Geophysical Research Letters

Enhanced detection capability of non-volcanic tremor using a 3-level vertical seismic array network, VA-net, in southwest Japan

Authors


Abstract

[1] This study evaluates a method that enhances detection of non-volcanic low frequency tremor (NVT) using VA-net, a 3-level vertical seismic array network that was recently constructed in southwest Japan. The method employs semblance analysis for continuous vertical array records and calculates semblance coefficients as a function of time and apparent vertical velocity. The sign of the best apparent velocity provides information about propagation direction, either upward or downward, allowing easy discrimination between seismic signals of natural origin and unwanted noise such as man-made vibrations. We applied the method to a NVT episode that occurred beneath the Kii Peninsula during November 2008, to demonstrate that NVT signals can be detected using only the 3-level array. Furthermore, we demonstrate that the method enables us to detect minor NVT activity which is not detected by a conventional envelope cross-correlation method, resulting in a dramatic improvement in detection capability.

1. Introduction

[2] Non-volcanic, low frequency tremors (NVTs) and/or short-term, slow-slip events (SSEs) have been detected in various subduction zones such as the Nankai [Obara, 2002] and Cascadia [Rogers and Dragert, 2003] subduction zones and in the strike-slip San Andreas fault zone [Nadeau and Dolenc, 2005] during the last decade. Seismological and geodetic observations suggest that these slow earthquakes occur as shear slip events between tectonic plates within the quasi-stable frictional regime below the shallower seismogenic-locked zone [e.g., Dragert et al., 2001; Shelly et al., 2006]. Rogers and Dragert [2003] observed that NVTs coincided temporally and spatially with SSEs in the Cascadia subduction zone, implying that they are closely linked. The same phenomenon has also been identified in other regions including the Nankai subduction zone [Hirose and Obara, 2010]. Ide et al. [2008] interpreted these phenomena as different manifestations of a single process, observed seismically as NVTs (or isolated events up to 200 s) and geodetically as SSEs, and suggested that NVTs could be used as a proxy for SSEs. In fact, Hiramatsu et al. [2008] found that the cumulative seismic moment estimated for NVTs agrees well with that estimated from geodetic observations and showed that NVTs are useful to monitor slip on the plate interface. Aguiar et al. [2009] and Obara [2010] demonstrated that the activity of NVTs scales directly with moment release of SSEs estimated from geodetic observations, providing a new way to assess small SSEs that cannot be detected geodetically. These small-sized SSEs would also cause stress to accumulate on the shallower locked portion of the fault and increase the likelihood of future large earthquakes. Therefore, detailed real-time monitoring of NVTs, including minor tremor activity, is of great importance for forecasting the occurrence of the next large earthquake.

[3] NVTs are characterized by long-duration signals with no clear P- or S-wave arrivals and low-frequency vibrations with predominant frequency content between 1 and 10 Hz [Obara, 2002]. These characteristics make it hard to apply conventional hypocenter determination methods using P- and S-phase picks to NVTs. The most widely used method to detect and locate NVTs is a signal envelope cross-correlation method (ECM) developed by Obara [2002]. The ECM is based on the smoothed envelope pattern obtained at network stations generally separated by a few tens of km, thus the spatial and temporal resolution of this method as well as the detection level of NVTs is limited. Dense surface seismic array observations have been shown to be effective in investigating the complex wave fields of NVTs and resolving their fine-scale spatio-temporal evolution [e.g., Ghosh et al., 2009; McCausland et al., 2010]. Those studies also suggest that the array analysis is able to improve the detection capability of NVTs. For example, Ghosh et al. [2009] deployed a temporary 84-element small-aperture seismic array in the Cascadia subduction zone and showed that the array analysis detected NVTs for a time span about four times longer in duration than that by the ECM.

[4] In 2007, the Geological Survey of Japan, AIST (GSJ) started integrated borehole observation of subsurface water level, water temperature, crustal strain, tilt, and seismic waves in southwest Japan to monitor anticipated earthquakes on the Tokai, Tonankai, and Nankai megathrusts [Itaba et al., 2010]. At each observatory, we have drilled three boreholes to different depths (approximately 30 m, 200 m, and 600 m) and installed three-component, high-sensitivity seismometers at the bottom of each borehole. Data acquisition is continuous at either 100 or 200 samples per second (sps). Fourteen observatories are in operation as of August 2011 (inset of Figure 1a). This vertical seismic array network (VA-net) will also provide significant opportunities to enhance our understanding of the nature of slow earthquakes. In this study, we analyze vertical array data collected during the November 2008 NVT episode and show a dramatic improvement in NVT detection capability using only the 3-level array network.

Figure 1.

(a) Epicenter distribution of the NVT episode in November 2008 determined by the ECM. Station distributions are shown by squares and triangles. Time is color-coded to illustrate temporal evolution of tremor activity. As a reference, epicenters of low frequency earthquakes determined by JMA during the same period are shown by black Xs. The inset at the lower-right shows the tectonic setting around Japan and the distribution of the vertical seismic array network of Geological Survey of Japan, AIST (black squares). (b) Space-time plot of the NVT episode along the profile X-Y in Figure 1a. Note that station ANO was not in operation during this time period.

2. Data

[5] This study focuses on the NVT episode beneath the Kii Peninsula that occurred in November 2008 and lasted approximately two weeks, one of the longest episodes in this region. We first determined NVT locations using the ECM, by applying a band-pass filter of 2 to 10 Hz for a given 1-minute-long time segment and creating the envelope function from two horizontal-component waveforms with a smoothing time of 1 s. Data for the ECM analysis came from the regional high-sensitivity borehole seismic stations of NIED (National Research Institute for Earth Science and Disaster Prevention) and surface stations of JMA (Japan Meteorological Agency) and ERI (Earthquake Research Institute, the University of Tokyo), together with one level of the vertical array operated by GSJ (Figure 1a). Figure 1a shows NVT locations determined by the present study, where time is color-coded to illustrate temporal evolution of the NVT activity. A time-space plot is shown in Figure 1b. This episode primarily consists of two clusters of NVT activity. One took place between November 3 and November 5 beneath the station HGM. The second began on November 10 near the station ITA and lasted until November 18, migrating to the northeast and the southwest. The migration velocity is on the order of 10 km/day, a typical rate along the strike of subduction zones [e.g., Obara, 2002].

[6] Figure 2 shows a one-hour-long section recorded at the three borehole depths of station HGM (Hole 1 is the deepest hole) during the period when NVT activity was just beneath the station. The signal amplitude of the shallowest hole (Hole 3) is the largest, which is probably due to a near-surface amplification. Note that many tremor-like signals are visible, although the ECM was able to detect only five NVTs (five minutes cumulative) during the one-hour period. Portions plotted in blue correspond to times of detected NVT with epicentral distances of less than 50 km. In the next section, we show that the NVT activity occurred continuously during the one-hour period on the basis of vertical array analysis.

Figure 2.

Band-passed (4–8 Hz) signals at station HGM between 01:00 and 02:00 local time (LT) on 4 November 2008. Portions plotted in blue correspond to times of NVT signals detected by the ECM, whose epicentral distances are less than 50 km. This study detected many additional NVT signals, which are plotted in red. Hole 1, 2, and 3 are the deepest, middle, and shallowest boreholes, respectively.

3. Vertical Seismic Array Detection (VSAD) of Non-volcanic Tremors

[7] The semblance algorithm has a capability to detect coherent waves even if the signal-to-noise ratio is low [Neidell and Taner, 1971]. The semblance coefficient for a vertical array is defined by

equation image

where vapp is an apparent vertical velocity among stations, τ corresponds to the central time position in which the semblance coefficient is calculated, N is the number of stations (3 in the present case), W is a length of time window, and fi(t) represents a waveform of i-th level of the array at a time t. Di is a distance at i-th level of the array relative to the deepest one. The definition of Di suggests that positive and negative apparent velocities exhibit upward and downward incidence signals, respectively. On the basis of the sign of the best apparent velocity for that particular window, therefore, we can easily discriminate seismic signals of natural origin from unwanted noise such as man-made, wind- and rain-induced vibrations. The value of apparent velocity depends on the incidence angle of signals to the vertical array: vapp = v/cosθ, where v is a material velocity and θ is the incoming angle of incidence from vertical. If the signals are vertically upward incidence to the array, for example, then vapp = v. We calculated a semblance coefficient for a range of values in apparent velocity space using a one-minute-long moving window with 50% overlap. Each seismogram was filtered in the 4–8 Hz frequency band. We experimented with shorter windows down to 5 s and other selected frequency bands between 2 and 50 Hz. We confirmed that the general characteristics remain consistent, when we use a shorter window length and a higher frequency band up to ∼20 Hz.

[8] Figure 3a shows a result of semblance analysis for the one-hour-long section in Figure 2. For horizontal components, high semblance coefficients continuously appear at an apparent velocity of about 2 km/s in agreement with the S-wave speed estimated from the sonic log. The consistency between the best apparent velocity and the S-wave speed indicates that we observed vertically incident S-waves of NVT signals in horizontal components. This is not unexpected, because the NVT activity during this time period occurred directly beneath the station HGM (Figure 1a). A continuation of high semblance coefficients suggests that the NVT signals were continuously detected by the vertical array, which is uncertain in the case of the ECM. The best apparent velocity for the vertical component (3–5 km/s) is clearly larger than that for the horizontal components and almost equal to the P-wave velocity estimated from the sonic log. This observation indicates that the vertically incident P-wave came directly from NVT sources, although we cannot rule out the possibility that these P-waves were produced by SV- to P-wave conversions. Because this topic goes beyond the scope of this paper, we leave it for future research. Compared with horizontal components, the resolution of apparent velocity is poor and the semblance coefficient is small in the vertical component. This seems to be caused mainly by an insufficient sampling interval (100 sps) for P-wave velocity. It may also originate partially in the small radiation energy of P-waves relative to S-waves and partially in a mixture of P- and S-waves in the vertical component. We notice that slightly high semblance coefficients sometimes appear at negative apparent velocities for all three components. Because the semblance coefficients at negative apparent velocities tended to become large when the signal amplitude was high, the vertical array probably detected seismic signals reflected at surface of the Earth. Figure 3b shows the semblance analysis of daytime noise caused by nearby industry, where relatively high semblance coefficients appear at negative apparent velocities in contrast to the case of NVT, demonstrating that the vertical seismic array detection (VSAD) is an effective approach for distinguishing between deep seismic signals of natural origin (i.e. dominant positive apparent velocity) and shallow unwanted noises (i.e. dominant negative apparent velocity).

Figure 3.

Semblance analysis of the vertical seismic array at HGM during a period of (a) NVT activity and (b) artificial noise. Positive and negative apparent velocities show upward and downward incidence signals, respectively. See the caption of Figure 2 regarding the color plotted in the seismograms.

[9] We then applied the VSAD method to other vertical array observatories in the Kii Peninsula during the period from October 30 to November 20 to investigate the whole tremor episode. Only horizontal components were used to detect the signals of NVT, because the semblance analysis for horizontal components works better than for vertical as described above. We averaged semblance coefficients of two horizontal components near the S-wave velocity every 30 seconds and estimated a total duration per hour (hereafter, duration-rate) by counting the time that the averaged semblance coefficient exceeds 0.4. Here the criterion based on the magnitude of the apparent velocity plays a role in selecting energy coming only at a steep angle, so we can reject shallow coherent signals propagating upward (e.g., regular shallow earthquakes). In the case of the one-hour-long section in Figure 2, red portions were identified as NVT signals, together with blue portions which were detected by the ECM. Figure 4 represents duration-rate histories of NVT activity over the whole episode determined at each observatory. One notable feature is a pulsing pattern of NVT activity, which is discussed in the next section. The duration-rate generally depends on the signal-to-noise ratio at each site, so it tends to increase with decreasing hypocentral distance. This means that we can roughly infer a location of the source region of NVT episode on the basis of duration-rate patterns determined from multiple observatories. For example, the duration-rate was highest at HGM and lowest at ITA during the first NVT episode, which agrees with the observation that the activity during the period occurred just beneath HGM (Figure 1). The duration-rate also appears to correlate somewhat with the tremor migration during the second episode.

Figure 4.

Duration-rate histories of a NVT episode in November 2008, derived from ECM and VSAD technique. The bottom panel shows a volumetric strain calculated from a borehole strainmeter at MYM.

[10] We also produced a duration-rate history of NVT on the basis of our ECM catalogue, which is shown in the top of Figure 4. Note that the detection capability is significantly improved by the VSAD method. During the period from November 3 to 6, for example, the VSAD technique at HGM detected 2892 minutes of tremor, representing an over 10-fold detection rate increase as compared to the ECM detection (290 minutes). The VSAD method also succeeded in detecting minor NVT activities which were not detected by the ECM. For example, during the time period from the afternoon of November 7 to the morning of November 8, three observatories (ICU, HGM, and KST) showed a high duration-rate. The pattern of duration-rate histories indicates that the activity occurred near the station HGM as did the first episode.

4. Concluding Remarks

[11] We have shown that the VSAD technique enables detection of minor NVT activities which cannot be identified by the ECM, resulting in a dramatic improvement in detection capability. Furthermore, we obtained this result with fewer seismic sites (5 sites - 15 instruments for the VSAD versus more than 40 sites/instruments with the ECM). These results provide some guidance for future network design and installation. Here we discuss some possible applications of the VSAD method in seismic hazard assessments and seismic source studies.

[12] The likelihood of a large earthquake may increase during the occurrence of SSE [e.g., Dragert et al., 2001], so it is desirable to identify the initiation of SSEs as early as possible. As shown in Figures 1 and 4, the ECM started detecting NVTs on November 3, although seven tremors were detected on the morning of November 1 and four on the afternoon of November 2. In contrast, the VSAD method had already detected a large number of NVTs on the afternoon of October 31 (Figure 4). We frequently observed this same pattern in which the VSAD method detected many NVTs a few days earlier than the ECM. Thus the VSAD method will become a useful tool for early detection of SSE initiation.

[13] Near-source observations of ordinary earthquakes indicate that earthquakes initiate with a nucleation phase characterized by a low rate of moment release relative to the rest of the event [e.g., Ellsworth and Beroza, 1995]. Laboratory and theoretical models also predict the existence of a nucleation phase, which is considered to be the growth of slip instability on the fault prior to the earthquake (see Scholz [2002] for a review). Hirose and Obara [2010] determined source processes for seven repeating SSEs in the western Shikoku, southwest Japan, by analyzing the Hi-net tilt data and compared them with NVTs determined by the ECM. Their results imply that the SSEs nucleate with a small slip before the acceleration of moment release, similar to that of ordinary earthquakes. The onset of NVT activity is generally coincident with that of slow tilt deformation, but a majority of NVTs were detected after the acceleration of moment release. As described above, the VSAD method frequently detected many NVTs a few days earlier than that the ECM. We infer that the VSAD method has the capability to detect seismic signals associated with a small slip during a much earlier stage of nucleation, although such a small slip cannot be detected either geodetically or by the ECM. A detailed investigation of the nucleation process is needed to reveal the physical processes controlling SSEs, which may also contribute to a better understanding of ordinary earthquakes.

[14] Previous studies show that NVTs are strongly modulated by both earth and ocean tides with periodicities of about 12 and 24 hours that correspond to the principal lunar and lunisolar tides [e.g., Rubinstein et al., 2008]. The same pulsing patterns of NVT activities are clearly identified in duration-rate histories estimated at every observatory, and these generally correlate with dilation of volumetric strain as shown by a blue curve (Figure 4). The enhanced NVT detection capability of the VSAD technique, furthermore, makes it possible to detect shorter-term pulses over the time scale of a few hours. We infer that these pulses reflect smaller-scale spatial and temporal evolution of SSE [e.g., Ghosh et al., 2010; McCausland et al., 2010].

[15] The enhanced detection capability of NVT opens a new window for the study of deep deformation in the brittle-ductile transition zone, and its application to the occurrence of large earthquakes.

Acknowledgments

[16] Seismograph stations used in this study include permanent stations operated by NIED (Hi-net), JMA, and ERI. We are grateful to JMA for the hypocenter list of low frequency earthquakes. Comments by A. Guilhem, A. J. Ghosh, R. Harris, P. A. McCrory, and W. L. Ellsworth were helpful in improving the manuscript. Figures have been generated using the Generic Mapping Tools [Wessel and Smith, 1998]. This work was supported by KAKENHI (20340115), US-Japan Joint research project, and the Earthquake Research Institute cooperative research program.

[17] The Editor wishes to thank Aurélie Guilhem and Abhijit Ghosh for their assistance evaluating this paper.