Towards a continuous seismic wavefield scanning

Aurélie Guilhem, Doug S. Dreger and Bob Uhrhammer


The Mendocino Triple Junction (MTJ) region is tectonically very complex, and as a result it is the seismically most active region of Northern California, with earthquakes (including potential large earthquakes M$>$7) occurring on the Mendocino Transform Fault, along the Cascadia subduction zone, and in the offshore Gorda/Juan de Fuca plate. In addition to typical intra- and inter- plate seismic events, a variety of anomalous seismic events, including repeating earthquakes, slow/low-stress-drop earthquakes, and nonvolcanic tremors, are observed. In order to more effectively monitor the offshore region, we implement an automatic scanning of the continuous long-period ($>$10 sec) broadband seismic records to continuously detect, locate and determine moment tensors of events within minutes following their occurrence. Because the analysis will be done every two seconds, this scanning offers the possibility of rapidly identifying any large damaging and potentially tsunamigenic events.


We follow the algorithm proposed by Kawakatsu (1998) and used in Japan by Tsuruoka et al., 2009. A seismogram d recorded at a station k can be represented as the convolution of the response of the medium to an impulsive source G with the moment tensor elements m for a source s: $ \sum_{i}G_{i}^{sk}(t)m_{i}^{s}=d^{k}(t) $ (1). The normal equation based on (1) is given by: $ \sum_{i}A_{ji}^{s}m_{i}^{s}=b_{j}^{s} $ (2). A is a square matrix and the least-squares solution of the moment tensor $\widehat{m}$ is then given by : $\widehat{m}=A^{-1}b $ (3). The moment tensor m can be obtained by a simple matrix multiplication once b is calculated. The variance reduction (VR) obtained from the residual between data and synthetic seismograms is used to evaluate the fit for each calculation.

This analysis is performed over a grid search between 40.0$^\circ$ and 43.0$^\circ$ latitude (0.2$^\circ$ interval), -128$^\circ$ and -123$^\circ$ longitude (0.2$^\circ$ interval), and 5 to 38 km depth (3km interval) (Figure 2.20). For each point of the grid, a moment tensor is generated every two seconds, using 380 sec of broadband velocity seismograms filtered between 20 and 50 sec from four stations of the Berkeley Digital Seismic Network (BDSN) - HUMO, ORV, WDC, and YBH - and a previously generated catalog of Green's functions. The best calculated VR gives the preferred solution for an event.

Figure 2.20: Preliminary test for a $M_{w}$5.0 earthquake in the MTJ. The large dashed beach ball diagram represents the best mechanism (largest VR indicated on the map) determined for each window, and the large black one shows the reference solution from the Berkeley Moment Tensor catalog. The smaller beach ball diagrams represent the solutions obtained for each time period. The waveforms (vertical velocity seismograms) show the data used in the computation for each time window. A) Analysis performed for 380 sec starting 60 sec before the original time. B) Data starts at the origin time (ANSS catalog) of the earthquake. C) Best time given by the best VR computed for the grid search (here 6 sec after the ANSS catalog time). D) Window starts 60 sec after the origin time.
\begin{figure}\centering\epsfig{file=guilhem09_1.eps, width=8cm}\end{figure}

Preliminary results

We performed preliminary tests of the continuous seismic scanning using different magnitude earthquakes (from $M_{w}$4.2 to $M_{w}7.2$) located in the study region (Figure 2.20). We centered the grid search around the origin time of the events (about 90 sec before and 60 after), and we reduced the grid search to increase the computational timing of the calculations. For the five studied earthquakes we are able to retrieve their characteristics (origin time, location, magnitude, and mechanism) by searching for the maximum VR (between 69 and 80% agreement between data and synthetics). Additional tests have been performed for periods of seismic noise and teleseismic wave arrivals to evaluate the response of the designed technique when no event was generated in the study region.

Also, because the Mendocino region lies in the southernmost part of the Cascadia subduction zone, which is capable of great and potentially tsunamigenic earthquakes, we performed a synthetic test for a $M_{w}$8.2 reverse earthquake along the trench. We simulated synthetic waveform data for the four stations, filtered between 100 and 200 sec period. 480 sec were inverted, beginning 50 sec before the origin time and ending 50 sec afterward, and the best VR was determined for each point of the grid. Such a multi point-source inversion gives promising results for large earthquakes and can be employed in parallel to the single point-source analysis.

Implementation of the continuous scanning

The preliminary tests utilized archived seismic data rather than the realtime data that will be processed when the system is in service. We plan to use HH and HL channels (80 sps to 100 sps) from the four stations, decimated to 1 sps and filtered between 20-50 sec and 100-200 sec. Tests of the recursive filters, computed in only a fraction of a second for the 12 channels simultaneously, have been done, and they show very good agreement with similar filters computed in SAC. Also, because a realtime system might mean a lack of incoming data, we performed a jackknife test to measure the effects of missing channels on the stability of the moment tensor solution for a $M_{w}$5.0 earthquake (Figure 2.21). The mechanism and magnitude of an earthquake are stable. However the VR may change by tens of percentage points, which may result in the non detection of an event if the threshold is set at 65%. Also, missing two or three channels from different stations can worsen the result compared to missing one or two stations, especially if those missing channels contained most of the energy of the earthquake. As a result, we intend to remove all stations with missing channels from the computation and account for missing stations when computing the VR.

Figure 2.21: Test of the stability of the moment tensor with missing channel(s) and station(s). To symbolize the missing traces, we multiplied the data by zero. The beach ball diagrams on the right and left sides of the figure show the variation in the strikes. The central part of the figure shows the optimal solution obtained with the 12 working channels.
\begin{figure}\centering\epsfig{file=guilhem09_2.eps, width=8cm}\end{figure}


We are designing a continuous seismic waveform scanning system that will allow the detection, location, and determination of the magnitude and mechanism of any earthquakes of $M_{w}3.5$+ located in the vicinity of the MTJ. The combination of single point-source and multiple point-source inversions by the implementation of two parallel calculations (20-50 sec and 100-200 sec periods) will be useful for the detection of small to great earthquakes in the region. Such realtime monitoring will provide faster results than the currently used procedures. However, HUMO is currently in a triggered mode for the HH and HL channels. We will need to either obtain continuous datastreams or consider replacing it with another seismic station. Finally, the MTJ region is also known for the occurrence of unusual seismicity, in particular for slow/low-stress-drop earthquakes, and such a technique considering long period data may then help in the search for these unusual earthquakes.


Kawakatsu, H., On the realtime monitoring of the long-period seismic wavefield, Bull. Earthquake Res. Inst., 73, 267-274, 1998.

Tsuruoka, H., H. Kawakatsu, and T. Urabe, GRiD MT (grid-based real-time determination of moment tensors) monitoring the long-period seismic wavefield, Phys. Earth Plan. Int., 175, 8-16, 2009.

Berkeley Seismological Laboratory
215 McCone Hall, UC Berkeley, Berkeley, CA 94720-4760
Questions or comments? Send e-mail:
© 2007, The Regents of the University of California