SE803-05 September 30, 2003 16:4

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Chapter 5

Data acquisition and

analysis methods

This chapter discusses the methods of acquisition and analysis of microtremor

data suitable for the detection of surface waves by the frequency-wavenumber
( f -k) method and the spatial autocorrelation (SPAC) method.
The f-k method requires a relatively large number of observation stations.
Therefore, its field effort is more intensive than for the SPAC method. At the
same time, the spatial extent of the array becomes larger than that of the
SPAC method, resulting in the violation of the assumption of “horizontal
layers.” However, when there are several dominant modes of surface waves in
the microtremors, or when it contains strong body waves, this method theo-
retically has the potential to separate these waves. Additionally, by this method
the direction of the source of the wave can be estimated, and the phase velocity
of waves of a wide frequency range can be obtained.
The SPAC method is simpler in field data acquisition and data analy-
sis than the f-k method. Using an array of three-component seismometers,
the SPAC method can separate Rayleigh and Love waves. However, when
higher-mode surface waves are mixed in the data, the SPAC method cannot
separate the fundamental-mode surface wave. Such a case is rather rare, but
where such a condition exists, the f-k method is available.

67
SE803-05 September 30, 2003 16:4

68 CHAPTER 5. DATA ACQUISITION AND ANALYSIS METHODS

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

5.1 Observing microtremors

5.1.1 Observation array
5.1.1.1 For the f-k method
More flexibility is allowed in array design for the f-k method than for the SPAC
method. The ideal array spread should have a large lateral extent with varied
distance between seismometers. In general, topography and other constraints
on seismometer location dictate the shape and size of the array. By considering
the data analysis process, and taking the possibility of later application of the
SPAC method into account, a desired array configuration would be a seismo-
meter at the center of a spread with other seismometers arranged to form
equilateral triangles of various sizes concentric with the central station.
A larger number of observation stations provides greater precision in the
f-k spectra. For example, the gigantic LASA in America deploys 525 seis-
mometers in the array with a diameter of about 200 kilometers. However, con-
sidering restrictions in station setting and analysis time, around ten seismo-
meters may be a practical number for the purpose of exploring subsurface
structure.

5.1.1.2 For the SPAC method

The observation array to collect data for analysis by the SPAC method requires,
as seen in Figure 5.1 for example, at least four seismometers, three of them ar-
ranged on a circle of radius r and one at the center. It is expected that a larger
number of seismometers will give a better result.
If the three stations on the circle form a equilateral triangle, the directional
calculation becomes simpler, and the distance r in (3.80) can take two values,

Figure 5.1. A four-point array.

SE803-05 September 30, 2003 16:4

5.1. OBSERVING MICROTREMORS 69

√
Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

i.e., the radius r and the length of the side of the equilateral triangle 3r . If
two or more arrays with different r are spread at the same location, a wider
range of phase velocity can be estimated.
The extended spatial autocorrelation method does not require an equilat-
eral triangle array, and it allows an array similar to the frequency-wavenumber
method.

5.1.2 Data acquisition system

There are two types of data acquisition systems: one with all the stations of
an array connected and recorded with a multichannel recorder, and the other
with each station recorded independently.
The former system has the advantage that no time correction is necessary,
since all the stations are recorded against the same clock. This system is well
suited to a small array, for estimation of shallow structure. However, there are
operational disadvantages for large arrays; the difficulty of wiring restricts
selection of station sites, and the requirement for maintenance of the wire
throughout the observation.
The second system requires time calibration among the stations, for ex-
ample, by using a GPS clock installed in each recorder to synchronize the
records. However, it imposes little restriction in selection of the station sites.
Considering all the advantages and disadvantages, the independent system
seems more suited to the recording of microtremors.
Finally, it is desirable for the data to be in digital form to facilitate data
processing.

5.1.3 Time for data collection

The time of day for data collection, i.e., day or night, does not affect the content
of the data. However, it is best to avoid areas subject to noise, e.g., areas with
heavy traffic which disturbs the stationarity of microtremors required for data.
Data for many of the examples in this book were collected at night, the reason
being that there is little unstable cultural noise at night, and the microtremors
as a stationary process dominate in the data.
The data length is typically forty-five minutes to one hour for long period
microtremors. However, thirty minutes may be sufficient for microtremors
with a period shorter than one second.
SE803-05 September 30, 2003 16:4

70 CHAPTER 5. DATA ACQUISITION AND ANALYSIS METHODS

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Microtremor data recorded at independent stations require some prepro-

cessing. One of the requirements is synchronization of each station.

5.2 Data analysis

5.2.1 Data analysis by frequency-wavenumber
method
Figure 5.2 is a data processing flow chart using the frequency-wavenumber
method. Since the theory of the data analysis was explained in Chapter 3, only
the practical data processing sequence is presented here.
It is believed that the MLM is superior to the BFM in detecting surface
waves from microtremor records. Therefore, this chapter follows the data an-
alysis sequence of the MLM.
Equation (3.33) gives P  (k, ω), the estimated f-k power spectrum. Fol-
lowing Capon (1969), this may be transformed as

P  (k, ω) = (aT Φa∗ )−1 (5.1)

or
 −1
N 
N

P (k, ω) = φij (ω) exp[ik(ri − r j )]
i=1 j=1

N 
N
= Ai∗ (k, ω) A j (k, ω) Ŝ ij (ω) exp[ik (ri − r j )], (5.2)
i=1 j=1

where

N
qij (k, ω)
j=1
Ai (k, ω) = , (5.3)

N 
N
qij (k, ω)
i=1 j=1

and qij (k, ω) is an element of the inverse matrix of matrix

{ Ŝ ij (ω) exp[ik(ri − r j )]}.

Ai (k, ω) is called the optimum weighting function which is a kind of opti-

mization filter. This filter passes the plane waves with a velocity corresponding
to the peak of the f-k spectra, and attenuates waves with other velocities.
SE803-05 September 30, 2003 16:4

5.2. DATA ANALYSIS 71

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Array
Observation

1
2
3
4
5
Data Set 6
7
8
9
10

POWER SPECTRUM

Calculate
FFT
Power and
AR
Cross Spectra

FREQUENCY (HZ)

Calculate BFM 2π f
F-K Spectra MLM V=
k2x + k2y

3 AZIMUTH
N
VELOCITY (KM/S)

2
Estimate
Velocities & 1
E

Azimuths
0
0 1 2 3 4
FREQUENCY (HZ) 4.0 HZ

Inversion
S Wave Velocity

Vs1
Geological Vs2
Depth

Structure
Vs3

Figure 5.2. Flow of data analysis by the f-k method.

SE803-05 September 30, 2003 16:4

72 CHAPTER 5. DATA ACQUISITION AND ANALYSIS METHODS

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

The vector a in equation (5.1) is of the form

aT = {exp (ikr1 ) , exp (ikr2 ) , . . . , exp (ikr N )} , (5.4)

and * denotes the complex conjugate. The parameters are:

N : number of stations
Ŝ ij (ω): estimated cross-spectrum between the signals at the ith and jth
stations for the angular frequency ω = 2πf
ri : The position of the ith station
The calculation of the cross-spectra of equation (5.2), as described in
Chapter 3, uses the direct Fourier transform, or the block averaging method.
Smoothing of the raw spectra is required for the direct Fourier transform. The
Parzen window, shown below, can be used for smoothing:

πu f 4
3  sin
W (f) = u 2  ,
πu f 
(5.5)
4
2
where u is a constant in seconds. Figure 5.3 shows the shape of this window.
In the figure, b is the resolution bandwidth and is given as:

1
b= ∞ . (5.6)
2
W ( f)df
−∞

1.0u

0.5u

0 f
2 1 1 2
−u −u 0 u u
b

Figure 5.3. Parzen window (from Ohsaki, 1976).

SE803-05 September 30, 2003 16:4

5.2. DATA ANALYSIS 73

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

a) b)
N N

Ky (rad/km)
Ky (rad/km)
96

96
48

48
00

00
W E W E
−96 −48

−48
00h30m 00h33m
204.8s 204.8s

−96
−96 −48 00 48 96 −96 −48 00 48 96
S Kx (rad/km) S Kx (rad/km)
c) d)
N N
Ky (rad/km)

Ky (rad/km)
96

96
48

48

W E W E
00

00
−48

−48

00h36m 00h30m
204.8s 614.4s
−96

−96

−96 −48 00 48 96 −96 −48 00 48 96

S Kx (rad/km) S Kx (rad/km)

Figure 5.4. The f -k spectra at 0.833 Hz. Data lengths are 204.8 s for (a), (b), and (c);
614.4 s for (d).

Smoothing of the raw spectra S( f ) is performed by

 ∞
Ŝ ( f ) = S ( f ) ·W ( f − g) dg. (5.7)
−∞

Figure 5.4 shows the f-k spectra at 0.833 Hz in frequency (Matsushima

and Okada, 1989) of microtremors observed by eight stations that formed an
array of 2 km by 1 km on the campus of Hokkaido University. Setting the
maximum amplitude to be 0 dB, and contoured to –12 dB with a contour
interval 1dB. This figure shows the temporal variation of the spectra by using
three data sets, each 204.8 s long, lagged by 3 min (a, b, and c) and the spectra
calculated from the longer 614.4-s record. The shorter records show some
variation in spectra, but the position of the peak is approximately the same
regardless of the time of observation. In this calculation, the resolution band
of the Parzen window was set to 20f ,where f is the frequency interval of
the cross-spectra.
SE803-05 September 30, 2003 16:4

74 CHAPTER 5. DATA ACQUISITION AND ANALYSIS METHODS

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 5.5. The f -k spectra of four different periods: 2.27, 1.86, 1.57, and 1.20 s
(corresponding frequency 0.440, 0.538, 0.637, and 0.833 Hz, respectively).

Figure 5.5 shows examples of the f-k spectra for four different periods,
2.27, 1.86, 1.57, and 1.20 s (i.e., 0.440, 0.538, 0.637, and 0.833Hz in frequency,
respectively) calculated from the same data as Figure 5.4. Such f-k spectra
can be calculated for a wider range of frequencies at a finer interval. Then the
phase velocity of the dominant wave with each frequency component can be
obtained by reading the coordinates of the peak spectrum (k x0 , k y0 ) of each
frequency in each plot. Note the frequency interval cannot be finer than the
resolution bandwidth of the smoothing window, 20 f .

5.2.2 Data analysis by spatial autocorrelation method

The work flow for data analysis by the spatial autocorrelation method is shown
in Figure 5.6. The theory of the analysis was explained in Chapter 3, and several
problems in practical data analysis are presented in this section.
SE803-05 September 30, 2003 16:4

5.2. DATA ANALYSIS 75

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Array
Observation

1
2
3
4
Data Set 5
6
7
8
9

Power Spectrum SPAC Function

Calculate
Power spectra,
SPAC Function

Frequency Frequency

F=1.51 HZ V=0.36 KM/S

1.0
CORRELATION

0.5
Calculate SPAC ρ( f, r) = J0(2πf r )
SPAC Coef. 0.0
V
ESPAC
−0.5
0.0 0.1 0.2 0.3 0.4
DISTANCE (KM)

1.5
VELOCITY (KM/S)

1.0
Estimate
Velocities 0.5

0.0
0.0 1.0 2.0 3.0
FREQENCY (HZ)
Inversion
S-Wave Velocity

Vs1
Geological Vs2
Depth

Structure
Vs3

Figure 5.6. Flow of data analysis by the SPAC and ESPAC methods.
SE803-05 September 30, 2003 16:4

76 CHAPTER 5. DATA ACQUISITION AND ANALYSIS METHODS

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

5.2.2.1 Standardization of the spatial autocorrelation function

In observing microtremors with a circular array, it is ideal for all the seis-
mometers and recorders to have identical properties, such as frequency re-
sponse. In addition, it is desirable that the conditions of installation of seis-
mometers, e.g., coupling with the ground, be the same. In practice, it is rare
to have such an ideal situation. Usually the frequency characteristics of seis-
mometers and recorders vary, and conditions of installation may be different
from place to place. When calculating the spatial autocorrelation function
under such circumstances, it is necessary to standardize the spatial autocor-
relation function (3.77) by the power spectra density function of each station.
In other words, the spatial autocorrelation coefficient should be calculated by

1 2π
Ŝ (ω0 , r, θ )
ρ (ω0 , r ) = dθ, (5.8)
2π 0
Ŝ0 (ω0 ) Ŝr (ω0 )

where

Ŝ (ω0 , r, θ ) = E[ X̂ (t, ω0 , 0, 0) X̂ (t, ω0 , r, θ )],

Ŝ 0 (ω0 ) = E[| X̂ (t, ω0 , 0, 0)|2 ],
Ŝr (ω0 ) = E[| X̂ (t, ω0 , r, θ )|2 ]. (5.9)

In equation (5.9), X̂ (t, ω0 , 0, 0) and X̂ (t, ω0 , r, θ ) are the records of com-

ponents with the frequency ω0 at the central station (0,0) and the station (r ,θ ),
respectively. Through the characteristics of the recording system and ground
conditions, the original microtremors at the ith station Xi is magnified by ai
times and recorded as X̂ i (= ai X i ), where ai is a constant independent of fre-
quency, and the spectral density function is assumed to be constant throughout
the array, i.e.,  
E |X (t, r, θ )|2 = const. (5.10)
In short, the spatial autocorrelation coefficient of equation (5.8) is a direc-
tional (or azimuthal) average of the coherency between the records at the center
and at the circumference of the circular array for a certain frequency ω0 .

5.2.2.2 The size of sample N and the analysis interval Tr

There is no standard for the size of sample, i.e., the length of the data segment
for analysis. It is chosen based on the required precision for the estimated
SE803-05 September 30, 2003 16:4

5.2. DATA ANALYSIS 77

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

value. It may be practical to decide these values by considering parameters

such as the highest expected frequency ( f N ), standard error of the power
spectrum (εr ), the resolution bandwidth desired in the power spectrum analysis
(Be ), etc. However, the size of the sample may be dictated by other factors
such as limitation in the length of the record and precision of observations.
Exceptional cases aside, the following relationships may be used as rules of
thumb in data processing, using the power spectra analysis method (Bendat
and Piersol, 1986):

5.2.2.2.a For calculating the estimated value of the power spectra from
the estimated autocorrelation function. Lag in correlation function: The
maximum lag number m is chosen by
1
m= (h = t), (5.11)
Be h
where Be is the desired resolution bandwidth in the power spectra analysis.
Sample size and time window for analysis: The size of the sample size N
is chosen by

m
N= , (5.12)
εr2

where εr is the standard error desired in the calculation. The corresponding

minimum window for analysis (i.e., minimum record length) Tr is
Tr = N h = N t. (5.13)
Degree of freedom and standard error: The degree of freedom n is given
by

2N
n = 2Be Tr = , (5.14)
m

and the standard error is

 
1 m
εr = = . (5.15)
Be Tr N

Therefore, when N is constant, a smaller m gives a smaller εr .

SE803-05 September 30, 2003 16:4

78 CHAPTER 5. DATA ACQUISITION AND ANALYSIS METHODS

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

5.2.2.2.b For calculating the estimated power spectra by the Fourier

transform. Here, the commonly used smoothing in the frequency domain
is assumed. When the spectra are white noise within a limited range, the esti-
mated spectra at the frequency interval of 1/Tr essentially have no correlation.
Therefore, when l frequency components near the estimated values of the raw
spectra are averaged, the smoothed estimate value of the spectra Ĝ k results in
1 
Ĝ k = Ḡ k + Ḡ k+1 + · · · + Ḡ k+l−1 . (5.16)
l
This Ĝ k is a χ 2 variable with degree of freedom approximately n = 2l by
the addition theorem of χ 2 of independent variables. The resultant effective
resolution bandwidth is approximately Be = lBe , where Be = 1/Tr . Therefore
this averaging in the frequency domain gives

Be = l Be = l/Tr = l f ; (5.17)

N = 2Be Tr = 2l. (5.18)

The standard error is  

1 1
εr = = . (5.19)
Be Tr l

The estimated value Ĝ k is considered to represent the midpoint of the

frequency range from fk to fk+l−1 . Altogether N /l such estimated values are
calculated.

5.2.2.3 Calculation of the phase velocity

When the spatial autocorrelation function method is used, the phase velocity
at the array center is obtained by equation (3.80). From x0 = 2π f 0r0 /c ( f 0 ),
the phase velocity for the frequency f0 is
2π f 0r0
c ( f0) = , (5.20)
x0
where x0 is the variable of the Bessel function of the first kind of zero order.
The value of the spatial autocorrelation coefficient ρ 0 is satisfied by x0 for
a certain frequency obtained from a dataset observed by a circular array of
radius r0 .
In this case, there may be a frequency range in which the spatial autocor-
relation coefficient does not satisfy the Bessel function due to errors in data
SE803-05 September 30, 2003 16:4

5.2. DATA ANALYSIS 79

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

MED02B R=100.3 m ST : 1–5 MED02B R=100.3 m ST : 1–5 MED02B R=100.0 m ST : 1–7

NDATA=4096 DT=0.020 s NUM=24 NDATA=4096 DT=0.020 s NUM=24 NDATA=4096 DT=0.020 s NUM=24

1.0 1.0 1.0

SPAC FUNCTION

0.5 0.5 0.5

0.0 0.0 0.0

−0.5 −0.5 −0.5

−1.0 −1.0 −1.0

0.0 1.0 2.0 3.0 4.0 0.0 1.0 2.0 3.0 4.0 0.0 1.0 2.0 3.0 4.0
FREQUENCY (Hz) FREQUENCY (Hz) FREQUENCY (Hz)

MEDO2B R = 100.2m
NDATA = 4096 DT=0.020 s NUM=24
1.0

0.5
SPAC COEF.

0.0

−0.5

−1.0
0.0 1.0 2.0 3.0 4.0
FREQUENCY (Hz)

Figure 5.7. Top: the integrand in equation (5.8) (coherency); bottom: the result of the
integration (5.8) (spatial autocorrelation coefficient).

acquisition and processing and insufficient power in the microtremors. Esti-

mation is not possible in such cases. To fill the gap in the phase velocity, it
may be necessary to reprocess using the data from a different time window,
or re-acquire data with a different array radius.

5.2.2.4 An example of spatial autocorrelation function

The data in this example were collected by the array shown in Figure 5.1. The
example shown in Figure 5.7 is for r = 100 m. The three graphs on the top
row of the figure show spatial covariance function g (ω, r, θ ) calculated from
equation (3.67), each corresponding to the function for different directions
from the center station to the three apex stations. These show coherency
given by the integrand of equation (5.8). Here, a standardization is applied for
plotting.
The spatial autocorrelation coefficient ρ (ω0 , r ) is shown in the bottom row.
Theoretically this should be the Bessel function of the first kind of zero order.
In the high-frequency range, the maxima in the amplitudes of ρ tends to
become smaller than the value of the Bessel function. This is because either
SE803-05 September 30, 2003 16:4

80 CHAPTER 5. DATA ACQUISITION AND ANALYSIS METHODS

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

the power of the microtremors diminishes in that frequency range, severely

reducing correlation, or it is affected by spatial aliasing due to the number of
stations on the circumference. To increase precision in this area, it is necessary
to employ a smaller array size or to increase the number of stations on the
circumference.

5.2.3 Case history of estimating phase velocity

As discussed, during data acquisition for analysis by the spatial autocorrela-
tion (SPAC) method, the arrangement of stations is often restricted, because
it requires a special circular array. On the other hand, the extended spatial
autocorrelation (ESPAC) method assumes that the spatial covariance function
at a given frequency is approximately constant in every direction, eliminat-
ing the restrictions in observation and increasing freedom in array design.
For example, a dataset collected for the f-k method can be analyzed by the
ESPAC method. This section shows an example of the application of the
ESPAC method to records of the vertical component of microtremors, col-
lected by both a circular array and an arbitrary array. The aim here was to
estimate the velocity of the Rayleigh wave with the ESPAC method, then
compare the result with the f-k method. Also presented is an example of sepa-
rating the Rayleigh and Love waves from the data acquired by a circular array
of three-component seismometers.

5.2.3.1 Analysis of vertical component data from the circular array

Three-component microtremor data were collected in the Tokachi Plain in
eastern Hokkaido, using three semicircular arrays of different radii (Figure 5.8,
top right). The vertical component is used for the analysis in this example. The
duration of observation is 45 min for the large array and 30 min for the medium
and small arrays. Each array contains nine seismometers arranged in equal
angle spacing of 22.5 degrees. The radii of the array are 180, 90, and 50 m.
In addition, two triangular arrays (array-L and array-S), of ten vertical-
component seismometers, were used for data analysis by the f-k method. The
size of the arrays is about 1100 and 400 m, respectively (Figure 5.8, bottom),
and the record length is 45 min.
Figure 5.9a shows a sample of the microtremor data collected by the large
semicircular array. Figure 5.9b is a 2500-s running power spectrum at Station 1
of the large array. From the running power spectrum, it is clear that the record
of microtremors is very stable for the duration of the 2500-s sample. The
SE803-05 September 30, 2003 16:4

5.2. DATA ANALYSIS 81

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

JAPAN SEA

HOKKAIDO

100 m
N

PACIFIC OCEAN

250 m

Array-L Array-S
N 3 N 10
200 m 100 m

10 1

4 9 3 1
2
8
7 8
6 4 2

5 7
9 5 6

Figure 5.8. Various arrays used in microtremor observation in the Tokachi Plain,
Hokkaido.

running power spectra of other stations, whose figures are not presented here,
show very similar characteristics.
Three graphs, a, b, and c, in Figure 5.10 are the power spectra of mi-
crotremors recorded with the large, medium, and small arrays, respectively.
They show that the power is characteristically high at about 0.4 Hz and 2.5 Hz,
and low around 1.0 Hz. Two band-pass filters, centered at 1.0 Hz and 2.0 Hz,
were separately applied to the records. The results are shown in Figure 5.11.
Notice that while (b), with a 2.0-Hz filter, shows waves with good correlation
among the stations, (a), with a 1.0-Hz filter, does not show obvious correlation.
This suggests difficulty in estimating the phase velocity near 1.0 Hz.
Figure 5.12 shows examples of fitting the Bessel functions of the first
kind of zero order to the r ∼ρ (r ) curves for f = 0.5, 1.5, and 2.5 Hz. In
the figure, a, b, and c correspond to small, medium, and large arrays, re-
spectively, while d is the sum of the three. As is well known, the precision
of the calculation in fitting a Bessel function to the observed data is improved
by increasing the number of stations.
SE803-05 September 30, 2003 16:4

82 CHAPTER 5. DATA ACQUISITION AND ANALYSIS METHODS

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ST. 1

ST. 2
ST. 3

ST. 4
ST. 5

ST. 6
ST. 7

ST. 8
ST. 9

100 150 200 250

a) Time (s)

0.0

1.0
FREQUENCY (HZ)

2.0

3.0

4.0

5.0
0 250 500 750 1000 1250 1500 1750 2000 2250 2500

b) Time (s)

Figure 5.9. (a) A part of the microtremor record for the large semicircular array; (b)
running power spectra of the whole record of the large array at station 1.

100 100 100

a) b) c)
POWER SPECTRUM

POWER SPECTRUM

POWER SPECTRUM

10 −1 10 −1 10 −1

10 −2 10 −2 10 −2

10 −3 10 −3 10 −3
0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
FREQUENCY (HZ) FREQUENCY (HZ) FREQUENCY (HZ)

Figure 5.10. Power spectra of the three semicircular arrays.

SE803-05 September 30, 2003 16:4

5.2. DATA ANALYSIS 83

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

a) f = 1.0Hz

ST. 1
ST. 2
ST. 3
ST. 4
ST. 5
ST. 6
ST. 7
ST. 8
ST. 9

100 150 200 250

Time (s)
b) f = 2.0Hz

ST. 1
ST. 2
ST. 3
ST. 4
ST. 5
ST. 6
ST. 7
ST. 8
ST. 9

100 150 200 250

Time (s)

Figure 5.11. A part of the microtremor record collected by the large semicircular
array processed through band-pass filters (the central frequencies are: (a) f = 1.0
Hz; (b) f = 2.0 Hz.

This method can use a combination of data from different arrays or

recorded at different times, as long as the station is not moved. An exam-
ple of combined data analysis using the records from three different arrays is
shown in Figure 5.13, together with the Bessel function of best fit to ρ (r ).
The velocity values shown on the top of the graphs are the estimated opti-
mum phase velocities. As these graphs show, and as expected from the power
spectra, the power is weak in the vicinity of 1.0 Hz (e.g., 0.8 ∼1.2 Hz). So
the result in this range is poor as the correlation is poor. However, one Bessel
function fits almost all the other frequency ranges.
SE803-05 September 30, 2003 16:4

84 CHAPTER 5. DATA ACQUISITION AND ANALYSIS METHODS

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

a) b) c) d)
1.0

CORRELATION
V = 0.31 km/s V = 0.90 km/s
0.5 V = 0.94 km/s V = 1.17 km/s
f = 0.5 Hz
0.0

−0.5

1.0
CORRELATION

V = 0.37 km/s V = 0.39 km/s V = 0.38 km/s V = 0.38 km/s

0.5
f = 1.5 Hz
0.0

−0.5

1.0
CORRELATION

V = 0.29 km/s V = 0.31 km/s V = 0.30 km/s V = 0.29 km/s

0.5
f = 2.5 Hz
0.0

−0.5
0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4
DISTANCE (KM) DISTANCE (KM) DISTANCE (KM) DISTANCE (KM)

Figure 5.12. An example of fitting the Bessel function of the first kind of zero order
to the spatial autocorrelation coefficient ρ(r ). (a) Small array, (b) medium array, (c)
large array, and (d) a combination of the three arrays.

F = 0.40 HZ V = 1.06 KM/S F = 0.60 HZ V = 0.91 KM/S F = 0.80 HZ V = 0.69 KM/S F = 1.00 HZ V = 0.45 KM/S
1.0
CORRELATION
CORRELATION

0.5

0.0

−0.5

F = 1.20 HZ V = 0.45 KM/S F = 1.40 HZ V = 0.40 KM/S F = 1.60 HZ V = 0.37 KM/S F = 1.80 HZ V = 0.36 KM/S
1.0
CORRELATION

0.5

0.0

−0.5

F = 2.00 HZ V = 0.34 KM/S F = 2.20 HZ V = 0.31 KM/S F = 2.40 HZ V = 0.30 KM/S F = 2.60 HZ V = 0.29 KM/S
1.0
CORRELATION

0.5

0.0

−0.5
0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4
DISTANCE (KM) DISTANCE (KM) DISTANCE (KM) DISTANCE (KM)

Figure 5.13. An example of combined processing of spatial autocorrelation coefficient

from three semicircular arrays of different sizes. Triangles, squares, and circles are the
spatial autocorrelation coefficients from the data recorded by the small, medium, and
large arrays, respectively. Each graph shows the r ∼ ρ(r ) curve for f = 0.4, 0.6, . . . ,
2.6 Hz. The solid line is the Bessel function of best fit to ρ(r ).
SE803-05 September 30, 2003 16:4

5.2. DATA ANALYSIS 85

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

a) b) c) d)
1.5
VELOCITY (KM/S)

1.0

0.5

0.0
0.0 1.0 2.0 3.0 0.0 1.0 2.0 3.0 0.0 1.0 2.0 3.0 0.0 1.0 2.0 3.0
FREQUENCY (HZ) FREQUENCY (HZ) FREQUENCY (HZ) FREQUENCY (HZ)

Figure 5.14. Phase velocity estimated by the ESPAC method, obtained by fitting
Bessel functions to the ρ(r ) values from the data collected by the small array (a),
medium array (b), large array (c), and combination of the three arrays (d).

In the above example, a Bessel function was fitted to the observed data
by the least-squares method, as it is theoretically expected and justified. But,
as seen in the examples, around 1Hz, the correlation coefficient generally
contains some noise. With the expectation that such cases will occur, future
processing methods will require a quantitative method to assess the appropri-
ateness of such solutions, and this is an example in which an inappropriate
value of velocity is obtained at 1Hz, when the spectral power is low.
As seen in the above example, an important characteristic of the ESPAC
method is that it can easily detect an inappropriate value, which is impossible
in the f-k method.
Figure 5.14 shows the estimated phase velocities obtained by fitting Bessel
functions to the ρ(r ) values from the data collected by the arrays of various
sizes.
Figure 5.15 is a comparison of these phase velocities estimated by the
ESPAC method (circles) and by the f-k method (triangles). In the figure,
the ESPAC result is an estimate from the combination of the three arrays, while
the results from the f-k method for the small and large arrays are shown sepa-
rately. They are overplotted in the bottom row. From the figure, we can see that

i) the results are coincident in the frequency range f ≥ 1.2 Hz;

ii) both methods gave poor results near 1.0 Hz (0.8 Hz ≤ f ≤ 1.2 Hz),
perhaps due to the lower power;
iii) in the range f ≤ 0.8 Hz, the phase velocity from array-S (the size of
which is similar to the large semicircular array) via the f-k method is a
SE803-05 September 30, 2003 16:4

86 CHAPTER 5. DATA ACQUISITION AND ANALYSIS METHODS

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

1.5 1.5 1.5

VELOCITY (KM/S) F-K(Array-S) F-K(Array-L) ESPAC

1.0 1.0 1.0

0.5 0.5 0.5

0.0 0.0 0.0

0.0 1.0 2.0 3.0 0.0 1.0 2.0 3.0 0.0 1.0 2.0 3.0
FREQUENCY (HZ) FREQUENCY (HZ) FREQUENCY (HZ)

1.5 1.5
(a) (b)
VELOCITY (KM/S)

VELOCITY (KM/S)
ESPAC ESPAC
1.0 F-K(Array-S) 1.0 F-K(Array-L)

0.5 0.5

0.0 0.0
0.0 1.0 2.0 3.0 0.0 1.0 2.0 3.0
FREQUENCY (HZ) FREQUENCY (HZ)

Figure 5.15. Comparison of phase velocities estimated by the ESPAC method and by
the f -k method.

larger estimate than the one via the ESPAC method (c ≥ 1.5 km/s). The
reason for this is believed to be that this frequency range is outside the
limit for which phase velocity can be estimated; a limit imposed by the
size of the array.

By using Array-L (the size of which is about three times as large as

the large semicircular array), a result similar to the ESPAC method is ob-
tained. This demonstrates the superiority of the ESPAC method in estimating
subsurface structure over the f-k method for arrays of the same size. This is
concordant with the numerical simulation by Ling (1994).
In calculating ρ(r ) by equation (4.5), it is desirable to have an array
with many widely spaced stations. A large number of stations allows a large
number of pairs of stations with varying distance r and a wider distribu-
tion of azimuth. As a result, the subsurface structure can be estimated more
SE803-05 September 30, 2003 16:4

5.2. DATA ANALYSIS 87

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

50 M

Figure 5.16. All the possible combinations of two stations.

precisely. In this case history, we had enough instruments for nine stations.
For this array, a diagram of all the possible pairs of stations is shown in Fig-
ure 5.16. Table 1 summarizes the number of pairs and the azimuths for each
possible distance between pairs of stations.
As seen in Table 1, the semicircular array provides a large number of pairs
separated by the distance of the radius (r = 180 m) and a large number of
pairs of adjacent points (r = 70 m); the array’s azimuth coverage approaches
180 degrees. Where the distance between stations is larger than the radius of
the array, there are fewer possible combinations and the azimuth coverage is
narrow. As a result, the accuracy of the spatial autocorrelation coefficient may
suffer.

Table 5.1. All the possible combinations of two stations, distance between
the stations, number of combinations and azimuth of the line connecting the
stations.

r(m) N Angle (degree)

70 7 −56.25 −33.75 −11.25 11.25 33.75 56.25 78.75
138 6 −45.00 −22.50 0.00 22.50 45.00 67.50
180 8 22.50 45.00 67.50 90.00 112.50 135.00 157.50 180.00
200 5 −33.75 −11.25 11.25 33.75 56.25
255 4 −22.50 0.00 22.50 45.00
299 3 −11.25 11.25 33.75
333 2 0.00 22.50
353 1 11.25
SE803-05 September 30, 2003 16:4

88 CHAPTER 5. DATA ACQUISITION AND ANALYSIS METHODS

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

1.00 f = 0.5 Hz 0.25 f = 1. 5 Hz

R = 0.353 KM 0.0 0.0
1.00 0.25
R = 0.333 KM 0.0 0.0
1.00 0.50
R = 0.299 KM 0.0 0.0
1.00 0.25
R = 0.255 KM 0.0 0.0
1.00 1.00
R = 0.200 KM 0.0 0.0
1.00 1.00
R = 0.180 KM 0.0 0.0
1.00 1.00
R = 0.138 KM 0.0 0.0
1.00 0.50
R = 0.070 KM 0.0 0.0

−90 0 90 180 −90 0 90 180

AZIMUTH (DEC.) AZIMUTH (DEC.)

Figure 5.17. Spatial autocorrelation function calculated for all the combinations of
stations for 0.5 and 1.5 Hz (sorted by distance).

Figure 5.17 shows the autocorrelation functions at f = 0.5 and 1.5 Hz

plotted against the distance r of all the combinations of the stations. The dotted
lines in the figure indicates the averages. Where an autocorrelation value is
absent due to insufficient combinations for a particular r, the dotted line of the
average assumes the missing value. This problem of dependency of azimuth
coverage on r may be eliminated by using a circular array.
As discussed before, the ESPAC method allows the combination of several
data sets obtained at different times. The example here demonstrates that the
data sets observed at different times by the three arrays, small, medium and
large, could lead to a Bessel function that fits a curve representing the spatial
autocorrelation coefficient for the corresponding frequency. This shows that
several data sets separated by a considerable time lag could be jointly analyzed
by the ESPAC method. This is a merit of the ESPAC method that has not been
available in previous methods. With the ESPAC method, even with only a
few seismometers available, one can obtain the same result as that found with
many seismometers, by repeating observations.

5.2.3.2 Analysis of data from the arbitrary array

The regular semicircular array discussed in the previous section may not be
easily arranged at the desired site. In such a case, an arbitrary array may be
SE803-05 September 30, 2003 16:4

5.2. DATA ANALYSIS 89

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

1 OMB
6

9
12 11 3
13 8
4
10
7

200 M
2

Figure 5.18. Arrangement of the stations of array OMB.

used by virtue of the ESPAC method. A range of distances [r0 −  r, r0 +

 r ] is, in this case, represented by r0 . This section is a case history of this
application.
Figure 5.18 is a diagram of the observation array used in Oma, Aomori
Prefecture, at the northern tip of Honshu Island. An analysis of these same
data with the f-k method has been shown in Chapter 3. A separate analysis
by the ESPAC method is demonstrated here. For arbitrary arrays like this, the
conventional SPAC method cannot be applied.
Figure 5.19 shows r ∼ ρ(r ) curves for f = 0.4, 0.6, . . . 2.0 Hz.
Figure 5.20 is a comparison between the ESPAC method and the f-k
method. It is clear from the figure is that the two results are almost identical
for frequencies higher than 1.0 Hz, and the f-k method gives considerably
higher velocity at frequencies below 1.0 Hz. It is considered that the ESPAC
method gives a velocity closer to the real value. However the real velocity is
not independently known at this stage.
From the above, the following conclusions for the ESPAC method were
reached:

1. The ESPAC method gives the same result as the f-k method in the higher
frequency range.
2. The ESPAC method is superior to the f-k method in that it
a) gives a good stable result for the lower frequency ranges,
b) the computation of the correlation function is simpler, and suitable
for a PC,
SE803-05 September 30, 2003 16:4

90 CHAPTER 5. DATA ACQUISITION AND ANALYSIS METHODS

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

F = 0.40 HZ V = 1.820 F = 0.60 HZ V = 2.030 F = 0.80 HZ V = 1.930

1.0
CORRELATION

0.5

0.0

−0.5

F = 1.00 HZ V = 1.700 F = 1.20 HZ V = 1.230 F = 1.40 HZ V = 0.910

1.0
CORRELATION

0.5

0.0

−0.5

F = 1.60 HZ V = 0.770 F = 1.80 HZ V = 0.670 F = 2.00 HZ V = 0.610

1.0
CORRELATION

0.5

0.0

−0.5
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
DISTANCE (KM) DISTANCE (KM) DISTANCE (KM)

Figure 5.19. An example of fitting the Bessel function of the first kind of zero-order
to the spatial autocorrelation coefficient from an arbitrary array. Symbols: the cross
is the spatial autocorrelation coefficient calculated for each ri ; the circle is the spatial
autocorrelation coefficient calculated and averaged over the distance [r0 − r, r0 +
r ], where r is 40 m; the solid line is the best-fit Bessel function to ρ(r ).

3.0 3.0 3.0

F-K ESPAC F-K
VELOCITY (KM/S)

VELOCITY (KM/S)

VELOCITY (KM/S)

ESPAC
2.0 2.0 2.0

1.0 1.0 1.0

0.0 0.0 0.0

0.0 1.0 2.0 3.0 0.0 1.0 2.0 3.0 0.0 1.0 2.0 3.0
FREQUENCY (HZ) FREQUENCY (HZ) FREQUENCY (HZ)

Figure 5.20. Comparison of phase velocities estimated by the ESPAC method and by
the f-k method (data acquired by the array OMB).
SE803-05 September 30, 2003 16:4

5.2. DATA ANALYSIS 91

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

c) a reasonable result can be obtained with fewer stations,

d) several data sets collected at different times can be jointly processed,
and
e) inappropriate values as calculated by the f-k method can be detected
by the ESPAC method.

On the other hand, the ESPAC method cannot detect the direction of the
source of the wave. Also, if there are surface waves of higher mode mixed in
with the data, the ESPAC method cannot detect it. In these respects, the f-k
method is superior.

5.2.3.3 Analysis of data from the circular array with three-component

seismometers (identification of the Love wave)

In Section 3.5, we discussed how the SPAC method can detect the Love wave
from three-component data while the f-k method cannot, and the theory and
the method was presented in Section 3.6. This section shows an example of
its application to long-period microtremor data in order to identify Love wave
energy (Matsushima and Okada, 1990b).
The data were collected in a suburb of Hachinohe City, Aomori Prefec-
ture. Two datasets were collected with arrays of two different radii, 266 m
and 501 m. Each array comprised four stations of three-component seis-
mometers; one at the center and three on a circle with a 120-degree inter-
val. The horizontal-component seismometers are set to record north-south
and east-west components. Each dataset was collected for 46 min at night.
The relative location of the arrays is shown in Figure 5.21. The seismome-
ters are models PELS74V for the vertical component, and PELS74H for the
horizontal component. Both have a natural period of 8 s.
The data were converted to the radial and tangential components, i.e., X r
and X θ in the left-hand side of equation (3.81), respectively. Applying narrow
band-pass filters with various central frequencies to these and the vertical
component data, the spatial autocorrelation coefficients, ρr and ρθ of equation
(3.96), were calculated.
Figure 5.22 shows the autocorrelation coefficients of the vertical, radial
and tangential components of the data for the array of 868-m diameter (which
is the length of the sides of the triangle of the circle with the diameter of
SE803-05 September 30, 2003 16:4

92 CHAPTER 5. DATA ACQUISITION AND ANALYSIS METHODS

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

N
2’

1’

1 3’
3

0 500 m

Figure 5.21. Arrangement of three-component seismometers to form two equilateral

triangle arrays. The radii of the arrays are 266 m and 501 m.

501 m), plotted against the central frequency of the band-pass filter. The
different symbols indicate the different components.
Figure 5.23 is the relationship between the phase velocity of the Rayleigh
wave and its period calculated from the vertical component of microtremor
records by various array radii using equations (3.78) or (3.80). In this graph
the horizontal axis is converted to period from frequency. The symbols in the
figure correspond to the radii of the array from which the phase velocities were
calculated. The subsurface structure estimated from the relationship between
the phase velocity and period is shown in Table 5.2.

1.0 SPATIAL AUTOCORRELATION COEFFICIENTS

R=868m

1.0 1.2
0.0
0.2 0.6
FREQUENCY (Hz) VERTICAL
RADIAL
TANGENTIAL
−1.0

Figure 5.22. Spatial autocorrelation functions for an array with a radius of 868 m.
SE803-05 September 30, 2003 16:4

5.2. DATA ANALYSIS 93

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

RAYLEIGH WAVE
2

PHASE VELOCITY (km/s)

1 RADIUS
266 m
460 m
501 m
868 m

0 1 2 3
PERIOD (s)

Figure 5.23. Phase velocities of Rayleigh waves obtained from the vertical component
of microtremors. The solid line indicates the theoretical phase velocity curve of
Rayleigh waves calculated for the structural model in Table 5.2.

The curve in Figure 5.23 is the phase velocity of the fundamental-mode

Rayleigh wave theoretically calculated for the subsurface structure model of
Table 5.2.
Figure 5.24 shows the phase velocity of the Love wave plotted against its
period, calculated by equation (3.98) from the spatial autocorrelation function
[equations (3.93), (3.94), and (3.95)], and using the parameter λ (ω, r ) derived
from equation (3.100).
The curve in Figure 5.24 is the phase velocity of the fundamental-mode
Love wave calculated from the S-wave velocity of the subsurface structure of

Table 5.2. Model parameters of geological structure at the test site esti-
mated by an inversion technique using phase velocities of Rayleigh waves
in microtremors.

S-wave P-wave Density

Layer Thickness (m) velocity (km/s) velocity (km/s) (g/cm)
1 44 0.15 1.58 1.70
2 1256 1.50 3.00 2.22
3 ∞ 2.40 4.27 2.43
SE803-05 September 30, 2003 16:4

94 CHAPTER 5. DATA ACQUISITION AND ANALYSIS METHODS

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

LOVE WAVE

PHASE VELOCITY (km/s)

2

1 RADIUS
266 m
460 m
501 m
868 m

0 1 2 3
PERIOD (C)

Figure 5.24. Phase velocities of Love waves obtained by the extended SPAC method.
The solid line indicates the theoretical phase velocity curve of Love waves calculated
for the structural model in Table 5.2.

Table 5.2. This curve generally gives slower velocities than the ones calculated
from the observation. However, the general trends of the calculated and ob-
served data are consistent; in particular we see the steep increase in phase
velocity through the periods 1 to 1.2 s, and the gradual increase with the
periods greater than 1.3 s.
 The relationship between period and the power  fraction of Love waves
= h 0 (ω) /H0 (ω) [see equations (3.95) to (3.97)] in microtremors is shown
L

in Figure 5.25. At the period 1 s, the power of the Love wave is about 50%,
and the fraction increases at longer periods.
100
POWER (%)
50

RADIUS
266 m
460 m
501 m
868 m
0 1 2 3
PERIOD (s)

Figure 5.25. Power fractions of Love waves in the horizontal component of

microtremors.
SE803-05 September 30, 2003 16:4

5.2. DATA ANALYSIS 95

Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

AMP. RATIO (H/V)

10

RADIUS
266 m
1 460 m
501 m
.5 868 m

0 1 2 3
PERIOD (s)

Figure 5.26. Amplitude ratios of vertical and horizontal components of Rayleigh

waves. The solid line indicates the theoretical curve calculated for the structural
model in Table 5.2.

Figure 5.26 shows the amplitude ratio between vertical and horizontal
components (H/V) computed from the ratio of the powers of the vertical
component and of the Rayleigh wave contained in the horizontal component.
The variation of the amplitude ratio against period generally agrees with
the values calculated from the estimated structure (inserted curve), although
the peak in the data from observation at 1.1 s slightly offsets from the calculated
peak at 1.23 s.
As seen, this comparison between the power ratio or amplitude ratio of
the Rayleigh wave can be used to improve or evaluate the estimated subsur-
face structure model. This demonstrates an advantage of the SPAC method
achievable by using three-component data.
 Downloaded 09/30/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
SE803-05
September 30, 2003
16:4

96