7 to 8 Hz.

  • Fig. 4 Earthquake recorded by phone and classifying earthquakes.

    (A) Example of 12-hour three-component acceleration record from a private/personal Samsung Galaxy S4 phone starting at 4:00 p.m. (23 August 2014). It shows the accelerations of everyday human motions for the first ~8 hours, then appears stationary during the night. The red box at the end of the figure highlights the time window of (B). (B) One minute of data from the period shown in (A) at the time of the M6 Napa earthquake 38 km from the phone. The earthquake occurred at 3:20:44 a.m. local time. (C) Scaled feature plot showing IQR versus ZC for the classifier training data set. The blue dots are the centroids of human activities, and the red dots are the earthquake features. (D) 3D plot of the three features we used to distinguish earthquakes. Adding the CAV to IQR and ZC drags some of the human activates (blue dots) to the third dimension but not the earthquake data, which helps improve the results. EW, east-west; NS, north-south; UD, up-down.

  • Fig. 5 Estimated magnitude.

    Comparison of our estimated magnitudes with the real magnitude for earthquakes in Japan using phone-like data. The green line is the 1:1 line, and the two gray lines are the 1 magnitude unit shift. Each blue point is the magnitude estimate at a single simulated phone. The red pluses are the average event estimates, which is the average of multiple single phone estimates.

  • Fig. 6 Snapshots of trigger detections for the 2014 M5.1 La Habra earthquake simulation at 3, 5, and 7 s after the event origin time.

    Gray dots are stations, and pink indicates a trigger. The true earthquake (EQ) location is the red star with circles at 10-, 20-, and 30-km radius. The blue star represents the estimated event location first detected at 5 s. The magnitude estimate at each point in time is shown in the upper right.

  • Table 1 Performance of the ANN algorithm.

    Performance of classifiers when applied to earthquake and non-earthquake data not used to train the ANN algorithm. In the case of earthquake data, the percentage of records that were correctly classified as earthquakes is shown along with the number of records (in parentheses) for various earthquakes recorded within various distances of the epicenter. For the everyday human activity data, the percentage correctly identified as non-earthquake and falsely identified as earthquakes is shown.

    Earthquake classificationWithin 10 kmWithin 20 kmWithin 30 km
    1989 Loma Prieta M7100% (2/2)100% (4/4)100% (11/11)
    1994 Northridge M6.7100% (4/4)100% (19/19)100% (29/29)
    2004 Parkfield M695% (19/20)90% (35/39)86% (36/42)
    2014 Napa M6100% (2/2)75% (6/8)42% (10/24)
    2014 La Habra M5.1100% (13/13)42% (22/52)25% (30/120)
    Human activity classificationNon-earthquake (correct)Earthquake (false)
    20150201-2015022893% (3562/3823)7% (261/3823)

Supplementary Materials

  • Supplementary material for this article is available at http://advances.sciencemag.org/cgi/content/full/2/2/e1501055/DC1

    Data collection: The MyShake application

    Classifier analysis: Detecting earthquakes on a phone

    Network detection algorithm

    Estimate waring time for Katmandu, Nepal

    Fig. S1. MyShake activity 1 November 2014 to 28 February 2015.

    Fig. S2. Example earthquake record used to train the ANN classifier algorithm.

    Fig. S3. Structure of ANN classifier algorithm.

    Fig. S4. Receiver operating characteristic curve.

    Fig. S5. Phone trigger times versus epicentral distance.

    Table S1. Accuracy score for ANN classifier with 10-fold cross-validation.

    Table S2. Simulated network detection performance for U.S. earthquakes.

    Table S3. Simulated network performance for various phone densities.

    Movie S1. Network detection animation for the 2014 M5.1 La Habra earthquake.

    Movie S2. Network detection animation for the 2004 M6.0 Parkfield earthquake.

    Movie S3. Simulation of network detection with no earthquake.

    Movie S4. Simulation of network detection for an M6 earthquake.

    References (3439)

  • Supplementary Materials

    This PDF file includes:

    • Data collection: The MyShake application
    • Classifier analysis: Detecting earthquakes on a phone
    • Network detection algorithm
    • Estimate waring time for Katmandu, Nepal
    • Fig. S1. MyShake activity 1 November 2014 to 28 February 2015.
    • Fig. S2. Example earthquake record used to train the ANN classifier algorithm.
    • Fig. S3. Structure of ANN classifier algorithm.
    • Fig. S4. Receiver operating characteristic curve.
    • Fig. S5. Phone trigger times versus epicentral distance.
    • Table S1. Accuracy score for ANN classifier with 10-fold cross-validation.
    • Table S2. Simulated network detection performance for U.S. earthquakes.
    • Table S3. Simulated network performance for various phone densities.
    • Legends for movies S1 to S4
    • References (34–39)

    Download PDF

    Other Supplementary Material for this manuscript includes the following:

    • Movie S1 (.mp4 format). Network detection animation for the 2014 M5.1 La Habra earthquake.
    • Movie S2 (.mp4 format). Network detection animation for the 2004 M6.0 Parkfield earthquake.
    • Movie S3 (.mp4 format). Simulation of network detection with no earthquake.
    • Movie S4 (.mp4 format). Simulation of network detection for an M6 earthquake.

    Files in this Data Supplement: