You are on page 1of 10

GEOPHYSICS, VOL. 71, NO. 6 NOVEMBER-DECEMBER 2006; P. V153V162, 9 FIGS., 1 TABLE. 10.1190/1.

2345054

On bias and noise in passive seismic data from nite circular array data processed using SPAC methods

Michael W. Asten1

ABSTRACT
The nite nature of typical small seismic arrays used in conjunction with spatial autocorrelation SPAC processing for observing the microtremor waveeld causes predictable perturbations of the SPAC spectrum when sources of seismic noise are conned to a restricted range of azimuths. Such perturbations are especially evident at higher frequencies where wavelengths are on the order of the array radius. The effects are readily modeled and show that the triangular array geometries commonly used for microtremor studies require azimuthal distributions of wave energy on the order of 60 or greater to have a high probability of being free of such perturbations. The imaginary component of the SPAC spectrum, which is ideally zero for a sufciently dense circular array and/or a sufciently isotropic waveeld, is in practice often nonzero and provides three quality-control indicators: 1 an indication of insufcient spatial averaging, 2 an empirical measure of the level of statistical uncertainty in SPAC spectral estimates, and 3 an indication of departures from planewave stationarity of the seismic noise waveeld.

INTRODUCTION
When using the passive seismic microtremor method coupled with a modied spatial autocorrelation SPAC array processing method to obtain a shear-wave velocity prole of the earth, the technique of nite array sampling of the waveeld has some specic underlying assumptions that impose important limitations on the resulting proles. Microtremor and other passive seismic techniques are relatively new, and a consensus on terminology among interested groups worldwide has not been reached; so it is helpful to discuss the genesis of the various passive seismic methods as well as their limitations.

The microtremor method uses background seismic noise typically generated by cultural sources vehicle trafc, railways, machinery and natural sources wind noise, surf action at coastlines. Energy from such sources propagates principally as surface waves. Array types in common use detect only the vertical components of ground motion and therefore are sensitive to Rayleigh modes. Case histories published to date see references discussed below show that the fundamental Rayleigh mode is usually, but not always, dominant. This background seismic noise has historically been termed microseisms Toksoz, 1964, but the term microtremor has become more popular in the last 15 years, especially as introduced by Japanese authors, e.g., Nakamura 1989, Tokimatsu 1997, Satoh et al. 2001, and Okada 2003. Array methods utilizing microtremor sources extract depth information directly or indirectly from the dispersive frequency-phase velocity curves associated with surface-wave propagation. Recently, the European literature has used the term ambient noise rather than microtremor energy, especially in reports of the SESAME project, funded between 2001 and 2004 by the European Union SESAME, 2004. The microtremor method usually produces a shear-wave velocity versus depth prole for a layered or 1D approximation to the earth. This method was developed initially for earthquake hazard assessment studies shakeability of soft sediments, but increasingly it is being used in civil engineering site studies Watanabe and Takeuchi, 2004. In addition, the term passive seismic has been used in the petroleum industry to describe a quite different seismic concept, that is, when seismic arrays are placed to detect body waves generated by microearthquakes associated with movement or cracking in hydrocarbon reservoirs during hydrofracturing or depletion Asanuma et al., 2004; Kapotas et al., 2004; Shapiro et al., 2004; Duncan, 2005. Arrays implementing this methodology use raypath concepts to map the distribution of body-wave sources within the specied earth volume. A third passive seismic method recently adopted within the petroleum industry Holtzner et al., 2005 uses signals detected by individual surface seismometers, processing them to obtain spectral characteristics that might detect hydrocarbons directly .

Manuscript received by the Editor August 25, 2005; revised manuscript received April 23, 2006; published online October 16, 2006. 1 Monash University, Centre for Environmental and Geotechnical Applications of Surface Waves CEGAS, School of Geosciences, Melbourne VIC 3800, Australia. E-mail: michael.asten@sci.monash.edu.au. 2006 Society of Exploration Geophysicists. All rights reserved.

V153

V154

Asten found to yield higher resolution Asten, 1976, 2001; Okada, 2003; Hartzell, 2005.

Although the three passive seismic methods appear to have some similarities in terms of data acquisition, their assumptions with respect to energy sources, methods of data processing, and interpretation are fundamentally different. In the remainder of this paper, we consider only the rst of these three passive seismic methods: microtremor or ambient noise. In the microtremor method, using Rayleigh-wave energy, four classes of array processing have been reported in the literature over roughly the past 50 years. The simplest processing method is singlestation analysis of three-component data to give spectra for the ratio of horizontal to vertical particle motion, a method based on a study by Nogoshi and Igarashi 1971 that was later popularized by Nakamura 1989. Bard 1998 reviewed the method in detail, and it is now widely used for qualitative or semiquantitative mapping of sediment thickness over bedrock, particularly in earthquake hazard zonation studies Lachet and Bard, 1994; Lermo and Chavez-Garcia, 1994; Ibs-von Seht and Wohlenburg, 1999; Scherbaum et al., 2003; Lang and Schwartz, 2005. The second and third processing methods are conventional and high-resolution beam-forming or frequency-wavenumber f - k array methods pioneered by Lacoss et al. 1969 and Capon 1969. They have been used recently by Liu et al. 2000, Scherbaum et al. 2003, Wathelet et al. 2004, Ohrnberger et al. 2004, and Kind et al. 2005. An important variant on these beam-forming methods is a frequency-slowness transform method McMechan and Yedlin, 1981, adapted by Park et al. 1999 for active surface-wave studies; it was further developed by Louie 2001 for use with microtremor data acquired with linear rather than 2D arrays. In addition, Park et al. 2005 applied the frequency-slowness method to data from circular arrays in combined active and passive seismic surface-wave methods. A fourth class of array processing methods is the SPAC method, pioneered incredibly, in a predigital age by Aki 1957, 1965 and generalized by Henstridge 1979 and Cho et al. 2004. In the context of modern data processing methods, this method can be more descriptively named the spatially averaged coherency method, with the advantage that the traditional acronym still applies. As discussed by Asten 2003a and in this paper, the SPAC method has an inherent advantage for processing seismic noise in that the method extracts the scalar wavespeed of seismic energy propagating across the array, irrespective of its direction and irrespective of the existence of single or multiple sources. This insensitivity to wave azimuth presents a signicant advantage when sources of seismic noise by denition are undened. In contrast, beamforming methods seek vector-wave velocity and are limited by the resolution of the array when multiple sources are distributed in azimuth. Where both f - k and SPAC methods have been applied to the same data, SPAC methods have been

THE SPAC ARRAY PROCESSING METHOD The basic concept


The conceptual basis for the SPAC array processing method is illustrated in Figure 1, which shows a circular array of n geophones with an additional central geophone. The process of modeling SPAC for an azimuthal distribution of noise-free plane waves, as observed by a set of n geophone pairs distributed in azimuth around the circle, may be expressed as a summation of complex coherencies c jc f , each having amplitude unity, with phase given by

c jc f = expir jck cos jc ,

where c jc f is a coherency measured by a single pair of geophones, r jc is the displacement of the jth geophone relative to the center at azimuthal angle jc, k is the spatial wavenumber at frequency f , and is the azimuth of propagation of a single plane wave across the array. We assume that the wave energy is unimodal, i.e., it has a unique scalar wavespeed v f at any given frequency. In the ideal case of a single plane wave observed by an innite number of geophones placed around a circle centered on a single reference geophone, the summation is expressed as an integration an azimuthal average that yields a purely real SPAC spectrum; Aki 1957 originally derived this relation, and it was given by Okada 2003, equation 3.72 as

1 c f = 2

expirk cos d

= J0rk = J0

2 fr , v f

Figure 1. Common array geometries used for microtremor studies using SPAC methods. a Six-station hexagonal array. b Threestation triangular array.

where c f is a spatially azimuthally averaged coherency, J0 is the Bessel function of the rst kind of zero order with the variable rk, f is the frequency, k is the scalar wavenumber, v f is the required phase velocity dispersion curve, and r is the station separation in the circular array. In the notation used here, the absence of the double subscript in the symbol c f denotes the spatially i.e., azimuthally averaged parameter. The same result is obtained for omnidirectional plane waves an innity of source azimuths observed by a single pair of geophones i.e., integrated around the circle over with xed, which was rigorously derived by Capon 1973. Note that these azimuthally averaged coherencies yield a purely real number; all phase data are lost, and scalar wave velocity is extracted from the amplitude of J0 kr for known values of r and f . This last result provides a basis for obtaining the azimuthally averaged coherency spectrum from more limited array types. Where energy sources are truly omnidirectional or can be considered omnidirectional when averaged over a sufciently long time, the averaged spectrum may be obtained at a range of r values from single pairs of geophones Ohori et al., 2002; Morikawa et al., 2004; the technique is reviewed and evaluated in detail by Chavez-Garcia et al. 2005. Such methods should be treated with caution, however, because omnidirectional wave energy is an ideal that may be achieved by recording over an extended period of time e.g., two days as used by Chavez-Garca et al., 2005, so they may not apply to

Bias and noise in passive seismic data data acquired over shorter recording times. Given the application of microtremor methods to engineering and site studies where survey logistics and/or economics demand faster data acquisition, in general, this ideal of omnidirectional source distribution is unlikely to be achieved. Thus, I consider here some consequences of limited azimuthal distribution of wave energy on the SPAC method.

V155

SPAC with a nite array


For a real array recording for a limited period of time, the number of geophones is nite, and the azimuthal distribution of sources may be narrow as in the case of a localized piece of vibrating machinery, moderate a dominant nearby road, or omnidirectional an urban area surrounded by roads. The response of a nite station spacing to an angular spread of wave energy can be modeled by a numerical summation Asten, 2003b; Asten et al., 2004 or analytically Henstridge, 1979; Cho et al., 2004; Okada, 2006. The numerical approach can be implemented by making the lefthand side of equation 1 a double summation over a nite number of m plane waves and n geophone pairs, as follows:

J0 kr to the third secondary maximum kr 20, while ci f is an oscillating function with an envelope of amplitude ranging from near zero to order 0.5, depending on the azimuthal symmetry of the incident wave relative to the array Figure 2e and f. For a six-station hexagonal array for any plane-wave sources, cr f is identical to cr f for the three-station triangular array for the same source, while ci f is zero everywhere Figure 2g and h. This pair of results also follows from the simple consideration of array symmetry and the fact that pairs of stations rest on each of two radii extending in opposite directions i.e., two radii along the same diameter of the circle. This geometry will produce complex conjugates of coherency. The results remain true for any even number of stations equispaced on the circumference, provided the waveeld remains coherent and nonattenuating across the array. Figures 3 and 4 illustrate the limitations of three array types when the incident plane-wave energy is nearly-unidirectional:

c f = c r f + c i f 1 = nm
n j m i

expir jck 3

cos jc i .

This spatially averaged coherency for a nite sum is in general a complex number with real and imaginary parts cr f and ci f . We proceed to compare both the real and imaginary parts of the summation in equation 3 with the theoretical ideal of the purely real function J0 kr, shown in equation 2. Asten 2003b and Asten et al. 2004 numerically model equation 3 for nite circular arrays of two to seven stations and azimuthal spreads of wave energy, = 5, 30, 60, and 90, where an array with n + 1 stations is dened to consist of n stations on the circumference plus one central station. The azimuthal spread was modeled in these numerical exercises as a summation of plane waves with azimuthal increments of 1. Figure 2 shows selected examples that illustrate the range of behavior attributable to the nite array. In summary, for stations equispaced on the circumference with an angular spread of incident unimodal waves, we can make the following generalizations: For a three-station triangular array and = 30 azimuthal spread of plane-wave energy, cr f approximates the theoretical J0 kr to the rst minimum kr 3.8, while ci f is an oscillating function with an envelope of amplitude ranging from near zero to about 0.8, depending on the azimuthal symmetry of the incident wave relative to the array Figure 2ad. For the same three-station triangular array and = 60, cr f approximates the theoretical

Figure 2. ad Modeled SPAC for a triangular array, n = 3, = 30, for four different dominant directions of wave propagation. The solid line is the real part of the SPAC spectrum cr f , the dashed line is the imaginary part ci f , and the thin solid line is theoretical J0 kr.At the top right of each plot is a diagram of the array geometry, together with an arc depicting the range of wave azimuths summed in the model. e,f As for a and b but with azimuthal spread of wave propagation increased to = 60. g,h Modeled SPAC for a hexagonal array, n = 6, = 30, for two dominant directions of wave propagation corresponding to c and d. Note that the thin solid line for J0 kr in a, c, e, and f is coincident with and overprinted by cr f .

V156

Asten For a three-station triangular array and = 5, cr f for most azimuthal directions departs from the theoretical J0 kr before the rst minimum kr 3.8, while ci f is an oscillating function with an envelope of amplitude ranging from near zero to 1, depending on the azimuthal symmetry of the incident wave relative to the array Figure 3ad. The chosen examples show the limits of departures of cr f from the theoretical J0 kr curve. For a two-station linear cross array and = 5, cr f departs from the theoretical J0 kr before the rst cross-over kr 2, while ci f is an oscillating function with an envelope of amplitude ranging from near zero to 1, depending on the azimuthal symmetry of the incident wave relative to the array Figure 3eh. For a six-station semicircular array and = 5, cr f follows the theoretical J0 kr up to at least the third cross-over kr 9, while ci f is an oscillating function with an envelope of signicantly reduced amplitude Figure 4. We can conclude from the above that, provided the azimuthal spread of wave energy is on the order of 30 or greater, the common triangular array geometry provides sufcient azimuthal averaging for SPAC data to be useful at least to the rst minimum of the SPAC spectrum for all wave azimuths. For some azimuths, this geometry will be useful to higher-order i.e., higher-wavenumber maxima. As discussed below, perturbations in the SPAC data attributable to the combination of a restricted source distribution and nite number of array stations produce biases in phase-velocity estimates that in turn may result in biased esti-

Figure 3. Modeled SPAC with = 5 for four examples of dominant directions of wave propagation: ad n = 3 triangular array, e, f n = 2 linear cross array, g and h n = 6 semicircular array. The solid line is cr f , the dashed line is ci f , and the thin solid line is theoretical, J0 kr. At the top right of each plot is a diagram of the array geometry together with an arc depicting the range of wave azimuths summed in the model.

Figure 4. Modeled SPAC for n = 6 semicircular array, with = 5, for six examples of dominant directions of wave propagation. The solid line is cr f , the dashed line is ci f , and the thin solid line is theoretical J0 kr. At the top right of each plot is a diagram of the array geometry together with an arc depicting the range of wave azimuths summed in the model.

Bias and noise in passive seismic data mates of shear-velocity versus depth proles. For a narrow-azimuthal or unidirectional spread of wave energy, biased velocity estimates from SPAC coefcients may occur in the vicinity of the rst SPAC minimum, and large deviations from the theoretical J0 kr curve can occur at higher wavenumbers. As described in the next section, this result is important when using the standard SPAC method. The limitations associated with triangular array geometry when observing highly directional wave energy e.g., associated with vibrations from xed machinery indicate that in such cases an array with closer azimuthal sampling is to be preferred. From Figure 4, a six-station semicircle or a twelve-station circular array appears to be sufcient for SPAC processing of unidirectional wave energy regardless of azimuth. Chouet et al. 1998 give examples of semicircular arrays with six and ten stations on the semicircle. These are used to study surface waves from an active volcanic center, which is obviously most likely to be a highly directional source.

V157

Use of SPAC spectra over an extended frequency range

Okada 2003 points out the difculty of extracting unique estimates of phase velocity v f from SPAC data at wavelengths sufciently short that the wavenumber-radius product exceeds the criterion kr = 3.8 rst minimum of J0 kr. He then develops the extended SPAC ESPAC method that uses data from nested triangular arrays of multiple radii.Okada 2003 and other authors implementing the SPAC method Bettig et al., 2001; Ohori et al., 2002; Chavez-Garcia et al., 2005; Hartzell et al., 2005; Hayashi, 2005 have transformed the SPAC spectrum to the phase-velocity dispersion curve v f and then t the dispersion curve to a modeled dispersion curve computed for a layered earth. An alternative approach introduced by Asten et al. 2002 and used by Asten et al. 2004, Asten 2004, 2005 and Roberts and Asten 2004, 2005 ts modeled SPAC spectra directly to SPAC spectra estimated from array data at multiple values of r. Model SPAC spectra are obtained by rst computing phase velocity dispersion curves v f versus f for a layered earth using code from Herrmann 2001, which implements matrix methods from Saito 1979, 1988. These model phase velocities are then used with the right-hand side of equation 2 for a given station separation r or, more usually, for multiple values of r to yield model SPAC spectra. Henstridge 1979, and personal communication points out that the process of azimuthal averaging with a circular array does not necessarily require use of a central geophone, and azimuthal averaging over n /2 different values of r is possible using data from a circular array of n geophones e.g., distances r2, r3, r4 shown Figure 5. Comparison of SPAC spectra for triangular and hexagonal arrays. Real part on Figure 1a. These distances are achieved by sethick line and imaginary part thin bars of SPAC spectrum, with superimposed model lecting pairs of geophones from around the cirSPAC spectrum dashed line computed from the layered-earth model given in Table 1. Spectra are computed from a raw data time series of length 300 s. a Triangular array cumference of the circle; however, the sampling ACE. b Triangular array BDF, rotated in azimuth by 60. c Hexagonal array. df As of the waveeld in azimuth is exactly the same as for ac where the imaginary SPAC spectrum is roughened. The rms value for the roughit would be if a central geophone and circular arened imaginary SPAC spectrum 020 Hz is annotated. ray of radius r3 or r4 were used.

The combination of direct tting of model and observed SPAC spectra with multiple values of r has been termed the multiple mode SPAC MMSPAC method because it can identify frequency bands with higher-mode surface-wave energy Asten et al., 2004. In principle, simultaneous tting of SPAC coefcients for a given frequency for three different values of r permits the solution of two phase velocities and an energy partition coefcient Asten, 2001. In practice, such quantitative separation of modes is difcult because of the uncertainty as to which and how many higher modes are present. However, qualitative recognition of the presence of higher modes is easily achieved; Figure 7 of Asten et al. 2004 gives an example. Direct tting of model and observed SPAC spectra also has the particular advantage of facilitating the use of SPAC spectra through multiple maxima and minima of the J0 kr curve. In practice, it is common in MMSPAC processing to nd SPAC spectra useful to the third secondary maximum e.g., eld data in Figure 5, discussed below and occasionally to the fth maximum. This approach widens the frequency range over which the SPAC method may be applied on a given array; therefore, the SPAC method can be more versatile than f -k processing methods, which are usually limited by spatial aliasing to a much more restricted range of wavenumbers. However, this MMSPAC method also increases the need to understand the geometry of a nite array and limited angular distribution of wave energy, and how these factors may affect the SPAC spectrum at high frequencies. From the modeling shown in Figure 2, we conclude that an angular distribution of microtremor energy on the order of 60 or greater

V158

Asten

A. Complex coherencies computed for every pair i, j in the array of n stations may be expressed as an n n Hermitian matrix cij f , where each element of the matrix that estimated using equation A-1. Azimuthal averaging of coherencies is achieved by averaging those elements of the matrix that correspond to pairs of stations having the same scalar distance separation illustrated in Figure 1. The basic theory of the SPAC processing method invokes two fundamental assumptions. First, the spatially averaged coherencies are purely real numbers. Second, no incoherent noise is detected by the array.Athird assumption implicit in all array processing methods for microtremor data is that the waveeld is spatially stationary, i.e., plane wave and uniform across the array. This paper shows that the nonzero behavior of the imaginary curve ci f is a useful aid for quality control of eld data because it is only in the ideal case of perfect USE OF THE IMAGINARY COMPONENT azimuthal averaging of coherencies that the imaginary part is zero OF SPAC IN QUALITY CONTROL everywhere. Thus, any departure from zero carries information on the quality of the assumptions behind the SPAC method. Such deparIn the model SPAC spectra discussed above, the waveeld is astures fall into two categories: a slowly oscillating ci f versus f sumed to be noisefree, and the coherency between any pair of array curves, resulting from the nite nature of the array, and b statistical stations is by denition a complex number with amplitude unity and noise on coherency estimates; the noise varies randomly from point phase given by equation 1. With real eld data, complex coherencies to point on the SPAC spectra. are estimated using time series analysis as summarized in Appendix Both categories are studied using a sample of data from a hexagonal array used in a successful Table 1. Layered-earth model from interpretation of McEwan Park microtremor data, for layer parameters thickness H, compressional and blind trial of the MMSPAC method in Wellingshear velocity Vp, Vs, and density P. Parameters in parentheses for layer 3 ton, New Zealand. Interpretation of the data gave are the perturbed model used for Figure 5. [Note: Vp and P values are guesses shear velocities and depths that were accurate to based on probable depth to water table for sands and gravels (layers 1-5). For basement rock (layers 6-8), Vp and Vs are not resolved by these measurements better than 5% Asten et al., 2005. Figure 5a and but are included as assumed reasonable values for hard rock.] b shows SPAC spectra for two superimposed triangular arrays ACE and BDF, which together make a hexagonal array of the type shown in FigVsm/s P t/m3 Geology Layer H m V p m/s ure 1a. In both cases, cr f is readily compared with a model SPAC spectrum using the MMS1 2 500 130 1.8 Holocene PAC method noted above. The earth model used 2 2.8 1700 140 2.0 Silts and mud is shown in Table 1 documented by Asten et al., 3 8 6.5 1700 160 144 2.0 Silts and mud 2005, from the interpretation of data acquired 4 9 1700 160 2.0 Silts and mud with a hexagonal array. The curve ci f shows 5 400 1700 550 2.4 Glacial strongly developed oscillations maxima-minima gravels exceeding 0.5 in amplitude which, by analogy 6 1040 3880 2230 2.4 Basement with Figure 2, clearly indicate that microtremor unresolved sources are distributed in azimuth over less than 7 1800 4630 2680 2.4 60. It is immediately apparent that Figure 5a shows 8 inf 6040 3490 2.8 a poor t to the superimposed model SPAC spectrum; the observed SPAC spectrum has biased results in the 515-Hz frequency band. The imaginary curve serves as a warning that such bias is a risk when this data set triangular array ACE is used in isolation. Figure 6 quanties the effect of the biased SPAC spectrum of Figure 5a. A best-t model for array ACE requires a reduction in both shear velocity 10% and thickness 25% of layer 3 see Table 1, which differ signicantly from the veried results given in Asten et al. 2005. In contrast, array BDF does not show the Figure 6. SPAC spectra, triangular arrays, with perturbed model. Real part thick line bias in the SPAC spectrum and does not produce and imaginary part thin bars of SPAC spectrum. Dotted lines are the superimposed any changes in the layered-earth model. This ilmodel SPAC spectra for the true layered-earth model of Table 1, and dashed line for the lustrates one of the limitations of triangular array perturbed model altered to t the biased SPAC spectrum of triangle ACE. a Triangular array ACE from Figure 5d. b Triangular array BDF rotated in azimuth by 60 from geometries: The imaginary curve provides a Figure 5e. The standard deviation of the t between the real parts of eld and model cowarning of possible bias in the real curve, but bias herency SPAC spectra over 2.515 Hz improves from 0.147 to 0.116 for arrayACE, but does not necessarily result. degrades from 0.116 to 0.139 for array BDF. is required for MMSPAC processing to have a high probability of being effective up to the high frequencies represented by kr 20 the third secondary maximum of the J0 curve. In fact, the actual angular distribution of energy at a site is rarely known, but the perturbations departures from J0 kr in model SPAC spectra such as those shown in Figure 2a and b may be recognizable in eld data. James Roberts 2005, personal communication observes that although Figures 24 illustrate such perturbations, signicant changes may occur when the angular spread is not a simple submultiple of the station azimuths e.g., = 40 or 70. Therefore, it is useful to apply the modeling tool over a range of likely azimuthal wave distributions before drawing conclusions regarding any particular set of eld data.

Bias and noise in passive seismic data It is also worth noting that, in practice, the angular spread of sources may change with frequency; thus, the level of bias in the SPAC spectrum of eld data may be a frequency-dependent combination of the types of bias depicted in Figures 2 and 3. Figure 5c shows that use of the full hexagon to compute the SPAC spectrum leads to a signicant reduction, but not a total elimination, of the cyclic character of ci f . The residual cyclic character is deduced to be an indicator of the departure of the microtremor waveeld from a set of plane waves exhibiting spatial stationarity because the modeling above and in previous studies establishes that the imaginary curve is zero everywhere for the circular array with an even number of stations. Such a departure from stationarity could be a result of any or all of the following factors: 1 near sources that give rise to curved wavefronts; 2 attenuation that causes changes in the waveeld across the array; or 3 local lateral variations in geology beneath the array. The circumstances in which such departures from stationarity may bias SPAC spectra is a subject for further research. Here, I can only point out the role of the imaginary curve in detecting departures from the third of the basic assumptions of the SPAC method.

V159

Estimating uncertainty in SPAC models

Statistical noise is obvious on both the real and imaginary parts of the SPAC spectra in Figure 5. It is desirable to estimate the envelope of statistical noise error bars for any inversion routine, but for multi- or omnidirectional microtremor noise, it is not a straightforward process. Estimates of uncertainty of coherency amplitude values that are conned to the range zero to one, and based on analysis of two time-stationary series as done by Koopmans 1974, p. 351, cannot be applied to SPAC spectra because the multidirectionality of the waveeld can yield coherency amplitude estimates for individual pairs of stations in the range 0.5 to 1.0, even in noise-free data. It is therefore more desirable to construct uncertainty values directly from the data. Uncertainty is a function of the degrees of freedom in each coherency estimate, which can be computed from the length of the data sample and width of frequency windows as outlined in Appendix A. If the original time series has a pseudorandom character, then the real and imaginary parts of each complex coherency value are independent and subject to the same statistical uncertainty. The imaginary curve, which in general has much smaller systematic or oscillatory character than the real curve, provides a measure of random noise statistics that also apply to the real curve. In the ideal case, ci f in Figure 5c for a hexagonal array would be associated with random noise only. In the real world of triangular arrays and/or directional waveelds, some ltering of the imaginary SPAC spectral curve is needed to remove the systematic oscillatory character. The process can be described as roughening the SPAC spectrum because one rst smoothes the spectrum to isolate the systematic oscillatory component and then subtracts this smoothed part to Figure 7. SPAC spectra, with the imaginary part of roughened SPAC spectrum shown as leave the residual random noise. The process is rms values. Real part thick line and imaginary part thin bars of SPAC spectrum, with demonstrated on eld data in Figure 5. superimposed model SPAC spectrum dashed line computed from layered-earth model Figure 5df shows the results of roughening given in Table 1. a As for Figure 5f, where the roughened imaginary part is shown as rms values, smoothed over a running window of 11 f units, and plotted both positive and the imaginary curves of Figure 5ac by rst negative to represent the uncertainty range. b As for a, where the uncertainty bars smoothing with a Hanning bell lter. The cohercomputed from rms values of the roughened imaginary part are overplotted on the real ency spectrum is sampled at the frequency interpart.

val f a = 0.13 Hz see Appendix A, and the width of the Hanning bell lter is chosen to be about half the width of the lobes of the maxima and minima of the real curve that dene the cyclicity we wish to reject. The lter used here has a width of 11 f a. The roughened imaginary curve provides a useful measure of the statistical noise envelope of the SPAC spectrum. Root-mean-square values for this noise envelope computed for the entire frequency range, 020 Hz, are annotated on the plots. The effects of random statistical noise for the hexagonal array relative to the triangular arrays should be proportional to n1/2 with n as the number of stations in the array and the rms gures are consistent with this. On the hexagonal array, the noise envelope in Figure 5f shows the variations in statistical noise at different frequencies; the 36-Hz band is clearly lower in noise, with SPAC coefcients stable within an envelope of 0.03 coherency units. The noise envelope of the roughened imaginary curve can be recovered by computing the rms values in a rectangular-shaped running-average window of width 11 f a chosen to be the same width as for the high-pass lter. Plotting these rms values as uncertainty bars, both positive and negative, in Figure 7a clearly depicts the variation in the statistical noise envelope with frequency. Figure 7b shows an alternative presentation where the same uncertainty bars are plotted on the real part of the SPAC spectrum. The process is further tested in Figures 8 and 9 where a raw sample of array data of reduced length 15,000 time samples, selected from the center of the original length of 60,000 time samples is used to compute SPAC spectra mirroring the procedures used for Figures 5 and 7; however, for reasons of space, we restrict plots to the hexagonal array only. Coherencies are computed over approximately the same frequency window widths in hertz as for Figure 5, which reduces the degrees of freedom in the statistical estimate of coherency by the same factor of four Appendix A. When Figures 5c and 7a are compared, we can see that the imaginary curve shows the same cyclicity in the two plots as expected because this feature is a function of array and source geometry, not signal processing statistics. However, the same comparison shows a higher level of point-to-point statistical noise in Figure 7a also expected, because of the reduced degrees of freedom in the coherency estimates. Figure 8b is obtained after applying the same roughening lter to the imaginary part, as for Figure 5f. The computed rms value of the roughened imaginary SPAC spectrum is 0.09, as compared to 0.05 in Figure 5f. Figure 9 presents the statistical noise envelope as uncertainty bars following the form of

V160

Asten from plane-wave stationarity. Point-to-point variations in the imaginary SPAC spectrum can be separated by using a roughening lter on that part of the SPAC spectrum to yield statistical uncertainty bars for the SPAC spectral estimates. In addition, this method may provide quality control where single pairs of geophones are used over extended recording times. For example, Chavez-Garcia et al. 2005 use beam-forming methods to demonstrate the isotropic nature of the waveeld in their eld study, whereas plots of the imaginary part of the SPAC spectrum for each pair of stations may well provide the necessary conrmation.

Figure 8. SPAC spectra for the hexagonal array, using a reduced data length. Real part thick line and imaginary part thin bars of SPAC spectrum, with superimposed model SPAC spectrum dashed line computed from layered-earth model given in Table 1. a As for Figure 5c, where spectra are computed from a raw data time window of reduced length 75 s. b As for a where the imaginary SPAC spectrum is roughened. The rms value for the imaginary SPAC spectrum 020 Hz is annotated.

ACKNOWLEDGMENTS
Field data shown in Figures 48 were acquired in collaboration with the Institute of Geological and Nuclear Sciences, New Zealand; thanks to Bill Stephenson and Peter Davenport for their assistance in the data acquisition. Surface-wave dispersion-curve modeling software used in this project was supplied by R. B. Herrmann. The author thanks David Boore and W. J. Stephenson of Figure 9. SPAC spectra using reduced data length, with imaginary part of the roughened the U. S. Geological Survey USGS and James SPAC spectrum shown as rms values. Real part thick line and imaginary part thin bars Roberts for many discussions and manuscript reof SPAC spectrum, with superimposed model SPAC spectrum dashed line computed views in the course of this work. Three anonyfrom layered-earth model given in Table 1. a As for Figure 7a, where the roughened mous reviewers provided valuable suggestions to imaginary part is shown as rms values, smoothed over a running window of 11 f units, clarify and improve the text. and plotted both positive and negative to represent the uncertainty range. b As for a, where the uncertainty bars computed from the imaginary part are over-plotted on the real The author is supported by the USGS, Departpart. ment of the Interior, under USGS awards 04HQGR0030 and 05HQGR0030. The views Figure 7, and the increased envelope of statistical noise associated and conclusions contained in this document are those of the author with the reduced time-length of array data is clearly evident. and should not be interpreted as necessarily representing the ofcial In Figures 7 and 9, we now have, in the form of the statistical noise policies, either express or implied, of the U. S. Government. This envelope, an empirical representation of uncertainty in the estimated work has also been supported by a series of short-term collaborative SPAC spectrum that a is based on properties of the eld data, b research grants from Geoscience Australia. can be used regardless of the degree of angular spread of the wave energy, c shows variations in level of statistical noise across the APPENDIX A spectrum, and d after roughening, can be applied to SPAC data from triangular or other nonsymmetric arrays.

DATA PROCESSING SEQUENCE FOR COMPUTATION OF COHERENCY

CONCLUSIONS
The nite nature of typical small seismic arrays used in conjunction with SPAC processing to obtain the microtremor waveeld causes predictable perturbations of the SPAC spectrum, especially at higher frequencies where wavelengths are on the order of, or shorter than, the array radius. The effects are readily modeled and show that typical triangular array geometries require azimuthal distributions of wave energy on the order of 60 or greater to have a high probability of being free of such perturbations. Modeling predicts that a sixstation semicircular array is sufcient to avoid such perturbations over a wide range of frequencies up to kr20 when wave energy is unidirectional. The imaginary component of the SPAC spectrum is generally nonzero. Slowly varying cyclicity with respect to frequency can indicate that the spatial averaging is insufcient, and for arrays with an even number of geophones on the circle, that there are departures

Spectral analysis can be performed either by the block averaging method averaging cross-spectra computed for multiple time windows of multichannel data or by frequency averaging averaging cross spectra from a single time window over specied frequency windows. The two methods are statistically equivalent Koopmans, 1974. The following steps summarize processing of eld data by the frequency-averaging method as used for examples in this paper; they follow Koopmans 1974, except for the use of direct averaging over a spectral window in the frequency domain, rather than averaging the covariance function. A data window consists of n channels, length N points, over sample interval t. Each channel is weighted with a fader, usually a Hanning bell, to minimize leakage of energy between spectral maxima and minima. Raw spectra Si f for each ith channel are computed with a discrete Fourier transform, yielding spectral estimates at discrete frequency intervals given by the relation f = 1/ Nt.

Bias and noise in passive seismic data Coherency or normalized cross-power spectral estimates between each pair i, j of channels are computed using the relation

V161

cij f =

Si f Si f S j f S j f ,

S i f S j f

A-1

where cij f is complex coherency also called complex coherence by Koopmans, 1974, p. 137, * denotes complex conjugate, and the bar denotes averaging over a spectral window of xed width. The spectra are divided into bands for spectral averaging, each of constant width and having a frequency spacing given by f a = m f . Each coherency estimate for a pair of stations has a value for the maximum degrees of freedom given by

2m /Kb ,

A-2

where Kb is a factor determined by the shape of the fader or taper applied to the original time window of data. For a square time window, Kb = 1. Using the formula supplied by Koopmans 1974, p. 300, Asten 1976, Appendix 2 computes that Kb = 35/18 for the Hanning window. For practical purposes, we may use Kb = 2, and we have m for the Hanning fader. The degrees of freedom are a maximum when the input time series are pseudorandom but are reduced when the time series contain impulsive spikes or wavelets which give rise to correlated values in the Fourier spectrum. For Figure 5, N = 60,000, t = 0.005 s, m = 39; therefore, f a = 0.13 Hz and = 39. For Figure 6, N = 15,000, t = 0.005 s, m = 9; thus, f a = 0.12 Hz and n, where is the degrees of freedom for the coherency estimate for each pair of stations. When n coherency estimates are averaged around the circle to give the spatially averaged coherency SPAC, the degrees of freedom of the sum is additive Koopmans, 1974, p. 270, i.e., n.

REFERENCES
Aki, K., 1957, Space and time spectra of stationary stochastic waves, with special reference to microtremors: Bulletin of the Earthquake Research Institute, 35, 415456. , 1965, A note on the use of microseisms in determining the shallow structures of the earths crust: Geophysics, 30, 665666. Asanuma, H., Y. Kumano, T. Izumi, H. Kaieda, K. Tezuka, D. Wyborn, and H. Niitsuma, 2004, Passive seismic monitoring of a stimulation of HDR geothermal reservoir at Cooper Basin, Australia: 74th Annual International Meeting, SEG, Expanded Abstracts, 556559. Asten, M. W., 1976, The use of microseisms in geophysical exploration: Ph.D. thesis, Macquarie University. , 2001, The spatial auto-correlation method for phase velocity of microseisms Another method for characterisation of sedimentary overburden, in K. McCue, C. Bubb, D. Finlayson, G. Horoschun, and B. Butler, eds., Earthquake codes in the real world: Proceedings of the Australian Earthquake Engineering Society, Paper 28. , 2003a, Historical note and preface to translation of The Microtremor Method, in K. Suto, trans., Microtremor survey method, SEG of Japan, Geophysical Monograph Series 12, SEG. , 2003b, Lessons from alternative array design used for high-frequency microtremor array studies, in J. L. Wilson, N. K. Lam, G. Gibson, and B. Butler, eds., Earthquake risk mitigation, Proceedings the Australian Earthquake Engineering Society, Paper 14. , 2004, Passive seismic methods using the microtremor waveeld for engineering and earthquake site zonation: 74th Annual International Meeting, SEG, Session NSG-1. , 2005, An assessment of information on the shear-velocity prole at Coyote Creek, San Jose, from SPAC processing of microtremor array data, in M. W. Asten and D. M. Boore, eds., Blind comparisons of shear-wave velocities at closely spaced sites in San Jose, California: U. S. Geological Survey Open-File Report 2005-1169. Asten, M. W., T. Dhu, and N. Lam, 2004, Optimised array design for microtremor array studies applied to site classication Observations, results and future use: Proceedings of the 13th Annual World Conference of Earthquake Engineering, Paper 2903.

Asten, M. W., N. Lam, G. Gibson, and J. Wilson, 2002, Microtremor survey design optimised for application to site amplication and resonance modelling, in M. Grifth, D. Love, P. McBean, A. McDougall, and B. Butler, eds., Total risk management in the privatised era: Proceedings of the Australian Earthquake Engineering Society, Paper 7. Asten, M. W., W. R. Stephenson, and P. Davenport, 2005, Shear-wave velocity prole for Holocene sediments measured from microtremor array studies, SCPT, and seismic refraction: Journal of Engineering and Environmental Geophysics, 10, 235242. Bard, P.-Y., 1998, Microtremor measurements, a tool for site effect estimation?, State-of-the-art paper: 2nd International Symposium on the Effects of Surface Geology on Seismic Motion, 3, 12511279. Bettig, B., P. Y. Bard, F. Scherbaum, J. Riepl, F. Cotton, C. Cornou, and D. Hatzfeld, 2001, Analysis of dense array noise measurements using the modied spatial auto-correlation method SPAC Application to the Grenoble area: Bollettino di Geosica Teorica ed Applicata, 42, 281304. Capon, J., 1969, High resolution frequency-wavenumber analysis: Proceedings of the IEEE, 57, 14081418. , 1973, Signal processing and frequency-wavenumber spectrum analysis for a large aperture seismic array, in B. A. Bolt, ed., Methods in computational physics 13, Academic Press Inc. Chavez-Garcia, F. J., M. Rodriguez, and W.R. Stephenson, 2005, An alternative approach to the analysis of microtremors Exploiting stationarity of noise: Bulletin of the Seismological Society of America, 95, 277293. Cho, I., T. Tada, and Y. Shinozaki, 2004, A new method to determine phase velocities of Rayleigh waves from microseisms: Geophysics, 69, 1535 1561. Chouet, B., G. De Luca, G. Milana, P. Dawson, M. Martini, and R. Scarpa, 1998, Shallow velocity structure of Stromboli Volcano, Italy, derived from small aperture array measurements of Strombolian tremor: Bulletin of the Seismological Society of America, 88, 653666. Duncan, P. M., 2005, Is there a future for passive seismic?: First Break, 23, 111115. Hartzell, S., D. Carver, T. Seiji, K. Kudo, and R. Herrmann, 2005, Shallow shear-wave velocity measurements in the Santa Clara Valley, Comparison of spatial autocorrelation SPAC and frequency wavenumber FK methods, in M. W. Asten and D. M. Boore, eds., Blind comparisons of shearwave velocities at closely spaced sites in San Jose, California: U. S. Geological Survey Open-File Report 2005-1169. Hayashi, K., 2005, The result of surface wave method in the Coyote Creek borehole Williams Park, in M. W. Asten and D. M. Boore, eds., Blind comparisons of shear-wave velocities at closely spaced sites in San Jose, California: U. S. Geological Survey Open-File Report 2005-1169. Henstridge, J. D., 1979, A signal processing method for circular arrays: Geophysics, 44, 179184. Herrmann, R. B., 2001, Computer programs in seismology An overview of synthetic seismogram computation: Version 3.1: Department of Earth and Planetary Sciences, St. Louis University. Holzner, R., P. Eschle, H. Zurcher, M. Lambert, R. Graf, S. Dangel, and P. F. Meier, 2005, Applying microtremor analysis to identify hydrocarbon reservoirs: First Break, 23, 4146. Ibs-von Seht, M. and J. Wohlenburg, 1999, Microtremor measurements used to map thickness of soft sediments: Bulletin of the Seismological Society of America, 89, 250259. Kapotas, S., G. Tselentis, and N. Martakis, 2004, The leap to passive seismic imaging, Is it time?: 74th Annual International Meeting, SEG, Expanded Abstracts, 576579. Kind, F., D. Fah, and D. Giardini, 2005, Array measurements of S-wave velocities from ambient vibrations: Geophysical Journal International, 160, 114126. Koopmans, L. H., 1974, Spectral analysis of time series: Academic Press Inc. Lacoss, R. T., E. J. Kelly, and M. N. Toksoz, 1969, Estimation of seismic noise structure using arrays: Geophysics, 34, 2138. Lachet, C., and P.-Y. Bard, 1994, Numerical and theoretical investigations on the possibilities and limitations of Nakamuras technique: Journal of Physical Earth, 42, 377397. Lang, D. H., and J. Schwarz, 2005, Identication of the subsoil prole characteristics at the Coyote Creek Outdoor Classroom CCOC, San Jos, from microtremor measurements A contribution to the CCOC blind comparison experiment, in M. W. Asten and D. M. Boore, eds., Blind comparisons of shear-wave velocities at closely spaced sites in San Jose, California: U. S. Geological Survey Open-File Report 2005-1169. Lermo, J., and F. J. Chavez-Garcia, 1994, Are microtremors useful in site response evaluation?: Bulletin of the Seismological Society of America, 84, 13501364. Liu, H., D. M. Boore, W. B. Joyner, D. H. Oppenheimer, R. E. Warrick, W. Zhang, J. C. Hamilton, and L. T. Brown, 2000, Comparison of phase velocities from array measurements of Rayleigh waves associated with microtremor and results calculated from borehole shear-wave velocity proles: Bulletin of the Seismological Society of America, 90, 666678. Louie, J. N., 2001, Faster better shear-wave velocity to 100 meters depth from refraction microtremor arrays: Bulletin of the Seismological Society

V162

Asten
Exploration Geophysics, 35, 1418. , 2005, Estimating the shear velocity prole of Quaternary silts using microtremor array SPAC measurements: Exploration Geophysics, 36, 3440. SESAME, 2004, Site effects assessment using ambient excitations: Final Report, European Commission Research General Directorate, Project EVG1-CT-2000-00026 SESAME. Saito, M., 1979, Computations of reectivity and surface wave dispersion curves for layered media, Part I Sound wave and SH wave: ButsuriTansa, 32, 1526. , 1988, Compound matrix method for the calculation of spheroidal oscillation of the earth: Seismological Research Letters, 59, 29. Satoh, T., H. Kawase, T. Iwata, S. Higashi, T. Sato, K. Irikura, and H. Huang, 2001, S-wave velocity structure of the Taichung basin, Taiwan, estimated from array and single-station records of microtremors: Bulletin of the Seismological Society of America, 91, 12671282. Scherbaum, F., K.-G. Hinzen, and M. Ohrnberger, 2003, Determination of shallow shear wave velocity proles in the Cologne, Germany area using ambient vibrations: Geophysical Journal International, 152, 597612. Shapiro, S. A., M. Parotidis, S. Rentsch, and E. Rothert, 2004, Reservoir characterization using passive seismic monitoring, Physical fundamentals and road ahead: 74th Annual International Meeting, SEG, Expanded Abstracts, 2541-2544. Tokimatsu, K., 1997, Geotechnical site characterization using surface waves, in K. Ishihara, ed., Earthquake geotechnical engineering: Balkema. Toksoz, M. N., 1964, Microseisms and an attempted application to exploration: Geophysics, 24, 154177. Watanabe, F., and T. Takeuchi, eds., 2004, Application of geophysical methods to engineering and environmental problems: Advisory Committee on Standardization, SEG of Japan. Wathelet, M., D. Jongmans, and M. Ohrnberger, 2004, Surface-wave inversion using a direct search algorithm and its application to ambient vibration measurements: Near Surface Geophysics, 2, 211222.

of America, 91, 347364. McMechan, G. A., and M. J. Yedlin, 1981, Analysis of dispersive waves by waveeld transformation: Geophysics, 46, 869874. Morikawa, H., S. Sawada, and J. Akamatsu, 2004, A method to estimate phase velocities of Rayleigh waves using microseisms simultaneously observed at two sites: Bulletin of the Seismological Society of America, 94, 961976. Nakamura, Y., 1989, Amethod for dynamic characteristics estimation of subsurface using microtremors on the ground surface: Quarterly reports of the Railway Technical Research Institute Tokyo, 30, 2533. Nogoshi, M., and T. Igarashi, 1971, On the amplitude characteristics of microtremor, Part II: Journal of the Seismological Society of Japan, 24, 2640. Ohori, M., A. Nobata, and K. Wakamatsu, 2002, A comparison of ESAC and FK methods of estimating phase velocity using arbitrarily shaped microtremor arrays: Bulletin of the Seismological Society of America, 92, 23232332. Ohrnberger, M., F. Scherbaum, F. Krger, R. Pelzing, and S.-K. Reamer, 2004, How good are shear wave velocity models obtained from inversion of ambient vibrations in the Lower Rhine Embayment N. W. Germany?: Bollettino di Geosica Teorica ed Applicata, 45, 215232. Okada, H., 2003, The microtremor survey method, SEG of Japan, K. Suto, trans: Geophysical Monograph Series 12, SEG. , 2006, Theory of efcient array observations of microtremors with special reference to the SPAC method: Exploration Geophysics, 37, 7385. Park, C. B., R. D. Miller, N. Ryden, J. Xia, and J. Ivanov, 2005, Combined use of active and passive surface waves: Journal of Environmental and Engineering Geophysics, 10, 323334. Park, C. B., R. D. Miller, and J. Xia, 1999, Multichannel analysis of surface waves MASW: Geophysics, 64, 800808. Roberts, J., and M. W. Asten, 2004, Resolving a velocity inversion at the geotechnical scale using the microtremor passive seismic survey method:

You might also like