Documentos de Académico
Documentos de Profesional
Documentos de Cultura
_
V
F(t t
, r r
)G(t
, r
)dt
dr
(1.1)
(Sato and Fehler, 1998, p. 175).
F(t, r) =
1
4v
0
r
2
H(t)(t r/v
0
)e
v
0
got
(1.2)
1.2. ENVELOPE MODELING 11
is the Greens function for the coherent wave energy. Here v
0
and g
0
denote average
velocity of S-waves and total scattering coecient respectively. r = [r[ is the source
receiver distance and H is the Heaviside step function. The exact solution of eq. 1.1
involves a two dimensional Fourier transform (Zeng et al., 1991; Sato and Fehler,
1998, p. 177).
To speed up the inversion an analytic approximation to the solution of 1.1 is
used that was obtained by Paasschens (1997) by interpolating between the explicit
solutions of the Boltzmann equation in 2D and 4D. The solution of Paasschens
(1997) reads
G(t, r)
e
v
0
tg
0
4r
2
(r v
0
t) +
(1 r
2
/(v
2
0
t
2
))
1/8
(4v
0
t/(3g
0
))
3/2
e
v
0
tg
0
K
_
v
0
tg
0
_
1
r
2
v
2
0
t
2
_
3/4
_
H(v
0
t r) (1.3)
where K(x) e
x
_
1 + 2.026/x. Recently this interpolation was used by Abubakirov
(2005) in a MLTWA. It is a good approximation to the solution of eq. 1.1 given by
Zeng et al. (1991). The deviation from the exact solution is below 5% where larger
deviations occur in the tail of the direct wave. The accuracy of the approximation
for the direct wave and its tail is reasonable in the present context because only an
average value in a short time window following the direct wave will be used. This
time window is largely dominated by the energy of the unscattered wave which is
correctly described in equation 1.3.
The boundary condition for radiative transfer in a half-space is zero vertical
energy ux at the surface. Assuming total reection, this condition can be accounted
for by introducing a mirror source above the surface for which the energy density
Greens function is G(t, r
+
), with r
+
being the distance between receiver and mirror
source. At the surface the upward ux of the real source equals the downward
ux of the mirror source, thereby satisfying the boundary condition. The energy
density Greens function of the half-space is thus G
h
(t, x) = G(t, r) + G(t, r
+
). In
the following it is assumed that the receiver is at the surface where r = r
+
. In this
case G
h
(t, x) = 2G(t, r) (c.f. Wegler, 2004). With this treatment only the boundary
condition for energy transfer is considered. Specic eects arising from boundary
conditions of the wave equation like mode conversions, surface waves and angle
dependent reection coecients are not accounted for.
The energy density for an arbitrary source can be obtained by convolution
with the source function. In this analysis I assume a source function of the form
W(r)(t) where W is the spectral source energy in J/Hz. Intrinsic absorption can
be accounted for with an additional time dependent factor e
bt
with the intrinsic
absorption parameter b. Finally I obtain
E
mod
(t, x) = WR
i
G
h
(t, x)e
bt
(1.4)
= 2WR
i
G(t, r)e
bt
(1.5)
for the energy density E
mod
(t, r) of the half-space model at time t and distance r
from the source of energy W. R
i
is the site response at station i.
12 CHAPTER 1. ACOUSTIC RADIATIVE TRANSFER THEORY
13-4-1992
20-6-1995
21-10-1997
29-11-1997
23-6-2001
22-7-2002
22-2-2003
22-3-2003
22-5-2004
20-10-2004
5-12-2004
BFO
BRNL
BUG
CLZ
FUR
GRB1
GRC1
HAM
TNS
WET
GRFO/GRA1
BRG
CLL
MOX
STU
GSH
GEC2
IBBN
BSEG
RUE
HLG
RGN
NRDL
UBBA
4
4
8
8
12
12
16
16
48 48
50 50
52 52
54 54
Figure 1.1: Map of Germany and neighboring countries. Gray triangles: broadband
stations of the German Regional Seismic Network, the Gr afenberg array and the GERESS
station GEC2 (station codes are added); Black stars: earthquakes used in this study (dates
of occurrence are added).
1.3 Data and regional setting
Data for this study was recorded by the German Regional Seismic Network (GRSN),
the Gr afenberg array (GRF) and the reference station of the GERESS-array. The
stations are equipped with STS-2 or STS-1 (GRF stations) broadband seismometers
and traces are sampled at 80 Hz. Refer to Korn (2002) for a detailed discussion from
the GRSN. For this work data of 25 dierent stations is used (gure 1.1).
As sources I use local and regional events from Germany and adjacent areas in
the period from the installation of the GRSN in 1991 until December 2004. Events
with local magnitudes estimated by the German Central Seismological Observatory
(SZGRF) larger or equal to 4 are chosen. To reduce scatter in the estimates of the
parameters events from the Alps are excluded because of dierences in the geological
conditions. One event that occurred on 18 October 1994 in the North Sea is excluded
1.4. INVERSION 13
because I observed suspicious dierences in the arrival times, perhaps due to location
errors or an unsuitable velocity model. In the end 11 earthquakes (gure 1.1) are
used and more than 100 3-component broadband records. Locations, magnitudes
and depths as provided by the SZGRF are listed in table 1.1. With the possible
exception of the Roermond 1992 earthquake, where depths even below 20 km have
been estimated (Korn, 2002), the events occurred in the upper crust. The last three
columns of table 1.1 contain seismic moments estimated by waveform modeling and
the method presented here. They will be discussed later.
Preparation of data involves ltering of the seismograms in eight frequency
bands centered at 0.1875, 0.375, 0.75, 1.5, 3.0, 6.0, 12.0 and 24.0 Hz with a narrow
normalized Gaussian lter. Choosing the lter such that
_
[B(f)[
2
df = 1 where
B(f) is the frequency response, ensures that the spectral energy density is preserved
in the lter process (Wegler et al., 2006a). Envelope sections are selected according
to the following criteria:
1. S/N - ratio greater than 4
2. No obvious disturbances e.g. aftershocks (checked manually)
3. Hypocentral distance greater than 60 km.
The hypocentral distance range is limited to distances greater than 60 km because of
the simple half space velocity model, with a mean shear wave speed v
0
= 3.5 km/s,
which is inappropriate for shorter distances.
The observed energy density E
obs
(t, r), i.e. the seismogram envelopes, is ob-
tained from the bandpassed velocity seismogram u(t, r) using
E
obs
(t, r) =
0
2
u
2
(t, r) +H
2
( u(t, r))
2
. (1.6)
Here
0
is the mean density of the medium and H denotes the Hilbert transform.
0
= 2700 kg/m
3
is used throughout this study.
1.4 Inversion
In the inversion scheme I estimate values for the parameters g
0
, b, W and R
i
that
minimize the mist function
=
N
i=1
end
i
j=start
i
_
log
_
E
obs
(t
j
, r
i
)
E
mod
(t
j
, r
i
)
__
2
. (1.7)
Here N denotes the number of stations. start
i
and end
i
correspond to the indices
of the rst and the last sample in the coda time window (gure 1.2) at the i th
station. The coda time window starts after twice the S-wave travel time and ends
when the S/N drops below 4. All available traces of a certain event in one frequency
band are simultaneously inverted for g
0
, b, W and R
i
. The inversion scheme is a
combination of a one dimensional grid search for g
0
and a linear inversion for the
remaining parameters. Figure 1.2 illustrates the constraints of this process which
works in the following ve steps.
14 CHAPTER 1. ACOUSTIC RADIATIVE TRANSFER THEORY
0 100 200 300 400
10
15
10
10
10
5
direct
Swave seismic coda
lapse time [s]
E
[
J
m
3
H
z
1
]
envelope
model
inversion
time windows
Figure 1.2: Example of an observed seismogram envelope (thin curve) on a logarithmic
scale with the best tting model (bold black curve). Time windows used in the inversion
are indicated at the bottom of the gure. Note that in the time window of the direct
S-wave only the average values (black dot) of envelope and model are tted.
1. A value of g
0
is picked (a range of g
0
is probed in a grid search) and the
energy density Greens functions G(t, r
i
) is calculated for all stations i, using
the approximation in eq. 1.3.
2. I equate E
obs
(eq. 1.6) and E
mod
(eq. 1.5) and rearrange terms like
ln
_
E
obs
(t
j
, r
i
)
G(t
j
, r
i
)
_
= ln (2WR
i
) bt
j
. (1.8)
Here ln denotes the base e logarithm. By tting a linear curve to the left
hand side of eq. 1.8 as function of time the appropriate values of b and (WR
i
)
are obtained. To reduce the eect of direct surface waves and anisotropic
scattering the graph is only tted in the coda time window (gure 1.2)
starting at twice the S-wave travel time. Additionally I apply the constraint
that average values of the modeled and observed envelopes within the direct
S-wave time window (gure 1.2) are equal. At distance r from the source this
time window contains the samples with lapse time t in the range 3.5 km/s >
r/t > 3.0 km/s. This constraint is applied to resolve the tradeo between g
0
and (WR
i
) that exists in the later coda.
3. With the values of b and (WR
i
) obtained in step two E
mod
is calculated ex-
plicitly (eq. 1.5) to estimate the mist (eq. 1.7).
4. Steps one to three are repeated for a range of g
0
values. The nal model for
one event and one frequency band is that with the smallest value of .
5. For the best tting model W and R
i
are calculated explicitly by simply taking
W to be the logarithmic average of all (WR
i
) values. This way it is implicitly
1.5. RESULTS 15
assumed that the average of the logarithms of the R
i
is zero. This means
that site amplications are measured relative to the network mean which is
assumed to be one.
The constraint that makes use of the direct S-wave time window originates
from the following problem. In a weakly scattering medium the dierence between
the coda of two models with dierent W and g
0
can be very small provided the
product Wg
0
is equal for both models. The single scattering approximation (Sato
and Fehler, 1998, p. 47) which is valid for small g
0
predicts that the energy density
E
mod
is proportional to Wg
0
. In this range it is inherently impossible to separate
W and g
0
. But also for moderate scattering that is not adequately modeled in the
single scattering approximation it can be practically impossible to separate W and
g
0
due to noisy data and other simplications in the model if only information from
the coda is used.
In contrast the ballistic energy density is proportional to We
(v
0
g
0
+b)t
. Here
a tradeo is observed between g
0
and b which corresponds to the known fact that
intrinsic and scattering attenuation can not be separated using unscattered energy
only. However, the energy W and total attenuation can be separated because of the
dierent time dependence.
In the inversion the envelopes are tted in the coda time window indicated
in gure 1.2. But I restrict the possible models to these which correctly predict the
ballistic energy in the direct S-wave time window (gure 1.2). I want to stress
that in a weakly scattering medium it is on the one hand not possible to reliably
estimate W from the coda alone without a priory knowledge of g
0
. On the other
hand it is impossible to estimate W solely from the ballistic wave because in order
to apply radiative transfer theory which provides the source energy W I need to
separate g
0
and b. I note that information from the unscattered wave is implicitly
evaluated also by MLTWA to separate g
0
and b since the rst time window of the
MLTWA generally contains the ballistic energy.
1.5 Results
As all available recordings are inverted simultaneously I implicitly assume that there
are no lateral variations in the structural parameters g
0
and b. The results represent
integral values averaged over possible regional variations. An example of the t that
is achieved with the model is given in gure 1.3. The observed envelopes are plotted
as thin lines whereas bold lines represents the model envelopes. Average energy
density values in the direct S-wave time window are indicated as gray and black
dots for observed and modeled envelopes respectively. The model generally ts the
envelopes in the coda time window. Deviations in the direct S-wave window
are probably due to directionality of the source.
In the following I will rst present the results of the attenuation parameters
as they are fundamental for the estimated source parameters that are presented
thereafter. I will also briey mention the estimates of the site amplications.
16 CHAPTER 1. ACOUSTIC RADIATIVE TRANSFER THEORY
10
16
10
14
10
12
10
10
10
8
BFO
r=123km
0 200 400 600
BRG
r=594km
lapse time [s]
0 200 400 600
BSEG
r=671km
10
16
10
14
10
12
10
10
10
8
BUG
r=346km
10
16
10
14
10
12
10
10
10
8
CLZ
r=469km
FUR
r=342km
0 200 400 600
GEC2
r=520km
GRA1
r=364km
10
16
10
14
10
12
10
10
10
8
GRB1
r=382km
e
n
e
r
g
y
d
e
n
s
i
t
y
E
[
J
m
3
H
z
1
]
GRC1
r=363km
GRFO
r=364km
IBBN
r=446km
MOX
r=439km
0 200 400 600
10
16
10
14
10
12
10
10
10
8
RUE
r=681km
STU
r=191km
TNS
r=245km
WET
r=464km
Figure 1.3: Comparison between the envelopes (thin curves) of the 22-2-2003 earthquake
with the model (bold black curve) in the 1.5 Hz band on a logarithmic scale. Time windows
used in the inversion are indicated with gray bars at the bottom of the plots. Note that
in the time window of the direct S-wave only the average values of envelope and model
are tted. These levels are indicated as gray and black dots for envelope and model
respectively.
1.5. RESULTS 17
1.5.1 Attenuation parameters
Scattering strength can be expressed in terms of the total scattering coecient g
0
which is the inverse of the mean free path or in terms of scattering attenuation
parameter
Sc
Q
1
. The relation between both is
Sc
Q
1
=
g
0
v
0
2f
(1.9)
where f denotes frequency. Figure 1.4 displays g
0
as function of frequency. Small
gray dots in gure 1.4 indicate measurements of individual earthquakes. The scat-
ter of these values is due to uncertainties of the measurements but additionally
regional dierences of the scattering strength will increase the variations in the g
0
values. Black dots with error bars indicate the logarithmic averages of the individ-
ual measurements together with their 95% condence limits. Individual measure-
ments range between 10
7
m
1
and 10
5
m
1
. Logarithmic averages range between
910
7
m
1
and 310
6
m
1
. There is no signicant frequency dependence of g
0
and
logarithmically averaged over all measurements irrespective of frequency a value of
1.45 10
6
m
1
is obtained corresponding to a mean free path of 690 km. I compare
my estimates with Abubakirov and Gusev (1990); Fehler et al. (1992); Mayeda et al.
(1992) and Lacombe et al. (2003) who studied coda waves in Kamchatka, Japan,
Hawaii/Long Valley/central California and France respectively. Table 1.2 summa-
rizes the results of the dierent studies in frequency bands centered at 3.0 Hz. Refer
to Bianco et al. (2002) for graphical representation of most of these results. The
value obtained in the present study is low in this comparison but the dierence can
be attributed to geological distinctions since the central European intraplate region
is likely less heterogeneous than the volcanic areas listed in table 1.2. Compared
to the results from France obtained by Lacombe et al. (2003) the dierence might
partially be due to a dierent model setup.
The intrinsic attenuation parameter
I
Q
1
is related to the absorption parameter
b as
I
Q
1
=
b
2f
. (1.10)
Figure 1.5 displays the results of individual measurements with small gray dots and
logarithmic averages in the individual frequency bands together with 95% condence
limits as black dots with error bars. Averaged over individual measurements in
the separate frequency bands values of
I
Q
1
between 2.6 10
4
and 2.2 10
3
are
obtained. For frequencies below 3 Hz
I
Q
1
is approximately constant at 2 10
3
.
Above approximately 3 Hz a power law decrease like
I
Q
1
f
1
is observed. These
values are within the range of results from dierent regions listed in table 1.2 and
very close to the results from France Lacombe et al. (2003). However, there may
be some bias because regional records are inverted with a half space model. I
probably overestimate
I
Q
1
because leakage of energy through the Moho into a
weakly scattering mantle is mapped into attenuation (Margerin et al., 1998; Wegler,
2004).
18 CHAPTER 1. ACOUSTIC RADIATIVE TRANSFER THEORY
10
1
10
0
10
1
10
8
10
7
10
6
10
5
10
4
f [Hz]
g
0
[
m
1
]
individual events
mean with
95% confidence limits
Figure 1.4: Results for the total scattering coecient of S-waves. Gray dots: scatter-
ing coecients measured by inversion of individual events; Black dots with error bars:
logarithmic mean of individual measurements with 95% condence limits.
1.5.2 Source energy
Since the present approach is on one hand based on a physical model of energy
propagation and on the other hand does not require a coda normalization, the
inversion parameter W can be interpreted directly. By measuring the source energy
at various frequencies I obtain the source energy spectrum W() of radiated S-
waves. Assuming particle motion is caused by a double couple and is observed in
the far eld one can obtain a relation between W() and the seismic moment by
integrating the energy ux density over a sphere containing the source. Sato and
Fehler (1998, p. 152) state
W() =
4
[M()[
2
10
0
v
5
0
(1.11)
where v
0
,
0
and M() denote mean S-wave velocity, mean density, and the Fourier
transform of the moment time function respectively. From eq. 1.11 one obtains the
source displacement spectrum [M()[.
In gure 1.6 the source displacement spectra of the events listed in table 1.1
are shown. Error bars correspond to the resolution of the inversion process and
not to the scatter of individual measurements as in gure 1.4 and 1.5. The mist
in eq. 1.7 is the variance of the logarithmic dierence between the samples of
the observed and modeled energy densities. I use the F-test to decide whether two
models are signicantly dierent based on the ratio of their variances . The error
bars mark the range of values for the displacement spectra that can be obtained for
models which cannot be distinguished from the best model according to the F-test
(Buttkus, 2000, p. 231) with a 5% signicance level.
Most of the curves show the expected characteristics of a displacement spec-
trum. A at region can be observed towards the low frequency limit whereas the
1.5. RESULTS 19
10
1
10
0
10
1
10
5
10
4
10
3
10
2
f [Hz]
I
Q
1
individual events
mean with
95% confidence limits
Figure 1.5: Results for the attenuation parameter of intrinsic absorption. Gray dots:
attenuation parameter measured by inversion of individual events; Black dots with error
bars: logarithmic mean of individual measurements with 95% condence limits.
displacement decays above a corner frequency. A dierent behavior is found for the
October 2004 event that occurred in Northern Germany which is tectonically quiet.
Neither low frequency plateau nor corner frequency can be observed. I regard this
as an evidence for a complex source process.
The low frequency plateaus can be compared with seismic moments M
wm
0
in-
dependently obtained from waveform inversions. For most of the events the so-
lutions of the Swiss regional moment tensor catalog that is available on-line at
http://www.seismo.ethz.ch/mt/ are used. Refer to Braunmiller et al. (2002) for ex-
amples and a description of the method. M
wm
0
estimates for the three events before
2000 can be found in Braunmiller et al. (1994) and Braunmiller (2002). For the event
in 1995 there is no seismic moment available. The waveform inversion estimates of
M
0
are plotted as thick lines in gure 1.6. Obviously the low frequency values of the
displacement spectra obtained in this study match the moments from the waveform
inversion well. Assuming that the corner frequencies of the events used here exceed
1.5 Hz I take the logarithmic average of the displacement spectra in the four lowest
frequency bands as a coda based estimate of the seismic moment denoted M
coda
0
. In
gure 1.7 the values of seismic moment obtained in this study (M
coda
0
) are plotted
against the values from waveform modeling (M
wm
0
). On average there is a dierence
of 36% between the waveform and the coda results.
Another reference that is interesting to compare with the displacement spec-
tra is the local magnitude scale. Recently Reamer and Hinzen (2004) investigated
earthquakes in the Northern Rhine Area (around station BUG in gure 1.1) and
established the relation
log
10
M
Ml
0
= 1.083Ml + 10.215 (1.12)
between the local magnitude Ml and the seismic moment M
Ml
0
in Nm. The super-
20 CHAPTER 1. ACOUSTIC RADIATIVE TRANSFER THEORY
10
11
10
13
10
15
10
17
1992_04_13 1995_06_20 1997_10_21 1997_11_29
10
11
10
13
10
15
10
17
2001_06_23
s
o
u
r
c
e
d
i
s
p
l
a
c
e
m
e
n
t
s
p
e
c
t
r
u
m
|
M
(
)
|
[
N
m
]
2002_07_22 2003_02_22
10
1
10
0
10
1
2003_03_22
10
1
10
0
10
1
10
11
10
13
10
15
10
17
2004_05_22
10
1
10
0
10
1
2004_10_20
f [Hz]
10
1
10
0
10
1
2004_12_05
Figure 1.6: Double logarithmic plots of source displacement spectra of the 11 earth-
quakes used in this study. Numbers within the charts indicate the dates of the events
in yyyy mm dd format. Dots with continuous curve: source spectra estimated in this
study. Error bars correspond to 95% condence limits of the resolution of the inversion
(see text for explanation). The low frequency part shown in black is averaged to obtain
M
coda
0
. Thick horizontal line: seismic moment M
wm
0
estimated independently using wave-
form modeling techniques (Braunmiller et al., 2002). Thin dashed curve: omega - square
spectra independently calculated from local magnitudes using empirical relations. Low
frequency limit corresponds to M
Ml
0
.
script Ml indicates that the seismic moment M
Ml
0
is estimated from local magni-
tudes. Figure 1.7 also shows the comparison of M
Ml
0
with the results from waveform
modeling (M
wm
0
). On average the dierence between the magnitude-derived and the
waveform-derived moments is 57%. The scatter in the magnitude derived estimates
is thus larger than that of the coda derived estimates. However, according to the
F-test with 10 degrees of freedom the dierence between this 57% and the 36% mean
deviation of the coda-derived moment is not signicant at a 95% condence level.
This means the coda derived estimates of the seismic moment are at least as good
as the estimates that can be obtained directly from the local magnitudes.
Assuming an omega-square source model according to [M
Ml
()[ = M
Ml
0
/(1+
f/f
c
)
2
(Aki and Richards, 1980a, p. 823) one can use equation 1.12 to calculate
an approximate source displacement spectrum if the corner frequency f
c
is known.
1.5. RESULTS 21
10
14
10
16
10
18
10
14
10
15
10
16
10
17
10
18
M
0
f
r
o
m
c
o
d
a
/
m
a
g
n
i
t
u
d
e
[
N
m
]
M
0
wm
from waveform inversion [Nm]
M
0
coda
M
0
Ml
Figure 1.7: Scatter plot of the seismic moments obtained from seismic coda (M
coda
0
)
in this study and moments estimated from local magnitudes (M
Ml
0
) against estimates of
seismic moments from long period waveform modeling (M
wm
0
). Straight line indicates the
locations of perfect agreement of the dierent estimates.
Further assuming a stress drop of 10 MPa, it can be obtained from f
c
=
3
_
/M
Ml
0
0.49v
0
(Hough et al., 2000). These approximate source spectra are plotted as thin
dashed lines in gure 1.6. Note that these approximate source spectra are solely
based on the local magnitude estimated by the SZGRF.
Figure 1.6 shows that also the local magnitudes can give reasonable estimates
of the seismic moment. There are some dierences between the magnitude- and
coda-derived spectra in the corner frequency and in the high frequency asymptote.
[M
coda
()[ shows a steeper decline than the omega-square model.
1.5.3 Site response
Site amplication factors for velocity amplitudes range between 0.3 and 3. Mostly
they can easily be related to local geology (table 1.3). In the 0.375 Hz band for ex-
ample we nd amplications of crystalline sites only below 1 (logarithmic mean: 0.8)
whereas stations on sediments show amplications exclusively above 2 (logarithmic
mean: 3.0). Sites on sedimentary rocks occupy an intermediate range between 0.75
and 1.75 (logarithmic mean: 1.1). As examples the site factors for the stations BUG
and FUR in gure 1.8 are shown. BUG is situated on a clastic sedimentary rock.
The site amplication is not frequency dependent with a mean of 0.9. In contrast
FUR is placed on moraine over molasse (i.e. sediments). Here a clear amplication
of a factor greater than 2 is seen. Response factors for all stations and frequencies
are listed in table 1.3, together with information about geology.
22 CHAPTER 1. ACOUSTIC RADIATIVE TRANSFER THEORY
10
1
10
0
10
1
BUG
s
i
t
e
a
m
p
l
i
f
i
c
a
t
i
o
n
f
a
c
t
o
r
R
10
1
10
0
10
1
10
1
10
0
10
1
FUR
f [Hz]
s
i
t
e
a
m
p
l
i
f
i
c
a
t
i
o
n
f
a
c
t
o
r
R
individual events
mean with 95%
confidence limits
Figure 1.8: Site response spectra for the stations BUG and FUR. BUG is placed on
sedimentary rock whereas FUR is situated on sediments.
1.6 Discussion
The model used in the present study is a rst-order approximation of S-wave scatter-
ing in the earth. It neglects anisotropic scattering and mode conversions, does not
pay attention to source radiation patterns and assumes a statistically homogeneous
half-space. Some of these eects could be incorporated into the model by solving the
radiative transfer equation with Monte-Carlo techniques. This allows to take into
account anisotropic scattering or mode conversion of elastic waves (Przybilla et al.,
2006). Even anisotropic sources and deterministic earth structure, like a depth de-
pendent velocity and scattering coecient can be simulated (Margerin et al., 1998,
chapter 2). The improvement of the model that can be achieved by incorporating
these eects will be variable and the successful application of more complex models
requires detailed knowledge about the medium and the source (Hoshiba et al., 2001).
The source radiation patterns for example will probably have a minor eect on the
present results because in the late coda the eect vanishes due to the averaging by
multiple scattering and the direct S-wave is measured and averaged over a lot of
1.6. DISCUSSION 23
stations. Note that S-waves always have less pronounced radiation patterns than
P-waves irrespective of the station conguration (Sato and Fehler, 1998, p.150).
The acoustic approximation is justied here because the dierent magnitude of
the conversion scattering coecients for S to P and P to S conversion (Aki, 1992)
leads to a stable partitioning of energy at large lapse times which is dominated by
S-waves. In this regime P-waves carry only about 10% of total energy (Ryzhik et al.,
1996, eq. 5.40). In order to model the early S- and P-wave coda the treatment of the
elastic scattering problem seems to be the most promising modication to the present
approach (c.f. chapter 2). The computational eort, however, is substantially higher.
A still unsolved problem is the treatment of conversion and scattering of surface
modes.
One consequence of the half-space model is the inability to handle Lg-waves.
Lacombe et al. (2003) model Lg-waves using a Monte-Carlo technique to separate
intrinsic absorption and scattering attenuation. Interestingly Lacombe et al. (2003)
could not satisfactorily separate elastic and anelastic attenuation without the use
of information from direct waves incorporated as total attenuation. This directly
corresponds to my observation that led to the constraint of the inversion with the
ballistic energy.
Nevertheless, the present results are encouraging. Though a very simple model
is used I can directly relate the shape and amplitude of seismogram envelopes to
the moment of the source. These moment estimates are in good agreement with
independent results from moment tensor inversions. I nd reasonable values for the
total scattering coecient g
0
and the intrinsic attenuation parameter
I
Q
1
and ob-
tain site response factors that reect local geology. I think that the approximations
are justied in order to obtain fast estimates of relevant parameters without a priori
knowledge about medium or source.
There are some technical parameters in the algorithm that might inuence the
results. First, the choice of the mist function might have an eect. I veried that
log(E
obs
/E
mod
) is approximately Gaussian distributed in the coda time window if
the model ts well. For comparison I inverted the dataset using dierent mist
functions including the L1-norm and found no signicant dierences.
Another inuence might come from the choice of the inversion time windows.
The seismic coda time window starts at two times the S-wave travel time. In
this part of the seismogram the distribution of energy is traditionally believed to be
close to the diusion limit (Rautian and Khalturin, 1978; Wu and Aki, 1985). This
is not required in the present method but also the assumption of isotropic scattering
is a better approximation in the later coda. Additionally direct surface waves are
excluded by tting only the late coda. The direct S-wave time window starts at
the arrival of the direct S-wave (v = 3.5 km/s) and lasts until waves with apparent
velocities above 3.0 km/s have passed the station. This is somewhat arbitrary but
the objective was to collect energy that was radiated as S-wave from the source in
the direction of the receiver. Taking into account nite source duration and envelope
broadening due to forward scattering (Sato, 1989; Saito et al., 2002) this appeared
to be a reasonable choice. By comparing the results with inversions done with
dierent lengths of the direct S-wave time window I veried that the inuence is
24 CHAPTER 1. ACOUSTIC RADIATIVE TRANSFER THEORY
small compared to the uncertainties of the estimates.
A new aspect of this method is the frequency range used for inversion. Typically
only frequencies above 1 Hz have been used in studies of seismic scattering (Fehler
et al., 1992; Mayeda et al., 1992; Bianco et al., 2002; Lacombe et al., 2003). On one
hand this is due to the short period data used in these studies. On the other hand
low frequency coda is believed to be inuenced by surface waves (Wu and Aki, 1985;
Abubakirov and Gusev, 1990). The latter argument appears to be not well founded.
Mode conversions between body and surface waves as well as surface to surface
wave scattering are not bound to frequencies below 1 Hz. Since it is important
for the purpose of estimating the low frequency level of the source spectrum, i.e.
the seismic moment of the earthquakes, the broadband data is further exploited to
obtain information below the corner frequencies.
The major advantage of this method is that I can easily obtain source parame-
ters like the seismic moment without a priori knowledge about attenuation functions
that are needed for magnitude estimations and without waveform modeled reference
events that are needed if envelopes are approximated on the basis of empirical for-
mulae. The approach is thus optimally suited for source parameter estimation with
temporal seismic networks.
Another possible application is related to the stability of the coda compared to
direct waves. If the medium parameters g
0
and
I
Q
1
and the site response factors
R
i
have been estimated once in a study like the present, the seismic coda can be
used to obtain magnitude estimates that are more stable than estimates from direct
waves because the coda is less sensitive to source radiation patterns. In fact Mayeda
et al. (2003) showed that coda amplitudes have 35 times less interstation scatter as
compared to direct phases. Just like the method of Mayeda et al. (2003) my approach
can be used to signicantly improve the magnitude estimates in permanent seismic
networks.
Compared to waveform modeling techniques the present method has a funda-
mental advantage for the estimation of the seismic moment. Modeling waveforms
requires knowledge about the Greens function of the propagation medium. Due
to small scale heterogeneities in realistic earth models this is impossible for high
frequencies. Therefore waveform techniques can not be applied to small (Ml < 3.5)
earthquakes. As I model energy densities rather than waveforms the approach only
requires knowledge of the energy density Greens function. In this case the phase
relation between waves that interacted with these small scale structures can be ne-
glected and it is possible to obtain the Greens function in a statistical sense. It is
thus possible with this method to obtain estimates of the seismic moment for small
events with high frequency sources.
1.7 Conclusions
A new method is presented to extract information from the seismic coda using ra-
diative transfer theory in the acoustic approximation. It is based on the modeling of
coda envelopes as S-waves that were isotropically scattered at randomly distributed
1.7. CONCLUSIONS 25
heterogenieties in a statistically homogeneous half-space. I do not apply coda nor-
malization but instead incorporate the eects of dierent sources and site responses
into the inversion scheme which allows to explicitly resolve these parameters together
with medium parameters.
Source spectra and seismic moments can thus directly be obtained from seis-
mogram envelopes and it is shown that there is excellent agreement between my
estimates and values independently obtained in moment tensor inversions. Addi-
tionally, site response factors are obtained and intrinsic and scattering attenuation
are separated.
2
6
C
H
A
P
T
E
R
1
.
A
C
O
U
S
T
I
C
R
A
D
I
A
T
I
V
E
T
R
A
N
S
F
E
R
T
H
E
O
R
Y
Table 1.1: List of earthquakes used in this study. Source time, location, depth, and local magnitudes are provided by the German
Central Seismological Observatory. Seismic moments estimated by modeling waveforms are listed in column 6 with respective references
in column 7. The last column lists seismic moments estimated in the present study.
Date Time Location Ml Depth M
wm
0
in Nm Reference M
coda
0
in Nm
(waveform modeling) (this study)
1992/04/13 01:20:03.1 51.14
o
N 6.05
o
E 6.0 15 9.210
16
Braunmiller et al. (1994) 9.310
16
1995/06/20 01:54:57.5 50.57
o
N 4.82
o
E 4.6 10 9.010
14
1997/10/21 16:44:39.4 48.78
o
N 9.72
o
E 4.0 10 2.310
14
Braunmiller (2002) 1.710
14
1997/11/29 20:06:09.3 50.31
o
N 8.35
o
E 4.0 7 3.010
14
Braunmiller (2002) 4.510
14
2001/06/23 01:40:05.3 50.87
o
N 6.17
o
E 4.3 10 9.2310
14
www.seismo.ethz.ch/mt 5.410
14
2002/07/22 05:45:03.8 50.91
o
N 6.17
o
E 5.2 10 8.6610
15
www.seismo.ethz.ch/mt 4.110
15
2003/02/22 20:41:05.5 48.35
o
N 6.68
o
E 5.7 10 1.6410
16
www.seismo.ethz.ch/mt 1.510
16
2003/03/22 13:36:16.3 48.25
o
N 8.98
o
E 4.7 10 8.5210
14
www.seismo.ethz.ch/mt 8.810
14
2004/05/22 05:19:03.3 50.38
o
N 7.44
o
E 4.0 10 2.7810
14
www.seismo.ethz.ch/mt 1.810
14
2004/10/20 06:59:16.0 53.04
o
N 9.54
o
E 4.5 5 3.4410
15
www.seismo.ethz.ch/mt 2.410
15
2004/12/05 01:52:38.8 48.12
o
N 8.04
o
E 5.1 10 8.0410
15
www.seismo.ethz.ch/mt 6.810
15
1
.
7
.
C
O
N
C
L
U
S
I
O
N
S
2
7
Table 1.2: Comparison of structural parameters g
0
and
I
Q
1
for dierent regions in frequency bands centered at 3 Hz.
Reference g
0
in m
1 I
Q
1
Region Distance range Model
Abubakirov and Gusev (1990) 6.710
6
0.0032 Kamchatka local half space
Fehler et al. (1992) 6.510
6
0.0026 Kanto-Tokai (Japan) local half space
Mayeda et al. (1992) 4.2610
5
0.00317 Long Valley local half space
2.8210
5
0.00336 Central California local half space
3.1910
5
0.00055 Hawaii local half space
Bianco et al. (2002) 2.7 10
6
0.0034 Italy local-regional half space
Lacombe et al. (2003) 410
6
0.0012 France regional layer over half-space
Abubakirov (2005) 5.310
6
0.0031 Kamchatka regional half space
this study 1.4510
6
0.0013 Germany regional half space
2
8
C
H
A
P
T
E
R
1
.
A
C
O
U
S
T
I
C
R
A
D
I
A
T
I
V
E
T
R
A
N
S
F
E
R
T
H
E
O
R
Y
Table 1.3: Site response factors of the 25 seismic stations that were used in this study. The sign denotes the logarithmic uncertainty.
The site factor is to be multiplied or divided by the uncertainty to obtain the limits of the 95% condence interval. If no uncertainty is
given only a single event could be inverted. The two stations marked with * are placed several 100 m below surface.
Station Geology 0.1875 Hz 0.375 Hz 0.75 Hz 1.5 Hz 3.0 Hz 6.0 Hz 12.0 Hz 24.0 Hz
BFO granite 0.71 1.61 0.76 1.26 0.72 1.16 0.73 1.20 0.64 1.28 0.57 1.18 0.94 1.54 1.06 2.85
BRNL cenozoic sediments 3.78 4.61 3.72 7.52 2.03
BRG slate 1.10 2.09 0.81 1.46 0.56 1.15 0.60 1.16 0.90 1.23 1.21 8.32
BSEG anhydrit 2.11 1.19 1.23 2.04 0.58 1.52 0.62 2.05 1.13 0.53
BUG sedimentary rocks 0.87 1.37 0.96 1.41 0.88 1.25 0.86 1.21 1.08 1.39 1.11 1.42 0.72 1.39 0.85 1.31
CLL Greywacke rocks 1.12 2.18 0.91 1.27 0.74 1.17 1.10 1.21 1.20 1.23 1.85 1.40
CLZ sedimentary rocks 0.87 1.20 0.86 1.22 0.91 1.13 0.84 1.09 0.91 1.12 1.37 1.12 2.13 1.37
FUR moraine over Molasse 2.80 1.37 3.45 1.14 2.49 1.11 1.89 1.14 1.89 1.34 1.06 2.99
GEC2 granite 1.11 1.27 0.77 1.17 0.65 1.20 0.68 1.13 0.75 1.45 1.30 1.57
GRA1 limestone 1.28 1.18 1.58 1.24 1.69 1.08 1.26 1.09 0.98 1.11 0.63 1.24
GRB1 limestone 1.04 1.19 1.32 1.17 1.61 1.09 1.73 1.11 2.97 1.18 2.51 1.23
GRC1 limestone 0.97 1.33 0.86 1.22 0.97 1.10 1.18 1.16 0.79 1.14 0.79 1.50
GRFO limestone 1.31 1.17 1.57 1.24 1.55 1.08 1.00 1.07 0.51 1.06 0.63 1.20
GSH slate 0.28 0.97 0.71 0.97 1.43 1.77 1.00
HAM cenozoic sediments 7.26 5.03 43.31 2.52 8.11 1.40 1.46
HLG sandstone 1.87 3.51 1.37 92.54 1.54 2.35 1.68 1.81 1.15 10.82 1.44
IBBN 1.21 1.32 1.23 1.60 1.17 1.37 1.02 1.26 1.22 1.42 1.60 1.23 2.40
MOX slate 0.98 1.35 0.87 1.19 0.65 1.09 0.69 1.07 0.78 1.08 0.94 1.19 1.39 4.65
NRDL Zechstein sediments* 1.16 1.74 0.61 0.36 0.29 0.29
RGN soft sediments 2.52 5.06 2.60 1.94 2.23 2.61 5.14 5.83
RUE limestone 1.83 1.53 1.16 1.52 0.93 1.31 0.81 1.61
STU hard marls 0.92 1.38 0.86 1.27 1.05 1.20 1.82 1.14 1.47 1.24 0.56 2.70
TNS quartzite 0.88 2.00 0.70 1.42 0.72 1.28 1.09 1.41 0.79 1.36 0.74 1.65 0.79 1.77 1.11 1.48
UBBA Zechstein sediments* 0.90 17.77 0.76 1.00 0.40 1.76 0.40 2.24 0.40 1.33 0.42
WET gneiss 0.89 1.68 0.69 1.17 0.54 1.10 0.63 1.17 0.78 1.17 1.54 1.18 2.16 2.09
Chapter 2
Monte-Carlo simulations of elastic
wave energy in the presence of
deterministic structure
29
30 CHAPTER 2. MONTE-CARLO SIMULATIONS
Abstract
In this chapter a short introduction into the transfer problem of elastic energy in
heterogeneous media is given. A comparison of dierent strategies for Monte-Carlo
simulations leads to the description of an algorithm that was developed to solve
the elastic radiative transfer equation in a model that is close to the situation of
the earth crust. The algorithm treats anisotropic scattering including conversions
between compressional and shearing modes. Deterministic large scale structure is
incorporated in terms of a free surface and an interface at depth that reect and
convert and in the case of the interface also transmit seismic energy. The horizontal
interface separates an innite half space which may have a vertical velocity gradient
from the overlaying layer of constant velocity.
INTRODUCTION 31
2.1 Introduction
In chapter 1 radiative transfer theory is applied in the approximation of multiple
acoustic and isotropic scattering. For the analysis of the S-wave coda and the es-
timation of basic medium and source/site parameters this approximation is useful.
The main shortcoming for practical applications is the inability of this approxima-
tion to model the P-wave and its coda. From the theoretical point of view, the
limitation to acoustic and isotropic scattering means a severe simplication of the
physics of wave propagation through a heterogeneous elastic medium. In this chap-
ter an algorithm is introduced that solves the radiative transfer equation with less
restrictive approximations. The fundamental improvement is the simulation of elas-
tic energy with conversion between P- and S-modes. This can be done in a model
that contains velocity gradients and discontinuities.
The radiative transfer of elastic energy in a model with deterministic velocity
structure is perhaps the most accurate approach to the description of the scat-
tering of seismic waves in the earths crust, that is available. Numerous approx-
imation to this situation can be found in the literature. Most simple approaches
assume isotropic single scattering of acoustic energy an innite medium (e.g. Aki
and Chouet, 1975). Isotropic means that the there is no preferred direction of scat-
tering. The acoustic approximation limits the scattering to S-waves. P-waves and
conversions are neglected. The single scattering approach assumes that waves are
scattered only once on their way from the source to the receiver. Since there is
strong evidence for multiple scattering of seismic waves in the crust (Hennino et al.,
2001; Larose et al., 2004), this approximation strongly limits the applicability of this
approach. The diusion approximation (Dainty and Toks oz, 1981; Wegler and L uhr,
2001) that assumes strong scattering is the other extreme, that is only applicable
for late lapse times of crustal seismic records. Multiple scattering is suitable for the
intermediate range of moderate scattering (e.g. Hoshiba, 1991).
The approximation of isotropic scattering was replaced by non-isotropic scat-
tering by Hoshiba (1995). The modeling of seismogram envelopes using elastic scat-
tering at distributed spheres is due to Margerin et al. (2000). This was extended
to random heterogeneity by Shearer and Earle (2004) and Przybilla et al. (2006).
Deterministic structure in terms of an interface was rst modeled by Margerin et al.
(1998) for acoustic waves.
Here, everything is brought together and an algorithm is presented that is able
to synthesize seismogram envelopes for multiple non-isotropic scattering of elastic
waves at randomly distributed heterogeneities in deterministic structure. Before the
details of the algorithms components are discussed, I present the radiative transfer
equation in the from that is used by the algorithm and briey compare dierent
Monte-Carlo techniques for their solution.
32 CHAPTER 2. MONTE-CARLO SIMULATIONS
2.2 Elastic radiative transfer theory
The central quantity of transfer theory is the radiance or specic intensity I(n, r).
It measures the energy ux from a radiating surface into a unit solid angle around
direction n. This angular spectrum of the energy ux may be a local quantity and
depend on position r. Alternatively to the energy ux described by the radiance
one may work with the local and directionally resolved energy density a(n, x) that
is related to the radiance by
a(n, x) =
1
v
I(n, x) (2.1)
with v being the velocity of the ray. Here, and in all what follows I implicitly
assume dependence of the radiance or related quantities on time and frequency.
This is justied because on one hand the energy transport in seismology is always
non-stationary which implies an intrinsic time dependence. On the other hand the
propagation medium is stationary which ensures that frequencies do not change with
time and remain uncoupled.
In the classical book of Chandrasekhar (1960), the radiative transfer equation
that describes the radiance is introduced phenomenologically to describe light prop-
agation through atmospheres. The phenomenological view of the transport problem
is based on considerations of energy conservation. Regarding the energy ux in
direction n at a xed point in space, the energy balance can be expressed as:
variation of energy
ux in direction n
=
decrement of energy ux due to ab-
sorption and scattering from direction
n into n
+
increment of energy ux in direction n
due to scattering from directions n
into
n
+
increment of energy ux due to radia-
tion of sources in direction n
These heuristic considerations are physically transparent and seem intuitively
convincing without further explanations. For several decades this eld developed
without complete physical foundations and independently of wave physics. A rigor-
ous statistical substantiation of this concept was achieved when the direct connec-
tion between the angular Fourier transform of the radiance and the spatial coherence
function of the wave eld was discovered. The coherence function dened as
(x, , t) = u(x + , t)u
I =
_
I
_
= c
_
A
) A
)
A
) A
)
_
(2.4)
where I dene the directions of the components indicated by the subscripts and
as the directions of increasing angles and of the polar coordinate system shown
in gure 2.2. Here A
and A
I
S
U
V
_
_
_
_
_
_
=
_
_
_
_
_
_
I
P
I
+ I
i(I
)
_
_
_
_
_
_
= c
_
_
_
_
_
_
[A
P
[
2
)
[A
[
2
)
[A
[
2
)
2ReA
)
2ImA
)
_
_
_
_
_
_
. (2.5)
with c being a vector of proportionality constants with components c
1
=
2
/2
and c
2...5
=
2
/2. I
S
and I
S
cos()) (2.6)
V = 2c
5
A
sin()) (2.7)
The Stokes vector provides a full description of the state of polarization in-
cluding linear and elliptic polarizations. The character of a seismic shear source as
detailed in section 2.4.3 and the structure of the scattering coecients (sec. 2.4.4)
allows to simplify the treatment of polarizations. A seismic shear dislocation point
source emits completely linearly polarized waves. For these waves the phase shift
is zero. Also during scattering events the two polarizations remain in phase. This
leads me to an approximation of the Stokes vector. In the following I assume that
S-waves are completely linearly polarized and no elliptic polarization occurs. This
allows to omit the V -component of the Stokes vector which is zero for linear polarized
waves (cf. equation 2.6 for = 0). The U-component is determined by the following
relation that holds for completely polarized waves (Apresyan and Kravtsov, 1996,
p. 74)
(I
S
+ I
S
)
2
= (I
S
I
S
)
2
+ U
2
+ V
2
. (2.8)
For V = 0, U is dened by I
S
and I
S
and I
S
. They fol-
low directly from the energy balance in a multi-mode system (c.f. Apresyan and
34 CHAPTER 2. MONTE-CARLO SIMULATIONS
Kravtsov, 1996, p. 84) and describe three propagating modes. The transport equa-
tion for I
P
reads:
_
0
t
+n
_
I
P
(n, r) =
_
g
PP
0
+ g
PS
0
+
0
I
Q
P
_
I
P
(n, r) (2.9)
+
_
4
g
PP
()I
P
(n
, r)d
n
+
_
4
g
SP
(,
)I
S
(n
, r)d
n
+
_
4
g
SP
(,
)I
S
(n
, r)d
n
.
Here I used the following notation:
0
is the mean P-wave velocity, g
AB
0
are the
total scattering coecients given by equations 2.47, and
I
Q
P
is the quality factor of
intrinsic P-wave attenuation (section 2.4.5). The scattering coecients g
AB
(, )
are discussed in section 2.4.4. These coecients are assumed to be given in the local
scattering coordinate system, i.e. is the scattering angle between n and n
and is
measured between the polarization direction of the incident wave and the scattering
plane spanned by n and n
. is dierent for I
S
and I
S
. Here d
n
= sin()dd.
In a similar notation the transport equations for I
S
and I
S
read
_
0
t
+n
_
I
S
(n, r) =
_
g
SS
0
+ g
SP
0
+
0
I
Q
S
_
I
S
(n, r) (2.10)
+
_
4
g
SS
l
(,
)I
S
(n
, r) cos
2
()d
n
+
_
4
g
SS
r
(,
)I
S
(n
, r) sin
2
()d
n
+
_
4
g
SS
l
(,
)I
S
(n
, r) cos
2
()d
n
+
_
4
g
SS
r
(,
)I
S
(n
, r) sin
2
()d
n
+
_
4
g
PS
()I
P
(n
, r) cos
2
()d
n
.
ELASTIC RADIATIVE TRANSFER THEORY 35
X Z
Y
I
S
l
I
S
r
Figure 2.1: Illustration of the S-polarization directions in the radiative transfer equations.
Angles are identied by color of the arcs. The red and blue arrows indicate directions n
0
t
+n
_
I
S
(n, r) =
_
g
SS
0
+ g
SP
0
+
0
I
Q
S
_
I
S
(n, r) (2.11)
+
_
4
g
SS
l
(,
)I
S
(n
, r) sin
2
()d
n
+
_
4
g
SS
r
(,
)I
S
(n
, r) cos
2
()d
n
+
_
4
g
SS
l
(,
)I
S
(n
, r) sin
2
()d
n
+
_
4
g
SS
r
(,
)I
S
(n
, r) cos
2
()d
n
+
_
4
g
PS
()I
P
(n
, r) sin
2
()d
n
.
The angle is measured between the scattering plane and the vertical plane that
contains the vector n. This form of the transfer equations is unusual but it best
refers to the way I solve the equations with the Monte-Carlo technique. There are
two rotations hidden in the integrands of these equations. The rst corresponds to
the notation of the scattering coecients in the local frame with orientation of the
polar axis along n
and
= /2 +
0
I
Q
P
(2.13)
+
l
N(n, r)
__
4
g
PP
()N
P
(n
, r)d
n
+
_
4
g
SP
(, )N
S
(n
, r)d
n
+
_
4
g
SP
(, )N
S
(n
, r)d
n
_
.
This is a probabilistic interpretation of equation 2.9. The left hand side of equa-
tion 2.13 is the fractional change of the number of particles (that move in direction
n) on a path element l along a ray with direction n. The right hand side has two
terms representing loss of particles. The rst describes the fraction of particles that
change the direction and leave the ray. It depends on the total scattering coe-
cients g
AB
0
. The probability of the new direction in which these particles continue
to propagate is determined by the scattering coecient of the accordant scattering
type:
AB
(, ) =
g
AB
(, )
4g
AB
0
. (2.14)
38 CHAPTER 2. MONTE-CARLO SIMULATIONS
The fraction of particles that die o because of intrinsic attenuation is described by
the second term containing /(
0
I
Q
P
).
The last term of equation 2.13 represents the gain of particles. It depends on
all particles at position r independent of their directions n
Figure 2.2: Illustration of the relations between the geographical, the Cartesian and the
polar coordinate systems as they are used in the algorithm. Angles are identied by color.
2.4.2 Description of a particle
For a complete description of a particle the following parameters are needed:
Coordinates in space and time (x, y, z, t)
Velocity vector (v
x
, v
y
, v
z
)
Stokes vector (I
P
, I
S
, I
S
, v
).
Polarization angle
Knowledge of the propagation direction in polar coordinates and of the po-
larization angle, dened as the angle between the polarization axis of a linearly
polarized S-wave and the vertical plane that contains the ray path, is useful for the
treatment of interface reections/conversions/transmissions and the transformation
of scattering angles from the local into the global coordinate system.
2.4.3 Particle emission from a point shear dislocation
For the emission of particles, a description of the source in terms of probability
densities for radiation directions and particle type with polarization is needed.
40 CHAPTER 2. MONTE-CARLO SIMULATIONS
A seismic source can be represented as a point shear dislocation if the source
dimensions are small compared to wavelength and distance. Mathematically this
can be described by a double couple of forces (Aki and Richards, 1980b, p. 113).
The far eld displacement vector of such a source is
u(x, t) = B
P
r
e
r
2
15
M(t r/)
4
3
r
. .
Pradiation
+
_
B
S
+ B
S
_
2
5
M(t r/)
4
3
r
. .
Sradiation
(2.15)
where x = (r, , ) is the position vector in spherical coordinates,
M is the time
derivative of the moment time function and e
r
, e
and e
15
2
sin
2
() sin(2)
B
S
=
1
2
_
5
2
sin(2) sin(2) (2.16)
B
S
=
_
5
2
sin() cos(2).
Here subscripts denote polarization direction and superscripts indicate wave type.
The probability distributions of the radiation directions are given by the squares of
the radiation pattern:
P
r
= sin
4
() sin
2
(2)
= sin
2
(2) sin
2
(2) (2.17)
= sin
2
() cos
2
(2)
The radiation pattern are normalized such that
_
B
P
r
2
d =
_
_
B
S
2
+
B
S
2
_
d = 4. (2.18)
By integrating the energy ux density v[ u[
2
with v being either P-wave velocity
or S-wave velocity over a spherical shell and time one can obtain the radiated P-
and S-energy respectively (c.f. Sato and Fehler, 1998, p. 149). The ratio of radiated
P- and S-energy is given by
W
S
W
P
=
3
2
5
. (2.19)
The ratio between the energies radiated in the two dierent S polarizations can
OUTLINE OF THE RadTrans ALGORITHM 41
be obtained by integrating e.g.
B
S
2
over the whole solid angle:
_
2
0
_
0
B
S
2
sin()dd =
5
2
_
2
0
_
0
sin
3
() cos
2
(2)dd
=
5
2
__
1
3
sin
2
()
3
2
_
cos()
_
0
_
2
+
1
8
sin(4)
_
2
0
=
10
3
(2.20)
Using this together with the normalization condition eq. 2.18 I nd the desired ratio:
W
S
W
S
=
1
5
(2.21)
Equations 2.17, 2.19 and 2.21 allow to simulate the energy radiation of a shear
source with fault normal in (1,0,0) direction and slip direction (0,1,0). For an ar-
bitrarily oriented source a rotation matrix is calculated from the given strike, dip
and rake angles. The rotation is rst performed for the rake around the axis along
(1,0,0) then for dip around (0,1,0) and nally for strike around (0,0,1). Both the
velocity and polarization vectors are multiplied with this rotation matrix to account
for the specic source orientation.
2.4.4 Scattering
The most interesting and complicated part in this algorithm is the simulation of scat-
tering of elastic vector waves by distributed heterogeneities. I follow the derivation
by Sato and Fehler (1998).
Random medium
Lets assume a certain property v of the medium is not constant in space but depends
on position x. This property may be the wave velocity, the density or a Lame
coecient. The eld v(x) can be written as the sum of a constant mean eld v
0
and
a perturbation v:
v(x) = v
0
+ v(x) = v
0
[1 + (x)]. (2.22)
(x) is the fractional uctuation. v
0
is chosen such that the average, denoted by
. . .), of an ensemble of heterogeneous media is:
v
0
= v(x)) and (x)) = 0 (2.23)
The random function (x) is best described by its spectral properties. For this
I introduce the auto-correlation function (ACF) as
R(x) (y)(y +x)). (2.24)
In the following it is assumed that the random function is isotropic and homo-
geneous which means that R does only depend on the distance r between y and x
(r [y x[).
42 CHAPTER 2. MONTE-CARLO SIMULATIONS
The strength of the uctuation is given by the mean square fractional uctua-
tion
2
R(0) = (x)
2
) (2.25)
and the spatial variation of the uctuations is characterized by the correlation length
a. The Fourier transform of the ACF R gives the power spectral density function
(PSD). For isotropic and homogeneous random media the PSD is a function of the
magnitude of the wavenumber vector m [m[. The most commonly used ACFs
and their corresponding PSDs are given below.
Gaussian ACF
R(r) =
2
e
r
2
/a
2
and PSD(m) =
2
3
a
3
e
m
2
a
2
/4
(2.26)
This ACF can be used for media that are poor in short wavelength components.
Exponential ACF
R(r) =
2
e
r/a
and PSD(m) =
8
2
a
3
(1 + a
2
m
2
)
2
(2.27)
The exponential ACF is much richer in short wavelength heterogeneities as it
decays for large wavenumbers as a power low with exponent 4.
von K arm an ACF
R(r) =
2
2
1
()
_
r
a
_
_
r
a
_
for = 0 0.5 (2.28)
PSD(m) =
8
3/2
2
a
3
( + 3/2)
()(1 + a
2
m
2
)
+
3
2
(2.29)
Here is the gamma function and K
F
AB
2
(2.31)
Here A and B indicate the wave mode which may be either P or S. v
A
and v
B
are
the mean velocities of the respective modes. If the inhomogeneities are not localized
but distributed over a larger volume the scattering coecient may be dened as 4
times the ensemble average of the scattering cross section of a region of size L.
g
AB
4
1
L
3
_
d
AB
d
_
= 4
1
L
3
_
F
AB
2
_
(2.32)
Thus the scattering coecients relate the incident energy ux to the angular de-
pendent energy ux that is scattered from distributed heterogeneities in a region of
unit size around the origin.
The main assumptions that are used for the derivation of the scattering coe-
cients are the following:
The heterogeneities can be described by eq. 2.22 with [[ 1 indicating that
the uctuation must be small compared to the average value of the quantity.
In order to be able to use a rst order perturbation method to decompose
the wave eld into a scattered and a homogeneous unscattered component
it is assumed that the scattered eld is much smaller in amplitude than the
incident eld. This approximation means that the scattering must be weak
and is known as the Born approximation.
To solve the wave equation for the scattered component of the wave eld the
Green function in homogeneous space is used.
It is further assumed that observations are made in the far eld (r a)
and the Fraunhofer zone (r a
2
k/) of the scatterer. This means that the
distance between scatterer and receiver on one hand and the wavelength on
the other hand have to be large compared to the spatial length scale of the
uctuations.
The uctuations of the wave velocities are equal. They are correlated with the
mass density of rock as stated by Birchs law (Birch, 1961). These correlations
reduce the uctuations to one single fractional uctuation that is common to
the wave velocities and the mass density.
(x)
(x)
0
=
(x)
0
=
1
(x)
0
(2.33)
and are the velocities of P- and S-waves respectively and is the mass
density. The parameter ranges between 0.78 < < 0.8 in typical lithosphere
(Sato and Fehler, 1998, p. 101).
44 CHAPTER 2. MONTE-CARLO SIMULATIONS
These are the formal assumptions and approximations used to derive the spe-
cic form of the scattering coecients that is used here. The range of applicability
of the radiometric concept is much wider. Przybilla et al. (2006) showed that even
with scattering coecients derived under these assumptions the correspondence be-
tween the radiative transfer equation and the full wave equation goes far beyond
what would be expected from the above conditions.
In spherical coordinates (, ) the scattering coecients are given by Sato and
Fehler (1998, p. 104):
g
PP
() =
l
4
4
X
PP
()
2
PSD
_
2l
0
sin
_
2
__
(2.34)
g
PS
() =
1
0
l
4
4
X
PS
()
2
PSD
_
1
0
_
1 +
2
0
2
0
cos()
_
(2.35)
g
SP
(, ) =
0
l
4
4
X
SP
(, )
2
PSD
_
1
0
_
1 +
2
0
2
0
cos()
_
(2.36)
g
SS
l
(, ) =
l
4
4
X
SS
l
(, )
2
PSD
_
2l sin
_
2
__
(2.37)
g
SS
r
(, ) =
l
4
4
X
SS
r
(, )
2
PSD
_
2l sin
_
2
__
(2.38)
measures the angle between the initial direction and the direction of the scat-
tered particle. This angle is usually called scattering angle. measures the angle
between the polarization direction of the incident particle and the scattering plane
spanned by the incident and scattered directions. Superscripts indicate the type
of scattering. SP for example stands for the scattering of a S-wave into a P-wave.
Subscripts r and l indicate the polarization of the scattered particle with respect to
the scattering plane. r and l stand for perpendicular (normal to scattering plane)
and parallel (in the scattering plane) respectively. This distinction is only neces-
sary for SS-scattering since the polarization in the other cases is non-ambiguous.
The polarization of P-waves is radial (in the direction of propagation) and S-waves
that were generated by scattering from P-waves have parallel polarization in the
scattering plane. The argument of the power spectral density function is the so
called exchange wavenumber expressed in terms of the scattering angle and the S-
wavenumber l. The terms denoted by X are the basic scattering patterns. They are
the same for all types of media. Modications of the angular patterns of the scatter-
ing coecients are solely introduced by the power spectrum of the heterogeneities.
OUTLINE OF THE RadTrans ALGORITHM 45
The scattering patterns are given by Sato and Fehler (1998, p. 102):
X
PP
() =
1
0
_
_
1 + cos() +
2
2
0
sin
2
()
_
2 +
4
2
0
sin
2
()
_
(2.39)
X
PS
() = sin()
_
_
1
2
0
cos()
_
0
cos()
_
(2.40)
X
SP
(, ) =
1
2
0
sin() cos()
_
_
1
2
0
cos()
_
0
cos()
_
(2.41)
X
SS
l
(, ) = cos() [ (cos() cos(2)) 2 cos(2)] (2.42)
X
SS
r
(, ) = sin() [(cos() 1) + 2 cos()] (2.43)
The scattering coecients 2.34 to 2.38 dene the probability distributions for
the directions of the scattered particles in a local frame with respect to the initial
direction and polarization.
Transformation of angles
The new propagation and polarization directions of a scattered particle are selected
in a local coordinate system dened by the incident direction and the direction of
polarization. For further computations the new directions have to be transformed
into the global coordinate system described in sec. 2.4.1. The geometry of this
transformation is depicted in gure 2.3. Let the direction of the incident particle be
given by the polar coordinates (
i
,
i
). The polarization of the incident particle is
given by
i
which is the angle between the polarization axis and the vertical plane
that contains the incident ray. The two angles that describe the scattering process
are the scattering angle and , the angle between the scattering plane and the
initial polarization direction. To express the polar angle
s
and azimuth
s
of the
scattered particle in terms of the direction of the incoming particle (
i
,
i
) and the
scattering and polarization angles (, ,
i
) one can use spherical trigonometry. The
polar angle of the scattered particle
s
can be expressed as
cos(
s
) = cos(
i
) cos() + sin(
i
) sin() cos(
i
+ ). (2.44)
The azimuth
s
of the scattered particle is given by
cos(
i
s
) =
cos() cos(
i
) cos(
s
)
sin(
i
) sin(
s
)
. (2.45)
To nd the polarization angle of the scattered particle the angle
s
between the
scattering plane and the vertical plane that contains the scattered ray is calculated
using
sin(
s
) =
sin(
i
s
)
sin()
sin(
i
). (2.46)
The polarization direction of the scattered particle is either
s
+ as given by eq. 2.46
if the polarization is parallel to the scattering plane or
s
+ /2 if the polarization
is perpendicular.
46 CHAPTER 2. MONTE-CARLO SIMULATIONS
X Z
Y
s
pol
Figure 2.3: Illustration of the scattering and the polarization angle and their relation to
the global polar coordinate system. Red and blue arrows indicates the directions of the
incoming and scattered particle respectively. The polarization direction of the incoming
S-particle is indicated by the bold black line.
Total scattering coecients
If a particle undergoes scattering, the probability distribution of the direction in
which the particle continues to propagate is given by the scattering coecients 2.34
to 2.38. The integral of the scattering coecients over the solid angle gives the total
scattering coecient g
AB
0
, which denes the probability that a certain scattering
event from mode A to B occurs while the particle moves one unit length (Sato and
Fehler, 1998, p. 43):
g
AB
0
1
4
_
g
AB
d =
1
l
AB
(2.47)
g
SS
0
is the sum of the total scattering coecients for the two dierent polarizations.
The inverse of g
AB
0
is the path l
AB
, i.e, the average length that a particle of type A
propagates until it is scattered into mode B. The length l
A
dened as
l
A
=
_
B
g
AB
0
_
1
(2.48)
is the mean free path of wave mode A. The length L
A
that a particle of mode A
propagates between two scattering events is determined by the exponential proba-
bility distribution
(L
A
) = (1/l
A
) exp(L
A
/l
A
) (2.49)
.
Random walk
In the previous sections I have outlined how the probability distributions for the
scattering events are related to the properties of the medium. Here I combine these
OUTLINE OF THE RadTrans ALGORITHM 47
information and explain how it is used to simulate the random walk of the particles.
This is largely analogous to Margerin et al. (2000).
The random walk is performed in incremental time steps t. Depending of
their velocity the particles move paths of length l in this time. The probability
of an incident particle of mode M
i
to be scattered while moving l is according to
equation 2.49
(M
i
) = 1 exp
_
l
l
M
i
_
. (2.50)
A rst uniform random number
1
(0, 1) is used to decide whether a particle is
scattered by comparing it to (M
i
). If
1
< (M
i
) scattering occurs.
In this case the mode of the outgoing particle M
o
is selected according to the
conditional probability distribution
(M
i
[M
o
) =
g
M
i
Mo
0
Mo
g
M
i
Mo
0
. (2.51)
It is the probability that a particle is scattered into mode M
o
given that the incident
mode is M
i
. For the selection of the outgoing mode a second uniform random number
2
(0, 1) is used. Here the two S-polarizations are considered dierent modes,
which means that in the case of an incoming S-particle it is decided at this point
whether the outgoing particle is polarized parallel or perpendicular to the scattering
plane, i.e. whether the new direction is selected using g
SS
l
or g
SS
r
. This choice also
determines the polarization of the outgoing particle.
Due to the special form of the scattering coecients (equations 2.34 to 2.38)
that are used to select the new directions of the particle, and the assumption of
linear polarization, the probability distributions of the scattering angle and angle
are independent. The probability of the angle under the condition that the
incoming particle of mode M
i
is scattered into mode M
o
is
(M
i
, M
o
[) =
_
g
M
i
Mo
(, )d
4g
M
i
Mo
0
. (2.52)
The probability of the angle is
(M
i
, M
o
[) =
_
g
M
i
Mo
(, ) sin()d
4g
M
i
Mo
0
. (2.53)
For fast evaluation of the probability distributions (M
i
, M
o
[) and (M
i
, M
o
[)
during the simulation I generate tables with random numbers drawn from these
distributions prior to the simulation (section 2.4.8). This means that tables with
values of are generated for the ve types of scattering that can occur, namely
P P, P S, S P S S
l
, and S S
r
. For the angle only a single
table has to be generated with numbers distributed according to sin
2
(). Numbers
with a cos
2
() distribution are obtained by adding /2 to the sin
2
() distributed
numbers. In the cases of P P and P S scattering the values for the angle
48 CHAPTER 2. MONTE-CARLO SIMULATIONS
are uniformly distributed in [0, 2). Two further uniformly distributed random
numbers
3
,
4
(0, 1) are used to select values for and from the appropriate
probability tables.
Since the radiative transfer equations in section 2.2 are written as a 3-mode
system, the I
S
and I
S
and I
S
0
I
Q
P
and E
S
(r, f) e
2rf
0
I
Q
S
. (2.54)
This decay is already contained in the transfer equations 2.9 through 2.11. In
equation 2.13 this decay is only described correctly for innitesimal l. For nite
size path length l the probability that a particle dies o can be calculated from 2.54
to be
(l) = 1 e
l
0
I
Q
P
and (l) = 1 e
l
0
I
Q
S
(2.55)
OUTLINE OF THE RadTrans ALGORITHM 49
for P- and S-waves respectively. These are also the factors that the Stokes vector has
to be multiplied with if the particles keep propagating. This is the strategy chosen
in the present algorithm because it accelerates the convergence of the energy density
estimates in the later part of the envelopes. However, it requires some considerations
of the numerical accuracy.
Total attenuation parameters Q
1
=
I
Q
1
+
Sc
Q
1
, with
Sc
Q
1
being the scat-
tering attenuation parameter, of the earths crust were estimated between 0.0005
and 0.05 (Sato and Fehler, 1998, g. 5.1 and 5.2). Assuming that
I
Q
1
and
Sc
Q
1
are of equal size, one can roughly approximate the maximum decay that can occur
during a simulation. For a large simulation length of 1000 s and an intrinsic quality
factor of 100 the energy can decrease by factor of 10
41
at 3 Hz whereas it decreases
only by a factor of 10
4
in a medium with
I
Q = 1000. Obviously there is a large
dynamic range required for the simulation of intrinsic attenuation. In a simulation
that consists of a low
I
Q crustal layer and an underlying mantle with high
I
Q the
energy of two particles of which one propagated in the mantle and the other in
the crust can dier by a factor of 10
37
after 1000 s. When these two particles are
recorded in the same cell the numerical accuracy of the oating point representation
can lead to a phenomenon known as absorption in computer science. The value
of the larger number does not change if the smaller is added. Because of this the
algorithm works with 64-bit integer numbers that have the highest dynamical range
with constant precision.
2.4.6 Ray tracing
Once the mode and direction of a particle are xed, it propagates through the
medium according to ray theory (e.g. Aki and Richards, 1980b). This includes the
interaction with interfaces at which the particles can be reected or transmitted
just like seismic waves. They can also change their mode if conversion occurs. In a
medium with constant mean velocity the particles move on straight lines. If there
are gradual changes of the mean velocity in the medium the paths are bent. The
algorithm is capable of dealing with any depth dependent velocity structure.
Interfaces
If a particle hits an interface between two media characterized by dierent mass
densities and wave velocities it can be reected, transmitted or converted. The
probabilities for the particle to undergo either of these events are given by the
respective energy reection coecients. These will be expressed in the notation of
Aki and Richards (1980b). Assuming horizontal interfaces acute and grave accents
are used to indicate up-going and down-going particles respectively. The general
form of an energy reection coecient is
(
A
B) =
o
v
o
cos(
o
)
i
v
i
cos(
i
)
_
A
B
_
2
. (2.56)
Here (
A
B) is the energy reection coecient i.e. the probability that an
incoming particle of type A with direction continues in direction as type B.
A
B
50 CHAPTER 2. MONTE-CARLO SIMULATIONS
is the reection coecient for displacement amplitudes. A and B stand for the wave
types P or S and and stand for the accents and ` indicating up- or down-going
directions. For an interface between two solids all 16 combinations of P, S, and` are
dened. and v indicate mass density and wave velocity respectively. is the angle
between the direction of the ray and the normal of the interface. Subscripts i and o
indicate the medium on the side of the incoming or outgoing particle respectively.
The velocity, however, does not only depend on the medium but also on the wave
type.
For horizontal interfaces the two dierent S-wave polarizations can be treated
separately. For horizontally polarized waves there is no conversion into other modes.
The reection coecients for displacement amplitudes for this case are given by Aki
and Richards (1980b) equation 5.33 p. 139. For vertically polarized S-waves and
P-waves which are coupled at the interface the amplitude reection coecients are
given by Aki and Richards (1980b) equations 5.40 p. 144-145. The interaction with
the free surface is a special case of these coecients for the case of zero velocity in
the upper layer. This causes total reection for horizontally polarized S-waves. The
displacement amplitude coecients for horizontally polarized S- and P-waves are
given in equations 5.27, 5.28 (p. 134), 5.31 and 5.32 (p. 136) by Aki and Richards
(1980b).
I treat supercritical reection i.e. S-energy reection for which the P-reection
angle would be greater than /2 as total reection similar to horizontally polar-
ized waves. Doing so I neglect phase shifts between the two S-polarizations upon
reection which could change the polarization type (e.g. from linear to elliptical).
Ray paths in a medium with vertical velocity variations
Ray theory predicts that in a laterally homogeneous medium the particles propagate
on curved rays in a vertical plane. In spherical coordinates the azimuthal angle
of the velocity vector remains constant and only the polar angle varies along the
path. The curvature of the ray is determined by the velocity gradient. In order to
use the usual notation of Snells law I introduce the incidence angle = for
as dened in gure 2.2. In this notation Snells law states that
sin()
v
= q = const. (2.57)
Here v is the wave velocity of either P- or S-waves and the constant q is called the
ray parameter. In order to keep q constant along a ray that passes through regions
with changing velocity the relative change of sin() must balance the relative change
of v. I thus have
d sin()
ds
sin()
=
dv
ds
v
(2.58)
where s parameterizes the length of the path. Considering the geometry of the ray
path in a medium with a vertical velocity gradient this leads to
d
ds
cos()
sin()
=
dv
dz
cos()
v
(2.59)
OUTLINE OF THE RadTrans ALGORITHM 51
and nally to
d
ds
=
dv
dz
sin()
v
(2.60)
where the last factor can be recognized as the ray parameter (equation 2.57). Equa-
tion 2.60 determines the change of the incidence angle along the ray. For a constant
velocity gradient dv/dz the ray path is a circle.
With equation 2.60 I can approximately model the particle propagation in
any depth dependent velocity structure by simply changing the component of the
velocity vector after the propagation of a path element l an amount of
dv
dz
sin()
v
l.
2.4.7 Recording of particles
The principal result of a simulation is the number density of particles which gives an
estimate of the energy density (equation 2.13). Since seismograms do not contain
information about the direction of seismic waves I do not accumulate information
about the direction of particles in the simulation. The total number of particles
is counted in a 4-dimensional array. This array represents rectangular boxes in 3-
dimensional space with a 4-th dimension for time. Snapshots of the energy eld are
taken in time intervals t. Like in a histogram the position of a particle is recorded
by adding its energy to the cell of the array in which the particle resides. This
way I obtain a spatio-temporal estimate of the local energy density as dened by
equation 2.1 for directions integrated over the whole solid angle.
2.4.8 Evaluation of probability distributions
A key point in Monte-Carlo simulations is the evaluation of probability distributions.
In the present algorithm random numbers have to be generated with probability
distributions for the direction of radiation from the source (eq. 2.17) and for the
scattering angles (eq. 2.34-2.38). Since, during a simulation a large quantity of
random numbers is needed, it is useful for the computational eciency to generate
tables with random numbers prior to the simulation.
I use an iterative trial and reject method to generate random numbers with
arbitrary probability distributions. This method does a random sampling of a given
probability distribution P(x). A set X of N random numbers x
i
[x
min
, x
max
],
i = 1, . . . , N having the distribution P(x) is selected as a subset of a uniformly
distributed set Y of random numbers y
j
[y
min
, y
max
], j = 1, . . . , M. The random
number y
j
belongs to X if z
j
P(y
j
)/P
max
with z
j
[0, 1] being another random
number from a uniformly distributed set. P(x)/P
max
is the normalized probability
distribution where P
max
is the maximum of P(x). As an approximate estimate of
P
max
I use P
max
= max
j=1,...,M
(P(y
j
)). This method is similar to the Metropolis-
Hastings algorithm (Hastings, 1970) with uniform proposal density.
The probability distributions encountered here depend on two angular vari-
ables. But as can be seen in equations 2.17 and 2.34-2.38 the two variables are
independent and the joint two-dimensional probability distribution is the product
52 CHAPTER 2. MONTE-CARLO SIMULATIONS
of the two one-dimensional distributions. However the trial and reject method as
explained above can be generalized to multi-dimensional probability distributions.
2.5 Conclusions
The algorithm is implemented in C programming language. The simple overall struc-
ture of the code allows to parallelize it by splitting the loop for the particles. Using
MPI it can be run in parallel on clusters of any size. The speed of the computation
strongly depends on the parameters of the random medium and it is impossible to
give general statements about the performance. The system requirements are de-
ned by the size of the 4-dimensional recording grid and the desired smoothness of
the energy density estimates in space and time. Models for regional wave propa-
gation over several hundred kilometers with a simulation length of several hundred
seconds require memory of a couple of GB. In combination with inversion procedures
that depend on the simulation of a large number of models, the application of this
algorithm certainly requires the use of large computer clusters.
The correctness of the implementation is veried in three dierent way. The
most simple property that is tested is energy conservation. For the Monte-Carlo
algorithm this means that the number of particles in the simulation remains con-
stant over time. In the case of zero intrinsic absorption this is conrmed also in the
presence of discontinuities. The second property that more deeply depends on the
physics of the radiative transfer is equipartition. It is veried that the energy ratio
between S- and P-energy converges to the ratio of 2
3
0
/
3
0
predicted by radiative
transfer theory (Ryzhik et al., 1996). The third test that is performed is a compar-
ison with nite dierence calculations. No signicant or systematic dierences are
observed between ensemble averaged energy density envelopes produced with nite
dierence calculations in full space and the envelopes produced with my algorithm.
Chapter 3
Elastic radiative transfer theory to
explain the Lg-blockage in the
Pyrenees
53
54 CHAPTER 3. LG-BLOCKAGE IN THE PYRENEES
Abstract
The RadTrans algorithm that is presented in chapter 2 is well suited to simulate
energy propagation in a model that contains both, deterministic structure and sta-
tistically described small scale heterogeneities. With this algorithm the propagation
of seismic waves through the Pyrenees is studied. Using a series of seismograms
the variations of the attenuation properties of the crust along the Pyrenean axis is
demonstrated. The western Pyrenees exhibit strong attenuation of crustal phases.
This phenomenon is known as Lg-blockage. Since the velocity structure of the west-
ern Pyrenees cannot account for the attenuation, the observation was unexplained.
Here I present a model that is able to explain the Lg-blockage. I propose that the
Lg-blockage is caused by a perturbed body in the crust that diers in the scattering
and attenuation properties from the surrounding crust. The velocities are identical.
The results show that it is possible to model the observed Lg-blockage with increased
scattering in the perturbed body. Stronger intrinsic attenuation than the surround-
ing crust improves the model further but intrinsic attenuation alone, cannot explain
the observations.
STRUCTURE AND EVOLUTION OF THE PYRENEES 55
3.1 Introduction
The Monte-Carlo algorithm that is introduced in chapter 2 allows to simulate full
seismogram envelopes of regional earthquakes in continental crust. It can be used
to improve existing methods for the estimation of medium or source/site parame-
ters like in chapter 1. Compared to such simple methods, the RadTrans algorithm
provides a much more realistic model, but needs more parameters to describe it. To
describe the elastic scattering in the way discussed in chapter 2, four parameters
are needed: the correlation length a, the strength of the uctuations , the quality
factor of intrinsic absorption Q, and to characterize the type of the auto-correlation
function in the case of a von K arm an medium, the parameter . This results in
eight unknown parameters to describe the energy propagation in crust and man-
tle. Due to this multitude of free parameters their precise estimation with sucient
condence appears very dicult.
Here a study is presented that has a dierent focus. The chapter starts with
an introduction to the geology and tectonics of the Pyrenees and describes the ob-
servation of anomalous attenuation of seismic waves in some parts of this mountain
range. The special capability of the RadTrans algorithm to simulate energy propa-
gation in a realistic crust-mantle structure is then used to test dierent models for
their ability to explain the observed attenuation.
3.2 Structure and evolution of the Pyrenees
The geological setting and tectonics of the Pyrenean range is characterized in the
review paper by Choukroune (1992) with the words: the Pyrenean belt cannot be
easily compared to other belts and is probably unique in its unusual structural style
and evolution. I therefore restrict to a simplied description which assumes an
east-west orientation of the tectonic units resulting from the collision between the
Iberian and European plates.
The central part of the Pyrenees is formed by the Paleozoic Axial Zone (PAZ).
This zone coincides with the main topographic expression of the Pyrenees. It is
composed of Hercynian rocks that were reactivated during the Alpine orogeny. This
zone is conned by the North Pyrenees Fault which marks the boundary to the
North Pyrenean Zone (NPZ) to the north. The NPZ is composed of Mezosoic ysch
deposits that are locally highly deformed and metamorphosed. Some large Paleozoic
outcrops form the North Pyrenean- and the Basque Massifs. Further north the NPZ
overrides the Aquitaine molasse basin along the North Pyrenean Frontal Thrust.
Almost symmetrically the South Pyrenean Zone (SPZ) connes the PAZ to the
south. The SPZ is composed of Mesozoic and Cenozoic sediments. At the South
Pyrenean Thrust the SPZ overrides the Ebro Basin.
Several seismic refraction and wide-angle reection proles revealed the Moho
structure beneath the Pyrenees. There is a jump in Moho depth between the 50 km
thick Iberian crust and the European crust which has a thickness between 29 km
and 35 km.
56 CHAPTER 3. LG-BLOCKAGE IN THE PYRENEES
Irrespective of this general north south structure that is common to the whole
range there are variations along the axis of the Pyrenees. The compressional motion
during the last 65 Ma is related to subduction of the Iberian lower crust beneath
the Aquitaine basin in the eastern and central parts of the Pyrenees (Souriau and
Granet, 1995). This observation, made with seismic tomography was conrmed by
magnetotelluric measurements (Pous et al., 1995) and gravity modeling (Vacher and
Souriau, 2001). In the western part the shortening has been accommodated in the
crust.
The lower crust in the western part of the NPZ has a much higher P-wave
velocity than the PAZ. Daigni`eres et al. (1981) speculate that this dierence is caused
by intrusion of mantle material into the lower crust. This dierence is not observed in
the eastern part of the range. The emplacement of lower crustal blocks into the upper
crust of the western Pyrenees has been detected by body wave tomography (Souriau
and Granet, 1995) and gravity modeling combined with petrological information
(Vacher and Souriau, 2001).
3.3 Data and observations
The geological variations in the Pyrenees along strike have an interesting manifesta-
tion in the attenuation properties of crustal waves. There is a pronounced dierence
between the western part and the eastern and central parts of the Pyrenees. The
western part strongly attenuates crustal waves (Chazalon et al., 1993). I illustrate
this observation using data from the seismic network run by the French Atomic
Energy Commission (CEA) and from the Spanish station PAB run by the Instituto
Geogr aco National, Spain. Figure 3.1 shows a map of the earthquakes and stations.
The bold red line with black dots along the Pyrenees marks the axis used to sort the
traces according to the place where the rays (blue lines) crossed the Pyrenees. The
dots indicate 50 km distance intervals. The corresponding waveforms are shown in
gure 3.2. They are band-passed between 2 and 4 Hz, scaled to group velocity, nor-
malized for their maximum and sorted for place of intersection with the axis of the
Pyrenees. The horizontal lines in gure 3.2 indicate the velocities of the dominant
regional seismic phases. These are the two crustal phases Pg (v
Pg
= 6.03km/s) and
Sg (v
Sg
= 3.56) and the two mantle phases Pn (v
Pn
= 8.16) and Sn (v
Sn
= 4.65).
The velocities are taken from the velocity model of the Laboratoire de Detection et
de Geophysique (LDG).
For wave paths that cross the eastern Pyrenees (east of the 250 km mark in
gure 3.2) seismograms are dominated by the crustal phases. The reverberations
in the crust form an extended Sg-arrival of high amplitude, the Lg-wave. The Sn-
arrival is barely visible and the Pn amplitude is small compared to the Pg phase.
In the western part of the Pyrenees the situation is dierent. The traces that cross
the Pyrenees west of the 200 km mark show almost no crustal phases. A strong
Pn-arrival is followed by Sn which usually has the highest amplitude. The Lg-wave
can hardly be recognized in the coda of the Sn-phase. This strong attenuation of
crustal phases, which is most obvious for the Lg-waves is referred to as Lg-blockage.
DATA AND OBSERVATIONS 57
6
o
W 4
o
W 2
o
W 0
o
2
o
E 4
o
E 6
o
E
39
o
N
42
o
N
45
o
N
48
o
N
51
o
N
Figure 3.1: Map of the study area with wave paths used in gure 3.2. Black triangles
indicate stations and red stars indicate earthquakes. Red line with black dots shows the
axis of the Pyrenees used to sort the traces in gure 3.2. The black dots indicate 50 km
intervals.
Another illustration of the Lg-blockage is presented in gure 3.3. Part A of
this gure shows records of six earthquakes that occurred north and south of the
Pyrenees made at the station SJPF inside the western Pyrenees. Part B shows
records of the same earthquakes made after the waves passed through the western
Pyrenees. A map with the events and the stations is shown in gure 3.4. The clear
Pg and Lg waves in gure 3.3A demonstrates that crustal phases are generated
by the sources and travel from both sides towards the station inside the Pyrenees.
Mantle phases are very weak. After passing through the Pyrenees the shape of the
seismograms is again completely dierent (gure 3.3B). Except for events 1 and
6 where a small amplitude increase is visible at the arrival time of the Lg phase,
there are no crustal phases. Using the interchangeability of source and receiver, this
exactly reects an observation made by the analysts at the LDG. Some earthquakes
in the western Pyrenees generate Lg-waves that can be recorded at both sides of the
range. But Lg-waves that where generated outside the Pyrenees do not propagate
through.
To generate traces that represent the main characteristics of the undisturbed
propagation through the easter Pyrenees and the propagation through the western
Pyrenees with blocked Lg-waves I use the following approach.
I split the collection of traces in gure 3.2 into groups according to the position
of intersection between the ray path and the axis of the Pyrenees. For the rst group
I select traces that cross the Pyrenean axis between the 100 km and 200 km marks.
I exclude wave paths west of the 100 km mark to not be biased by paths through
the Bay of Biscay. This group represents the propagation through the western
Pyrenees. The second group that represents the eastern part is made up of paths
east of the 300 km mark. The traces are band-passed between 2 and 4 Hz before the
envelopes are computed with a Hilbert transform. The envelopes are normalized for
their maximum, scaled to group velocity and stacked within each group. Figure 3.5
58 CHAPTER 3. LG-BLOCKAGE IN THE PYRENEES
0 50 100 150 200 250 300 350 400
8.16
6.03
4.65
3.56
3
2
g
r
o
u
p
v
e
l
o
c
i
t
y
[
k
m
/
s
]
position along axis of Pyrenees [km]
Figure 3.2: Collection of seismograms for wave paths through the Pyrenees. Traces are
sorted according to the position of the intersection between their paths and the axis of the
Pyrenees (red line in gure 3.1). The traces are band pass ltered between 2 and 4 Hz,
normalized for their maximum and scales to group velocity.
A
8.16 6.03 4.65 3.56 3 2
Event 6
Event 5
Event 4
Event 3
Event 2
Event 1
v [km/s]
B
8.16 6.03 4.65 3.56 3 2
v [km/s]
Figure 3.3: Seismograms of six earthquakes recorded within (A) the zone that blocks
Lg-waves and recorded after the waves passed through this zone (B). A map showing the
locations of the events and the stations is shown in gure 3.4.
shows the resulting envelopes. These clearly show the characteristics of Lg-blockage.
There are three mechanism known in the literature that could possibly explain
such an observation of variable amplitude ratio between crustal and mantle phases.
The mantle phases propagate in a high Q medium with little attenuation
whereas the crustal phases are stronger attenuated (Bormann, 2002). This
generally causes a distance dependence of the amplitude ratio. Indeed the
wave paths through the western Pyrenees are generally slightly longer than
the paths in the east but there are records with equal or reversed distance
relation. Also these traces with the same source and equal epicentral distance
DATA AND OBSERVATIONS 59
6
o
W 4
o
W 2
o
W 0
o
2
o
E 4
o
E 6
o
E
36
o
N
39
o
N
42
o
N
45
o
N
48
o
N
51
o
N
2
1
MFF
3
SJPF
4
5
6
PAB
Figure 3.4: Map of the study area with wave paths used in gure 3.3. Black triangles
indicate stations and red stars indicate earthquakes.
8.16 6.03 4.65 3.56 3 2
v [km/s]
w
e
s
t
e
r
n
P
y
r
e
n
e
e
s
e
a
s
t
e
r
n
P
y
r
e
n
e
e
s
Figure 3.5: Reference traces from the eastern and western Pyrenees Vertical lines indicate
arrival times of seismic phases. The upper black trace represents undisturbed propagation
through the eastern Pyrenees. It shows clear arrivals of crustal phases. The bottom
trace represents wave paths through the western Pyrenees. Here the crustal phases are
attenuated and only the mantle phases show clear arrivals.
show the Lg-blockage.
The radiation pattern of the source excites strong mantle phases and weak
crustal phases (Bormann, 2002). There are examples that conrm this situ-
ation but they seem rather rare. I use several earthquake sources to average
the observation over dierent sources in order to minimize this eect.
The velocity structure in the western Pyrenees causes the Lg-blockage.
The last point is worth a more detailed discussion.
The sensitivity of Lg-waves to crustal structure is known from the extinction of
this phase when passing through oceanic crust. Lg-blockage has also been been re-
ported from the Barents Sea (Baumgardt, 2001), eastern Mediterranean, the Caspian
60 CHAPTER 3. LG-BLOCKAGE IN THE PYRENEES
Sea, the Black Sea, the Red Sea (McNamara and Walter, 2001; Rodgers et al., 1997)
and the Lingurian Sea (Shapiro et al., 1996) and the Alpine range (Campillo et al.,
1993).
In most of these cases the attenuation is caused by soft marine sediments. They
form a layer of very low shear velocity and high attenuation that traps the energy
which propagates in the crust. This phenomenon of Lg-blockage by sedimentary
basins was modeled by Shapiro et al. (1996). Apart from minor details the velocity
structure explains the Lg-blockage. The Lg-blockage in the Alps and the Pyrenees
is dierent. There are sedimentary basins north and south of the Pyrenees but they
are not comparable in depth, extent and velocity to the marine situations. For the
specic situation of the western Pyrenean crust Chazalon et al. (1993) modeled the
wave propagation with a realistic velocity model including the Moho jump and high
velocity obstacles in the upper crust. Chazalon et al. (1993) conclude from their
modeling that neither the Moho jump nor the detailed crustal velocity structure can
cause the observed attenuation. The inability of the Moho jump to cause the Lg-
blockage is supported by the observed Lg propagation through the eastern Pyrenees
where the Moho jump is also present.
It can be summarized that that the observations reported in gure 3.2 reect
intrinsic properties of the propagation medium and not of the sources or the velocity
structure. There is no satisfactory explanation for the Lg-blockage by the Pyrenees
or the Alps. It was speculated by Chazalon et al. (1993) that the increased scattering
by small scale heterogeneities in the western Pyrenean crust causes the attenuation
of crustal phases.
3.4 Modeling
To test this hypothesis, the energy propagation through the Pyrenees is modeled
with elastic radiative transfer theory (RTT). This theory in the acoustic approxi-
mation was applied in a number of studies focusing on the separation of intrinsic
and scattering attenuation in simple half space models (e.g Hoshiba, 1991; Fehler
et al., 1992; Gusev and Abubakirov, 1996). Lacombe et al. (2003) use a model that
consists of a layer overlaying an innite half space for this purpose. Here the elastic
version of RTT is used in a complex geometrical model that consists of a scattering
layer over a scattering half space with a vertical velocity gradient. Additionally
a cuboid body with dierent scattering and attenuation properties is included in
the crustal layer to model the local eect of the Pyrenean crust. Because of this
complex model the main focus is not on the estimation of precise parameters of the
model. Instead I try to present a conceptual model to explain the phenomenon of
Lg-blockage in the Pyrenees.
The velocity model that is used for the simulation is based on the model of the
LDG. The crust is a constant velocity layer with density
crust
= 3.0 g/cm
3
, P-wave
velocity
crust
= 6.03 km/s and S-wave velocity
crust
= 3.56 km/s. The mantle
is represented by a half space with constant vertical velocity gradient and constant
density
mantle
= 3.5 g/cm
3
. At the Moho the P-wave velocity is
mantle
= 8.16 km/s
INVERSION 61
and S-wave speed
mantle
= 4.65 km/s. The velocity gradient for P-waves is
1
300
km/s
km
which is a approximate average of the velocity prole in the uppermost 600 km man-
tle in the Preliminary Reference Earth Model (PREM) (Dziewonski and Anderson,
1981). The velocity ratio is constant throughout the mantle. Small scale hetero-
geneities are described by an exponential autocorrelation function (equation 2.27)
with the parameters uctuation strength and correlation length a. The scattering
properties in crust and mantle are dierent but constant, i.e. the scattering proper-
ties do not vary with depth in the mantle. Intrinsic attenuation is described by the
quality factor. I x the ratio between the quality factor of P- and S-waves to the theo-
retical value that assumes that attenuation is limited to shearing: Q = Q
P
= 9/4Q
S
(Shearer, 1999, p. 114).
In this model I include an additional cuboid body in the crust that has the
same velocities and density as the surrounding crust but diers in the scattering
and attenuation properties. The body extends from the surface down to the Moho
and has a width in propagation direction of 100 km. This width corresponds to
the topographic expression of the Pyrenees. In the following I try to show whether
or not the attenuation and scattering properties of this body can account for the
Lg-wave blockage.
Since the reference traces are the result of averaging over several dierent
sources I use sources with isotropic radiation pattern. According to equation 2.19
the ratio between radiated P and S-energy is xed at 96% S-energy and 4% P-energy.
3.5 Inversion
The goal of this non-linear inversion process is to nd scattering and attenuation
parameters for crust, mantle and Pyrenean body that result in energy density traces
which t the eastern and western reference traces. Though some of the parameters
may be estimated independently the whole model has 9 free parameters, the uctu-
ation strength the correlation length a and the quality factor Q in three dierent
bodies. Here the type of the medium is already xed as I decided to restrict the
inversion to the exponential auto-correlation function. Because of the high number
of free parameters I apply a genetic algorithm for the inversion. The principal el-
ements of the algorithm are recombination of the parameters of the ttest models
and mutation to maintain diversity. Fitness of the models is linked to the mist
between the model and the reference trace. The mist is the sample wise average
of the absolute value of the dierence between the logarithm of the model envelope
and the logarithm of the reference. This is the L1-norm of the dierence between
the logarithmic traces. It ensures that the low amplitude coda has similar weight for
the mist calculation as the high amplitude direct arrivals and is not too sensitive
to uctuations in the data and the simulated envelopes.
62 CHAPTER 3. LG-BLOCKAGE IN THE PYRENEES
3.5.1 The undisturbed case
In a rst inversion I estimate the parameters of the background model. For this, the
crustal phases of the eastern reference trace are tted by adjusting the parameters
a, and Q of the crust in a model that has a transparent (no scattering) and
non-attenuating mantle and no additional body. The lapse time windows used to
calculate the mist in this inversion step are the Pg window between the Pg arrival
and the Sn arrival and the S-coda from the Sg arrival until a group velocity of
2 km/s. Similar to the generation of the reference traces I stack envelopes in a
certain distance range after scaling to equal group velocity. For comparison with
the eastern reference trace I stack records between 500 km and 800 km distance
from the source.
With this model of the crust I invert the western reference trace for scattering
and attenuation properties of the mantle. Because of the blockage of the crustal
phases the wave paths through the western Pyrenees oer the interesting possibility
to estimate mantle properties relatively independent of the crust. In this inversion
step I use time windows that contain the mantle phases, i.e. the Pn window be-
tween the Pn and the Pg arrivals and the Sn window between the Sn and the Sg
arrivals. Records between 600 km and 900 km are stacked in this case to account
for the slightly longer wave paths through the western Pyrenees. The results of this
inversion step are not well determined and a relatively wide range of values for the
parameters , a and Q can be used to t the reference trace. I therefore xed the pa-
rameter for the uctuation strength and correlation length of the mantle to = 0.02
and a = 2 km. This resulted in the quality factors Q
P
= 1070 and Q
S
= 475.
With these new parameters for the mantle I re-inverted the eastern refer-
ence trace for the crustal properties because the increased scattering in the mantle
changed the amount of energy that returns from the mantle into the crust. This
process raised the level of the coda and required changes of the parameters in the
crust. The nal parameters of the crustal material are = 0.021, a = 0.77 km,
Q
P
= 1400 and Q
S
= 623.
Figure 3.6 shows the result of this inversion. The general t of the model is
fairly good. Especially the decay of the S-coda is well tted. The kink in the S-coda
envelope shortly after 3 km/s originates from the reverberations in the crust. It
marks the end of the Lg-wave train which is well modeled by the algorithm. The
t is less good for the P-coda and the Pn-phase. The deviation from the shape of
the Pn-arrival is due to the fact that I estimated the mantle properties from the
western reference trace. I can better t the Pn-phase with stronger scattering in the
mantle but then the impulsive shape of the mantle phases in the western reference
trace cannot be reproduced. In fact this is an indication for slightly dierent mantle
properties under the eastern and western Pyrenees.
3.5.2 The disturbed case
To model the Lg-blockage I use the background model described in the previous
section and include an extra body with dierent scattering and attenuation prop-
INVERSION 63
8.16 6.03 4.65 3.56 3 2
v [km/s]
Figure 3.6: Simulation Results for the undisturbed propagation in through the eastern
Pyrenees. Bold red and black traces represent data and model respectively. Thin traces
indicate types of energy. Cyan trace shows radiated P-energy that arrives at the receiver
as P-energy. The dierence between the cyan and blue traces shows radiated P-energy
that arrives as S-energy. The dierence between blue and green traces shows radiated S-
energy that arrives as P. Radiated S-energy that arrives as S is indicated by the dierence
between the green and the black trace. Vertical lines indicate arrival times of seismic
phases
erties. To speed up the Monte-Carlo simulations I slightly modify this model to
enforce cylindrical symmetry which allows better averaging. In the modied model
the cuboid attenuating body is replaced by a cylinder ring centered at the source.
The radius of the inner boundary is 300 km and the width is 100 km. I veried that
the eect of this modication on the inversion is very weak.
Three dierent variants of the model are used for the inversion. In the rst
variant the uctuation strength of the small scale heterogeneities is adjusted to t
the western reference trace. The intrinsic attenuation and the correlation length
in the inclusion are equal to the surrounding crust. With = 11% the best t is
obtained (gure 3.7A). The general shape of the data is reasonably well tted. The
strong scattering in the inclusion eciently attenuates the crustal phases. There is
no Pg-arrival and the Sg-phase is only visible as a small bump in the Sn-coda. This
shows that scattering is a good candidate to cause the Lg-blockage in the Pyrenees.
But there is some mismatch between this model and the data in the S-coda which
has too much energy in this part of the envelope.
In a second variant of the model I assume that the Lg-blockage is caused by
intrinsic attenuation. The inclusion has the same scattering properties as the normal
crust but variable Q values. Again the ration between Q
P
and Q
S
is xed at 9/4.
The best t is obtained for Q
P
= 86 and Q
S
= 38. These values are extremely low.
Figure 3.7B shows the results of the model. The t is worse compared to the t
achieved by changing the scattering strength. The S-coda of the model is much too
low but there is still Pg-energy that traverses the inclusion. It is impossible to both
attenuate the Pg-phase strong enough but simultaneously increase the energy in the
64 CHAPTER 3. LG-BLOCKAGE IN THE PYRENEES
A
8.16 6.03 4.65 3.56 3 2
v [km/s]
B
8.16 6.03 4.65 3.56 3 2
v [km/s]
Figure 3.7: Simulation results for propagation through the western Pyrenees with
stronger scattering (A) and increased attenuation (B). For explanations see gure 3.6.
8.16 6.03 4.65 3.56 3 2
v [km/s]
Figure 3.8: Simulation results for propagation through the western Pyrenees with both
stronger scattering increased attenuation. For explanations see gure 3.6.
S-coda. Leaving the Q
P
/Q
S
-ratio free would allow to obtain a much better t. But
this would require much stronger P- than S-attenuation which is unrealistic. This
shows the attenuation alone does not suce to explain the Lg-blockage.
In the third model I allow for variable scattering and attenuation in the in-
clusion. With this model the best t is obtained (gure 3.8). The parameters are
Q
P
= 402, Q
S
= 179 and = 7.7%. With these parameters the crustal phases
can both be appropriately attenuated. The Pg-phase is weak enough not to the
stick out of the Pn-coda and simultaneously the small signature of the Lg-wave is
reproduced. Also the amplitude ratio between the Pn and Sn phases is reasonable.
For this model the parameters are listed in table 3.1.
I conclude that strong attenuation alone cannot explain the observation of Lg-
blockage in the Pyrenees. Increased scattering seems an essential ingredient of this
phenomenon but is not able to explain all details of the observation. In contrast the
observation can be well explained with increased scattering and stronger attenuation
DISCUSSION 65
Table 3.1: Parameters of the nal model used to explain the phenomenon of Lg-blockage
by the Pyrenees. The columns show values for the quality factor of P-waves (Q
P
) and
S-waves (Q
S
), the uctuation strength () and the correlation length (a) of the small scale
velocity uctuations and the transport mean free path l
S
(Turner, 1998). The values of
and a in the mantle are xed by hand. The correlation length a of the Pyrenean body
is assumed to equal the correlation length of the surrounding crustal material. Values in
brackets show representative numbers of the PREM (Dziewonski and Anderson, 1981).
The 750 km transport mean free path can be compared to the mean free path of 690 km
obtained in chapter 1.5 with the isotropic scattering model for the German crust.
Q
P
Q
S
a l
S
Crust 1400 (1350) 623 (600) 2.1% 0.77 km 750 km
Mantle 1070 (321) 475 (143) 2.0% 2.0 km 1530 km
Pyrenean body 402 179 7.2% 0.77 km 63 km
together.
3.6 Discussion
Compared to the waveform modeling by Chazalon et al. (1993), I use a simple struc-
tural model that consists of a at layer over an innite half space. It does neither
contain any deterministic velocity structure inside the crust nor the Moho topogra-
phy. Despite these limitations the shape of the seismogram envelopes in the eastern
Pyrenees can be reproduced. Including an extra body with stronger scattering and
attenuation the energy propagation can be inuenced to reproduce the seismogram
envelopes that are observed in the western Pyrenees. For the nal parameters I
illustrate the energy eld with two snapshots of the simulations in gures 3.9A and
B. These gures show the shadowing eect of the inclusion (marked by the black
box) in the western Pyrenees for crustal waves. The epicenter is indicated by the
star. The linear color-scale represents the square root of seismogram envelopes, i.e.
the fourth root of energy. Figure 3.9A displays the energy distribution at a lapse
time of 105 s. At this time the P-phases already passed the Pyrenees. The Pn
mantle phase which propagated north to about 47
and 45
i
curve. In the approach presented
here there is no need to adjust such algorithmic parameters. Since the distortion
along the trace is accounted for by the stretching or compression, the similarity
78 CHAPTER 4. PASSIVE IMAGE INTERFEROMETRY AT MT. MERAPI
between the two traces can be evaluated along the whole trace. This means that
the correlation coecient can be calculated in a single time window ranging from
the beginning of the trace until the signal reaches the noise level. This makes our
approach more robust and sensitive.
If the time delay is caused by a spatially homogeneous relative velocity change
v/v, the relative delay time, is independent of the lapse time at which it is measured
and v/v = /. In this case error estimates are obtained by choosing consecutive
non-overlapping time windows in which the correlations are calculated separately.
This gives independent estimates of the relative delay time and mean value and
standard deviation can be calculated.
4.5 Application to Merapi volcano
Passive image interferometry is applied to almost two years of nearly continuous
seismic records from Merapi volcano, Indonesia. Velocity changes are estimated for
each one-day GF (between stations GRW0 and GRW1) with respect to a reference
trace, assuming that v/v is spatially homogeneous. As reference trace a stack of all
one-day GF in January 1998 is used. Figure 4.6A shows the daily relative velocity
variation obtained from the LL component of the GT between 2 and 8 s lapse time.
I do not use lapse times smaller than 2 s because the surface waves contained in
the GFs prior to 2 s have a dierent spatial sensitivity to velocity variations. Gray
shading in gure 4.6A marks the standard deviation of independent measurements
in three consecutive 2 s time windows starting at 2, 4, and 6 seconds lapse time.
Seismic velocity shows temporal variations of a few percent with a clear seasonal
trend. Except for a constant oset, the curve only marginally depends on the period
used to generate the reference trace. However stacking the GFs from a longer period
of time may result in a poorer correlation of the reference trace with the daily
correlations because the velocity variations that occurred in the longer period would
degrade the reference trace. January 1998 is chosen because the velocity appeared
to be stable during this period.
Evidence for the fact that the inferred velocity variations are related to phys-
ical properties of the subsurface and not artefacts introduced by the sensors or the
recording system is provided by the analysis of records from dierent components
and stations. In gure 4.7A the velocity changes measured from the transverse com-
ponents are shown. The results are virtually identical to the ones obtained from the
LL-component. An interesting property of our new method for practical applica-
tions in geophysics is that it can be applied to records from a single station using
the auto-correlation function. Figure 4.7B shows the velocity variations measured
on the auto-correlation of the vertical component of station GRW0. Again the curve
is very similar to the one in gure 4.6A.
The blue curve in gure 4.6A shows the results of CWI measurements (Wegler
et al., 2006b) made at the same stations (GRW) but with an active source (BAT
in gure 4.1). Results from the southern ank of Merapi volcano (station KEN,
source BEB) are shown by the green curve. Considering the daily variability of the
4.6. HYDROLOGICAL MODELING 79
measurements and the source receiver distance of more than 3 km used for the active
measurement, I conclude that the present measurements are in agreement with the
results of the active experiment. The advantage of my new method is that changes
can be monitored continuously without an expensive active source.
Actually the velocity changes can directly be seen in the GFs. Figure 4.2B and
4.3B show close-ups around the phases marked by the arrows in gures 4.2A and
4.3A respectively. The coherent phases tend to arrive earlier in late May than in
early May 1999. This corresponds to the almost linear velocity increase of about
0.015% per day in May 1999 that can be observed in gure 4.6A.
4.6 Depth dependence of the velocity changes and
hydrological modeling
The periodicity of the velocity uctuations of approximately one year in this tropical
environment suggests a climatic inuence, most likely by precipitation. Another
observation also supports a connection with precipitation: The amplitude of the
velocity variations depends on the lapse time at which it is measured. This has
never been reported for CWI measurements before and indicates that the velocity
perturbations are spatially inhomogeneous. I can explain this observation together
with details of the temporal variations with a hydrological model of the ground
water level (GWL).
An illustration of the hydrological model and its inuence on the seismic veloc-
ity structure is shown in gure 4.8. I assume that drainage of ground water occurs
through a stationary aquifer that can approximately be described by Darcys law.
Then the drainage is proportional to the height of the ground water table which
results in an exponential decrease of the water level after rain events. A convolution
of the precipitation rates with this exponential function thus gives the ground water
level GWL (below surface) at time t
i
:
GWL(t
i
) = GWL
0
n=0
p(t
n
)
e
a(t
i
tn)
. (4.1)
Here is porosity, a is the parameter describing the decay, GWL
0
is the asymptotic
water level, and p(t
i
) denotes the daily precipitation. Good agreement of such a
model was found with water level measurements in a well (Akasaka and Nakanishi,
2000). Daily precipitation data were provided by the NASA/Goddard Space Flight
Centers Laboratory for Atmospheres (Human et al., 2001).
To relate the ground water level with delay times, I dene the depth (z) and
time (t) dependent relative slowness perturbation S(t
i
, z) and a reference water level
GWL
ref
that is chosen equal to the mean level of January 1998. Then S(t
i
, z) = s
for GWL(t
i
) < z < GWL
ref
, S(t
i
, z) = s for GWL
ref
< z < GWL(t
i
) and
S(t
i
, z) = 0 elsewhere. s is the relative slowness dierence between the states
80 CHAPTER 4. PASSIVE IMAGE INTERFEROMETRY AT MT. MERAPI
saturated and dry. The delay time at time t
i
measured at lapse time is then
(t
i
, ) =
0
_
z=
S(t
i
, z)K(z, )dz (4.2)
where K(z, ) is the sensitivity kernel (Pacheco and Snieder, 2005). I use the kernel
in the approximation of coincident source and receiver with a diusivity constant of
seismic energy of D = 0.05 km/s
2
as estimated at Mt. Merapi (Wegler and L uhr,
2001). The kernel decays with distance r from the station as r
1
exp(r
2
/D).
For the one dimensional structure of the velocity perturbation resulting from our
hydrological model, gure 4.9 shows the exact depth dependent sensitivity kernels
for lapse times = 3s and = 7s.
With a genetic algorithm I found the best model with the following parameters:
GWL
0
= 51 m, = 0.03, a = 0.008 d
1
, and s = 0.17. Figure 4.6B shows the daily
rainfall, the inferred GWL and the modeled apparent homogeneous velocity varia-
tion for the 24 s and 68 s lapse time windows together with the measured values.
Considering that the precipitation data are averages over an area of approximately
100 by 100 km and the simplicity of our hydrological model the velocity variations
can be explained in remarkable detail. The model also explains the dierence be-
tween the blue and the red curves in gure 4.6B that represent the measurements at
dierent lapse times. I am thus condent that the observed velocity variations are
related to precipitation via changes in the ground water level. The successful expla-
nation of the lapse time dependence of the relative delay times by depth dependent
velocity perturbations indicates that the coda in the frequency range above 0.5 Hz
is made up of scattered body waves rather than surface waves. In contrast to body
waves the depth prole of the sensitivity kernel of surface waves does not change
with lapse time.
For spatially inhomogeneous velocity perturbations, the accuracy of the method
can be characterized by the scatter of measurements around a common trend. In
rain-free periods like the rst three month in the study period I obtain a standard
deviation of 0.1%.
Assuming that the annual rainfall is similar each year the observation can
be compared with measurements from 1991 (Ratdomopurbo and Poupinet, 1995).
Indeed a velocity increase of a little more than 1% was observed during the period
from April until September 1991. During April to October 1998 and April to June
1999 we observe a similar increase. According to my interpretation these periods
of increasing velocities are caused by a falling ground water level in the dry season.
I conclude that the eect of stress changes inside the volcano on velocity changes
(Ratdomopurbo and Poupinet, 1995; Wegler et al., 2006b) is secondary.
4.7 Conclusions
The present study shows that the hydrological conditions can change the seismic
velocities by more than 10% in a depth range of several tens of meters. This eect
4.7. CONCLUSIONS 81
should be taken into account in studies of the local velocity structure. With the
hydrological model an independent check for PII is provided. It proves that de-
terministic information about the propagation medium is contained in the GFs at
large lapse times, i.e. one can retrieve scattered and reected waves by correlation of
seismic noise. This provides the basis for passive seismic imaging with non-ballistic
waves. I anticipate this observation to stimulate a variety of studies in reection
seismology using correlations of seismic noise.
The temporal velocity variations detected at Mt. Merapi in this study are
dominated by shallow perturbations and demonstrate the potential of PII. If the
scattering is weaker or the changes close to the stations are smaller, the sensitivity
kernel allows the detection of more distant changes. In a lower frequency range the
GF are dominated by surface waves. The depth sensitivity of surface waves is very
dierent from body waves and opens new possibilities for passive image interfer-
ometry. The is a variety of possible applications for PII (Snieder and Page, 2007).
Monitoring of the exploitation process of geothermal or hydrocarbon reservoirs with
PII seems possible as well as the monitoring of nuclear waste disposal sites. I think
that PII can be used for routine monitoring in various engineering, geotechnical,
and geophysical applications.
In a separate study a drop of seismic velocity associated with the Mid-Niigata
earthquake in Japan is reported that was detected with PII (Wegler and Sens-
Sch onfelder, 2007). Crucial for the results presented here is the precise relative
timing of the two sensors. But it is shown that similar results can be obtained from
just a single station and in chapter 5.2 introduces a technique to measure clock
errors of seismic stations. This means that PII can be applied to data from any
existing permanent seismograph station and we expect other interesting discoveries
from the application of PII in dierent environments.
82 CHAPTER 4. PASSIVE IMAGE INTERFEROMETRY AT MT. MERAPI
0.04
0.03
0.02
0.01
0
0.01
0.02
0.03
0.04
d
v
/
v
Jun Jul Aug
0.01
0.005
0
Sep Nov 1998 Mar May Jul Sep Nov 1999 Mar May Jul
40
30
20
10
0
0.02
0.01
0
0.01
0.02
G
W
L
[
m
]
v
/
v
0
30
60
p
r
e
c
i
p
i
t
a
t
i
o
n
[
m
m
/
d
]
A
B
Figure 4.6: Measured and modeled velocity variations at Merapi volcano. A: Red dots
mark the measurements in the 28 s time window. Individual measurements in the 24 s,
46 s, and 68 s seconds lapse time windows and their standard deviation are indicated
by light gray dots and gray shading respectively. Results of the active CWI experiment
(Wegler et al., 2006b) are indicated by green and blue lines. Inset shows a close up around
these active measurements. The blue line corresponds to the measurements at the GRW
station (used here). Green line shows measurements from station KEN on the southern
ank of Mt. Merapi. B: Bottom part: daily precipitation rate (blue) and modeled ground
water level (black). Top part: measured (light red dots) and modeled (dark red line)
velocity variations in the 24 s lapse time window. Same for the 68 s lapse time window
in blue.
4.7. CONCLUSIONS 83
0.03
0.02
0.01
0
0.01
0.02
0.03
d
v
/
v
Sep Nov 1998 Mar May Jul Sep Nov 1999 Mar May Jul
0.03
0.02
0.01
0
0.01
0.02
0.03
d
v
/
v
A
B
Figure 4.7: A: Velocity variations measured from the TT-component of the GT between
the stations GRW0 and GRW1 in the 2-8 s lapse time window. B: Velocity variation
measured on the auto-correlation of the Z-component from station GRW1.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
precipitation p(t)
drainage d(t)
v
s
d
e
p
t
h
0
G
W
L
v
Figure 4.8: Illustration of the hydrological model that is used to explain the observed
velocity variations. On the left the velocity structure is sketched that shows a discontinuity
at the depth of the ground water level. The hight of the ground water level on the right
hand side is determined by the interplay between precipitation and drainage though a
Darcy-like aquifer.
84 CHAPTER 4. PASSIVE IMAGE INTERFEROMETRY AT MT. MERAPI
0 2 4
x 10
4
0
0.5
1
K(z)
d
e
p
t
h
[
k
m
]
= 3s
= 7s
Figure 4.9: Amplitudes of the depth dependent sensitivity kernels for = 3 s and = 7 s
lapse times.
Chapter 5
Daily velocity variations and
instrumental errors at Merapi
volcano
85
86 CHAPTER 4. PASSIVE IMAGE INTERFEROMETRY AT MT. MERAPI
Abstract
Two case studies are presented that demonstrate the use of passive imaging for dif-
ferent seismological purposes. In the rst example the temporal resolution of Passive
Image Interferometry is improved to obtain measurements of temporal changes on
a sub one day scale. Periodic changes of the Greens functions are found that can
be related to temperature induced changes of the seismic velocities. In the sec-
ond example Greens functions on paths across Merapi volcano are retrieved. It is
shown how these Greens functions can be used to identify and correct instrumental
errors in a network of four stations. A polarity reversal of the vertical component
at station GRW0 between the 25th September and the 25th October is identied.
Measurements of the clock dierences between station pairs are performed and used
to obtain detailed information about the timing errors of individual stations. These
show short osets of about 1 s as well as periods of constant clock drifts. Relative
timing with this method can exceed the precision of an internal digitizer clock al-
ready after the clock drift of a single day. The results are extendable to networks of
arbitrary size.
IMPROVING PASSIVE IMAGE INTERFEROMETRY 87
5.1 Improving the temporal resolution of passive
image interferometry
The accuracy of techniques that infer information from passive imaging depends on
the convergence rate of the eld correlations toward a stable approximation of the
Greens function (GF). This rate is determined by the characteristics of the wave
eld that is used. The ideal eld would be equipartitioned in phase space meaning
that all modes and directions are equally present (e.g. S anchez-Sesma and Campillo,
2006). This is seldom fullled. Especially the low frequency ambient seismic noise,
that is often used for passive imaging in seismology, is excited predominantly in the
oceans (Sabra et al., 2005; Stehly et al., 2006). Thus the sources have preferred
locations and frequencies. To compensate for this shortcoming the correlations are
usually averaged over a long period of time. For structural investigations this does
not pose a problem because it is assumed that the medium does not change with
time. In monitoring applications, however, the time needed for the noise correlations
to converge denes the temporal resolution of the method. It is thus desirable to
decrease the necessary averaging time in order to increase the temporal resolution
of the monitoring.
5.1.1 Hourly Greens functions
In chapter 4 I have shown that a temporal resolution of one day can be achieved with
Passive Image Interferometry. The GFs obtained from 24 hours of continuous noise
records at stations GRW0 and GRW1 (gure 4.1), as they were used in chapter 4, are
shown in detail in gure 5.1. In contrast to the processing in chapter 4 I apply a one-
bit normalization (Larose et al., 2004) before the correlation. The noise records are
correlated in segments of one hour length. After correlation the 24 traces obtained
for each day are stacked. The section in gure 5.1 consists of more than 600 traces
(one for each day) with amplitudes mapped on a color scale. Several observations
can be made in this gure.
The shape of the traces is very similar during the whole period of time. Promi-
nent phases such as the ones at 3 s and 7 s lapse time can easily be traced
over the 2 years. This indicates that the medium remains stationary, i.e. the
scatterers that generate the coda do not move.
The velocity change reported in gure 4.6 can be observed in detail in the
GFs. The Phase at 7 s lapse time for instance, exactly resembles the shape of
the velocity variations shown in gure 4.6A.
The delays for positive and negative delay times are symmetric. This indicates
that the delays are indeed caused by medium changes and not by instrumental
errors.
The amplitudes of some phases in the GFs changes in July 1998. This change
probably reects variations in the characteristic of the seismic noise, that was
88 CHAPTER 5. DAILY VELOCITY VARIATIONS / INSTRUMENT ERRORS
l
a
p
s
e
t
i
m
e
[
s
]
Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun
8
6
4
2
0
2
4
6
8
Figure 5.1: Time section of 24-hour Greens functions between stations GRW0 and
GRW1. There is one Greens function for every day from beginning of August 1997 until
mid June 1999.
inuenced by rock falls and accelerated seismic activity during a volcanic crisis.
This change gradually disappears until beginning of January.
Figure 5.2 shows GFs obtained from correlations of one hour seismic noise.
This is similar to gure 5.1 except that no averaging of the 1-hour GFs is made.
There are more than 16000 GFs plotted in gure 5.2, one for each hour in which data
is available. The similarity between gure 5.2 and gure 5.1 is remarkable. This
indicates that even with averaging over only one hour of seismic noise, the correlation
function appears to be a reasonable approximation of the GF. I therefore perform
the same measurements as presented in chapter 4 on the one-hour GFs.
IMPROVING PASSIVE IMAGE INTERFEROMETRY 89
l
a
p
s
e
t
i
m
e
[
s
]
Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun
8
6
4
2
0
2
4
6
8
Figure 5.2: Time section of 1-hour Greens functions between stations GRW0 and GRW1.
There is one Greens function for every hour from beginning of August 1997 until mid June
1999.
5.1.2 Hourly velocity variations
In gure 5.3 the velocity variations are shown for each one-hour GF together with
the velocity variations obtained from GF that are averaged over one day. The blue
curve is similar to gure 4.6A, except that the measurements are performed for the
Z-component. The oset between end September and mid October 1997 is due to
a polarity reversal at station GRW0 (cf. chapter 5.2). Compared to these daily
measurements the red dots that represent the hourly measurements appear to show
stronger uctuations. A closer look at these uctuations reveals some periodicity.
Figure 5.4A shows an enlarged segment of the curve in gure 5.3. For comparison
the GF for this period are shown in gure 5.4B. The similarity between the the
measured velocity variations and the shape of the Greens functions in gure 5.4 is
not as striking as for the annual changes. This is due to the less averaging of the
90 CHAPTER 5. DAILY VELOCITY VARIATIONS / INSTRUMENT ERRORS
Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul
0.02
0.015
0.01
0.005
0
0.005
0.01
0.015
0.02
v
/
v
Figure 5.3: Hourly velocity variations measured between stations GRW0 and GRW1
from August 1997 until mid June 1999. Red dots are measured from 1-hour GFs and blue
line indicates the measurements obtained from GFs averaged over 24 hours.
A
03/15 03/16 03/17 03/18 03/19 03/20 03/21 03/22
15
10
5
0
5
x 10
3
v
/
v
B
l
a
p
s
e
t
i
m
e
[
s
]
03/15 03/16 03/17 03/18 03/19 03/20 03/21
8
6
4
2
0
2
4
6
8
Figure 5.4: A: Enlarged period (15-22 March 1998) of the hourly velocity variations
measured between stations GRW0 and GRW1. Red line is measured from 1-hour GFs
and blue line indicates the measurements obtained from GFs averaged over 24 hours. B:
1-hour Greens functions for this period.
IMPROVING PASSIVE IMAGE INTERFEROMETRY 91
0 2 4 6 8 10 12
velocity variations
0 2 4 6 8 10 12
air pressure variations
0 2 4 6 8 10 12
air temperature variations
0 2 4 6 8 10 12
precipitation rate
frequency [1/day]
Figure 5.5: Power spectra of hourly velocity variations and environmental processes that
might inuence the Greens functions. The vertical scale is logarithmic.
one-hour GFs. In fact also the amplitude of the GRs seems to uctuate periodically.
This indicates that the noise sources are not constant.
To identify the cause of these uctuations a spectral analysis of the velocity
variations is performed. For this analysis, velocity measurements are only used when
the correlation of the GF with the reference trace exceeds 0.4. Values with lower
correlation are substituted by linear interpolation of adjacent vales. In gure 5.5 the
power spectrum of the velocity variations is compared with dierent environmental
processes that might cause the changes in the GFs. These are rain, air temperature,
and air pressure for which data form Merapi was provided by Malte Westerhaus.
The spectrum of the velocity variations shows a dominant peak at one cycle per
day and 4 higher harmonics. There is a signicant dierence between the amplitude
of the 24 h period peak and the 12 h peak. The spectrum of the air pressure
uctuations has peaks at 24 h period and higher harmonics. In contrast to the
velocity variations the amplitudes of the 24 h and 12 h peaks are similar. It is a well
known characteristic of air pressure variations that the semi-diurnal oscillation has
equal or higher amplitude then the diurnal. Air temperature variations again show
dominant uctuations with 24 h period and higher harmonics. The 12 h period
peak is of considerably lower amplitude than the 24 h peak. This is similar to the
spectrum of the velocity variations. The precipitation rate does only show one peak
at 24 h period. Higher harmonics are missing. Since the 24 h periodicity seems
to be the fundamental oscillation in all these time series I stack each of the series
92 CHAPTER 5. DAILY VELOCITY VARIATIONS / INSTRUMENT ERRORS
0 5 10 15 20
t [h]
velocity variation
air temperature
air pressure
rain
Figure 5.6: Stack of the diurnal velocity variations, air-pressure, air temperature and
precipitation rate. Y-axis shows scaled amplitudes. Time axis is Coordinated Universal
Time (UTC) which is 7 h behind local time at Merapi.
with a 24 h periodicity to compare the phases. Therefor each time series is cut into
consecutive segments of 24 h length. The stacks of these segments are shown in
gure 5.6. Taking into account the 7 h dierence between UTC used in the gure
and local time at Merapi, on can recognize the expected shape of the air-temperature
and -pressure variations. The temperature reaches its maximum at about noon and
its minimum before sunrise. The frequency doubling of the pressure variations is
also evident. The precipitation assumes its maximum in the afternoon. The velocity
variations have a shape similar to the temperature variations and the oscillations
are approximately in phase. The exact phase dierence between the air-temperature
variations and the velocity variations is 1 h with the temperature variations being
in advance.
5.1.3 Conclusions
With the increased temporal resolution of PII I identied periodic variations in the
GFs. These variations concern the amplitude and the shape of the GFs. Applying
coda wave interferometry to these GFs reveals periodic uctuations of the seismic
velocity. The frequency spectra of the velocity variations and precipitation are
clearly dierent. The 12 h oscillation and the higher harmonics are missing in the
precipitation data. I conclude that the diurnal velocity variations are not caused by
or related to precipitation.
The spectra of air-pressure and -temperature variations are more similar to
the velocity variations. In particular the amplitude ratio between the 24 h and 12 h
peaks argues for a relation between the velocity variations and air-temperature. This
is conrmed by the shape of the diurnal stack of the data which shows similarity
between temperature and velocity data.
A possible mechanism that could cause variations of the seismic velocity as
response to temperature variations is thermal stressing. The associated temperature
changes in the subsurface are certainly limited to a very shallow layer but the stress
IMPROVING PASSIVE IMAGE INTERFEROMETRY 93
changes aect a larger volume. A combination of dierent causes like air pressure
and temperature can not be excluded with the present analysis. A detailed modeling
like in chapter 4 for the annual changes is necessary to clearly identify the cause of
the diurnal changes.
94 CHAPTER 5. DAILY VELOCITY VARIATIONS / INSTRUMENT ERRORS
5.2 Detecting instrumental errors with passive
imaging
One of the most dicult tasks of operating a seismic network is the verication of
data quality. Especially the data that is taken home from temporal network does
not always have the expected characteristics. An interesting example is presented by
Miller et al. (1998) who deployed a temporal network for the estimation of moment
tensors and tomography. They had to detonate an explosion to check the polarities
of the sensors of which 25% were wrong. Leaving these 25% uncorrected would be
disastrous for moment tensor estimation. Though tomography and event location
experiments require high timing precision, the timing problems especially of tempo-
ral stations are well known. Here a method is presented that uses passive imaging
to verify the timing and polarity of a seismic station directly from the data.
5.2.1 Cross-volcano Greens functions
To estimate GFs between between distant stations a lower frequency band is to be
used. The strong attenuation of high frequency waves destroys the correlation of
the waveeld on a larger distance. For the distance of 5 km between the Merapi
arrays GRW and KEN, the frequency band between 0.1 Hz and 10 Hz proved to be
useful for the short period stations and between 0.01 Hz 10 Hz for the broadband
stations. Since this frequency band contains the microseismic peaks there is some
variability in the power spectrum that is smoothed by spectral whitening (Bensen
et al., 2007) before the correlation. Figure 5.7 shows the color coded GF between
the stations GRW0 and KEN1 for the vertical components.
The following observations can be made.
1. The GFs show a good coherence over the whole period of time.
2. The GFs are not symmetric about zero.
3. Between end the 25th of September and the 15th of October 1997 the polarity
of one of the instruments is reversed.
4. Between March 1998 and October 1999 the individual phases show a saw-tooth
pattern.
Item one shows that there is a good reproducibility of the GFs. Their shape
does only marginally depend on the time of the year. This in agreement with Stehly
et al. (2006) who found stable noise sources in the frequency band 0.1 - 0.2 Hz.
GFs obtained from noise with this characteristics are well suited for the study of
temporal changes. The asymmetry of the GFs mentioned in item two above indicates
that either noise sources are unequally distributed or that the scatters which act as
secondary sources have preferred locations (Paul et al., 2005; Stehly et al., 2006).
The sign reversal mentioned in item three was already observed in chapter 5.1.2 for
correlations between stations GRW0 and GRW1. This means that the reversal most
DETECTING INSTRUMENTAL ERRORS WITH PASSIVE IMAGING 95
l
a
p
s
e
t
i
m
e
[
s
]
Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May
40
30
20
10
0
10
20
30
40
Figure 5.7: Time section of the GF between stations GRW0 and KEN1 for the period
August 1997 until May 1999. The amplitudes are linearly amplied from zero lapse time
towards the ends.
probably occurred at GRW0. In contrast to the symmetric changes observed in
section 5.1.1 the changes in gure 5.7 have the same sense for positive and negative
lapse times. They are thus of dierent origin.
All these observations can be made even at late lapse times such as 100 s.
Figure 5.8 shows a time section of a late phase in the GFs. This phase is found in
all GFs between a GRW-station and a KEN-station.
5.2.2 Instrumental errors
Temporal changes that aect positive and negative times in the same way are due to
errors in one of the stations clock (Stehly et al., 2007). To estimate the clock errors
of the individual stations I try to optimally use the conguration of the network. The
network consists of two arrays: GRW and KEN (gure 4.1). Each array consists of
four three-component stations (see gure 4.1 for the GRW array), that are pairwise
connected to a 6-channel digitizer. Thus there are four digitizers with independent
clocks and GPS timing system. Digitizers are named A with sensors KEN0 and
KEN1, B with sensors KEN2 and KEN3, C with sensors GRW0 and GRW1 and
D with sensors GRW2 and GRW3. To estimate to clock dierence between two
digitizers, e.g. A and C, I use the GFs between the four sensor pairs that are
connected to the two digitizers, i.e. A1-C1, A1-C2, A2-C1, A2-C2, where A1 and
A2 symbolize the two sensors connected with digitizer A.
I estimate the clock dierences in a multi step procedure.
1. Calculation of a reference trace by stacking all GF for each sensor pair
96 CHAPTER 5. DAILY VELOCITY VARIATIONS / INSTRUMENT ERRORS
l
a
p
s
e
t
i
m
e
[
s
]
Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May
120
115
110
105
100
95
90
85
80
Figure 5.8: Time section of the GF between stations GRW0 and KEN1 for the period
August 1997 until May 1999.
2. Measurement of the clock dierence as the position of the maximum of the
cross-correlation function between the individual GFs and the reference traces
for each sensor pair
3. Combination of the clock dierence measurements from the four sensor pairs
4. Calculation of a new reference trace for each sensor pair like in 1 but shifted
to account for the clock dierence
5. Measurement of the clock dierence for each sensor pair like in 2 but with the
new reference trace.
6. Combination of the new clock dierence measurements from the four sensor
pairs
Steps four up to six are necessary because the reference trace generated in step 1 is
altered by the clock errors. These steps can be repeated several times to improve the
reference trace, but after one iteration the improvements are marginal. For the cross-
correlation in items 2 and 4 I use the complete GFs between -100 and 100 s lapse time
including direct and scattered waves. Because of the larger amplitudes the cross-
correlation functions are dominated by the direct waves. For the combination of
the clock dierence measurements from the dierent sensor pairs a simple average is
calculated between measurements that fulll the following criterion: The correlation
between the reference trace and the GF, that was shifted to account for the measured
clock dierence exceeds 0.4. The number of traces that are used in this averaging
can be increased by using more components of the Greens tensors between the
DETECTING INSTRUMENTAL ERRORS WITH PASSIVE IMAGING 97
station pairs. In fact in the present case one could use 36 component pairs for this
averaging. Figure 5.9 shows the clock dierence measurements for all digitizer pairs.
Similar measurements were obtained for stations in Southern California by
Stehly et al. (2007). This information is useful for analysis of the station pairs, like
ambient noise tomography (Shapiro et al., 2005). But it does not help to improve
event localization that is strongly dependent on precise timing information of in-
dividual stations. To provide this information the following approach is used here.
The relation between the clock dierences shown in gure 5.9 and the clock errors
of individual stations can be written in matrix form:
_
_
_
_
_
_
_
_
1 1 0 0
1 0 1 0
1 0 0 1
0 1 1 0
0 1 0 1
0 0 1 1
_
_
_
_
_
_
_
_
_
_
_
_
D
_
_
_
_
=
_
_
_
_
_
_
_
_
AB
AC
AD
BC
BD
CD
_
_
_
_
_
_
_
_
(5.1)
This equation is valid for each point in time where the clock dierences are known.
Here
X
represents the clock error of digitizer X and
XY
represents the clock dier-
ence between digitizers X and Y. Equation 5.1 is a classical least squares inversion
problem of the form Gm = d with m being the unknown model vector and d be-
ing the known data. The system is solved in the least squares sense to obtain the
unknown clock osets. These are shown in gure 5.10. Since it is still possible to
add an equal oset to all clocks, the curves are tied to the timing of digitizer C. Its
clock error is set to zero. Actually gure 5.9 shows no signature of a clock correction
that is common to all clock dierence measurements that contain digitizer C. This
indicates that digitizer C either had a perfect clock or was, which is more likely,
continuously locked to the GPS signal.
The curves in gure 5.10 show the characteristics of the individual clocks. Dig-
itizer C probably has continuous GPS reception. Digitizer D also has a good GPS
but the worst clock. The GPS only lost the connection to the satellites in August
1998 but the clock has a drift of 6.71e-7 or of 1.74 s/month. Digitizers A and B
have clock drifts of 1.5e-7 and 9.33e-8 respectively. They lost their GPS connection
between March and October 1998. Between December 1997 and March 1998, the
digitizers A and B of the array KEN show frequent spikes of about 1 s amplitude
that can be attributed to instrumental errors. The two spikes of digitizer A are
common with digitizer B. This indicates that these errors do not occur randomly
but are externally triggered. Probably the meteorological conditions caused poor
GPS reception which led to misinterpretation of thepulseper second GPS signal
at both digitizers at KEN. The accuracy of the method depends on the distance
between the receivers, because of the damping of high frequency components over
longer distances. For stations pairs within one of the small arrays, i.e. with dis-
tances of about 200 m, the standard deviation of the clock dierence is about 10 ms.
For longer distances of about 5 km across the volcano the standard deviation of the
clock dierence measurements is of the order of 50 ms.
98 CHAPTER 5. DAILY VELOCITY VARIATIONS / INSTRUMENT ERRORS
Sep Nov 1998 Mar May Jul Sep Nov 1999 Mar May
1
0.5
0
0.5
1
c
l
o
c
k
d
i
f
f
e
r
e
n
c
e
[
s
]
C1A1
C1A2
C2A1
C2A2
average
Sep Nov 1998 Mar May Jul Sep Nov 1999 Mar May
1
0.5
0
0.5
1
c
l
o
c
k
d
i
f
f
e
r
e
n
c
e
[
s
]
C1B1
C1B2
C2B1
C2B2
average
Sep Nov 1998 Mar May Jul Sep Nov 1999 Mar May
1
0.5
0
0.5
1
c
l
o
c
k
d
i
f
f
e
r
e
n
c
e
[
s
]
C1D1
C1D2
C2D1
C2D2
average
Sep Nov 1998 Mar May Jul Sep Nov 1999 Mar May
1
0.5
0
0.5
1
c
l
o
c
k
d
i
f
f
e
r
e
n
c
e
[
s
]
D1A1
D1A2
D2A1
D2A2
average
Sep Nov 1998 Mar May Jul Sep Nov 1999 Mar May
1
0.5
0
0.5
1
c
l
o
c
k
d
i
f
f
e
r
e
n
c
e
[
s
]
D1B1
D1B2
D2B1
D2B2
average
Sep Nov 1998 Mar May Jul Sep Nov 1999 Mar May
1
0.5
0
0.5
1
c
l
o
c
k
d
i
f
f
e
r
e
n
c
e
[
s
]
A1B1
A1B2
A2B1
A2B2
average
Figure 5.9: Clock dierences between the digitizers of the GRW and KEN stations. Dots
indicate the measurements between the dierent sensor pairs of a given digitizer pair. Red
line shows the combination of the measurements from the corresponding sensor pairs.
DETECTING INSTRUMENTAL ERRORS WITH PASSIVE IMAGING 99
Sep Nov 1998 Mar May Jul Sep Nov 1999 Mar May
0.5
0
0.5
0.5
0
0.5
0.5
0
0.5
0.5
0
0.5
Digitizer D
Digitizer C
Digitizer B
Digitizer A
c
l
o
c
k
o
f
f
s
e
t
[
s
]
Figure 5.10: Clock errors of the four digitizer clocks
5.2.3 Conclusions
The method presented here provides information about the polarity of the seismic
records and about the clock osets. This information is very valuable for the veri-
cation and analysis of seismic data but it can also be obtained in a more traditional
way. The polarity can for example be checked by comparison of rst motions of
teleseismic arrivals or by explosions as done by Miller et al. (1998). In principle also
the curves that show the clock oset in gure 5.10 can be reproduced by analysis of
GPS-logles that usually record GPS reception and clock corrections. Nevertheless I
think it is very useful to be able to get all these information directly from the data,
without the need for external information from logles or seismological analysis.
Especially for the operation of temporal networks with dicult GPS conditions like
in high latitudes, mountainous areas or on the seaoor the possibility to correct the
timing may prove very valuable. In OBS networks one can obtain precise relative
timing and average the clock drifts of the whole network which will increase the
accuracy of the absolute timing. Onshore one can relate the need for GPS reception
at all stations since a single station without GPS can easily be synchronized to the
GPS-locked clocks of other stations. The accuracy of 10 ms for small receiver dis-
tances (200 m) and 50 ms for distances of 5 km can already outstrip the accuracy
of internal digitizer clocks after a single day.
100CHAPTER 5. DAILY VELOCITY VARIATIONS / INSTRUMENT ERRORS
Summary
This work contains a number of studies that extract information from scattered seis-
mic waves. In part I the amplitude of the seismic coda is analyzed with radiative
transfer theory. The rst study introduces an approach to simultaneously invert
the seismic coda for intrinsic attenuation, scattering strength, source spectrum, and
site amplication factors. This analysis is based on radiative transfer in the acous-
tic approximation with isotropic scattering. It is shown that the estimated seismic
moments are in good agreement with results from waveform modeling, that site am-
plications correspond to geology, that the transport mean free path in the German
is 690 km and that the quality factor of intrinsic attenuation is of the order of 500.
A Monte-Carlo algorithm is presented that was developed to solve the radiative
transfer equation for elastic waves. It accounts for non-isotropic scattering and
conversion between P- and S-modes. The large scale velocity structure of the crust
is taken into account in terms of dierent velocities in crust and mantle, and the
discontinuity in-between, that leads to conversion, reection, and transmission of
seismic energy. Reection and conversion at the free surface is also treated. A
vertical velocity gradient in the mantle can be used to generate mantle phases like
Pn and Sn.
The study on Lg-blockage applies this algorithm to explain the reason of the
strong attenuation of crustal phases in the western Pyrenees. In simulations of the
radiative transfer process in a 3-dimensional structure it is shown that increased
heterogeneity in a localized region below the western Pyrenees is able to explain the
Lg-blockage. The t of the model can be improved by additionally increasing the
intrinsic attenuation in the western Pyrenean body. Intrinsic attenuation alone can
not account for the observed attenuation.
The second part of this work uses the phase information of the seismic coda.
Additionally the seismic coda is not simply measured in a classical sourcereceiver
conguration, but retrieved from the ambient seismic noise by cross-correlation of
the signals from two receivers. This process of Greens function retrieval is based
on the phase information of the seismic noise.
In the rst study of data from Merapi Volcano it is shown that the passively
retrieved Greens functions can be retrieved from noise records of 24 h length. This
allows to obtain a seismic record independent of active sources once a day. Based
on this nding a monitoring technique named passive image interferometry is intro-
duced. It allows to precisely monitor velocity changes in the subsurface. Application
to noise records from Merapi volcano shows velocity variations of a few percent on
a time scale of weeks. A strong seasonal trend is observed. In a sophisticated in-
101
102 SUMMARY
version process, the parameters of a hydrological model are estimated that relates
the measured velocity variations with precipitation induced changes of the ground
water level. With the nal parameters the hydrological model can explain the shape
of the velocity variations and their lapse time dependent amplitudes solely based on
precipitation data.
The temporal resolution of passive image interferometry is increased in another
study from one day to one hour. Based on spectral comparison with various mete-
orological parameters, this study reports evidence for a relation of the fast velocity
variations with atmospheric temperature.
The last study introduces a method to assess the accuracy of instrument timing
and setup directly from records of seismic noise. The signature of timing errors on
the passively retrieved Greens functions is dierent from the signature of velocity
variations, that were analyzed in the studies above. The method above is applied
to data from four digitizers from the Merapi network. It is shown how the clock
dierence of each instrument pair can be estimated from the noise records. With a
least squares inversion the clock oset of each digitizer is obtained.
Bibliography
Abubakirov, I. and Gusev, A. (1990). Estimation of scattering properties of the
lithosphere of Kamchatka based on Monte-Carlo simulation of record envelope of
a near earthquake. Phys. Earth Planet. Int., 64(1):5267.
Abubakirov, I. R. (2005). Attenuation characteristics of transverse waves in the
lithoshere of Kamchatka estiamted from observations at the Petropavlovsk digital
broadband station. Izvestiya, Physics of the Solid Earth, 41(10):813824.
Akasaka, C. and Nakanishi, S. (2000). Correction of background gravity changes due
to precipitation: Oguni geothermal eld, Japan. Proceedings World Geothermal
Congress, pages 24712475.
Aki, K. (1980). Attenuation of shear-waves in the lithosphere for frequencies from
0.05 to 25 Hz. Phys. Earth Planet. Inter., 21(1):5060.
Aki, K. (1992). Scattering conversions P to S versus S to P. Bull. Seismol. Soc.
Am., 82(4):19691972.
Aki, K. and Chouet, B. (1975). Origin of coda waves: source, attenuation, and
scattering eects. J. Geophys. Res., 80(23):33223342.
Aki, K. and Richards, P. G. (1980a). Quantitative Seismology, Theory and Methods,
volume II. W. H. Freeman and Company, San Francisco.
Aki, K. and Richards, P. G. (1980b). Quantitative Seismology, Theory and Methods,
volume I. W. H. Freeman and Company, San Francisco.
Apresyan, L. and Kravtsov, Y. A. (1996). Radiation Transfer: Statistical and Wave
Aspects. Gordon and Breach, Amsterdam.
Baumgardt, D. R. (2001). Sedimentary basins and the blockage of lg wave propa-
gation in the continents. Pure Appl. Geophys., 158:12071250.
Bensen, G. D., Ritzwoller, M. H., Barmin, M. P., Levshin, A. L., Lin, F., Moschetti,
M. P., Shapiro, N. M., and Yang, Y. (2007). Processing seismic ambient noise data
to obtain reliable broad-band surface wave dispersion measurements. Geophys. J.
Int., 169(3):12391260. doi: 10.1111/j.1365-246X.2007.03374.x.
103
104 BIBLIOGRAPHY
Bianco, F., Del Pezzo, E., Castellano, M., Ibanez, J., and Di Luccio, F. (2002). Sep-
aration of intrinsic and scattering seismic attenuation in the Southern Apennine
zone, Italy. Geophys. J. Int., 150:1022.
Birch, F. (1961). The velocity of compressional waves in rocks to 10 kilobars, part
2. J. Geophys. Res., 66:21992224.
Bormann, P., editor (2002). New Manual of Seismological Observatory Practice.
GFZ, IASPAI, Potsdam.
Braunmiller, J. (2002). Moment tensor solutions of stronger earthquakes in Ger-
many with GRSN data. In Korn, M., editor, Ten Years of German Regional
Seismic Network (GRSN), number Report 25, pages 227235. Senate Commisson
for Geoscience of the Deutsche Forschungsgemeinschaft.
Braunmiller, J., Dahm, T., and Bonjer, K. (1994). Source mechanism of the 1992
Roermond earthquake from surface wave inversion of regional data. Geophys. J.
Int., 116:663672.
Braunmiller, J., Kradolfer, U., Baer, M., and Giardini, D. (2002). Regional mo-
ment tensor determination in the European-Mediterranean area - initial results.
Tectonophysics, 356:522.
Buttkus, B. (2000). Spectral Analysis and Filter Theory in Applied Geophysics.
Springer-Verlag Berlin, Heidelberg.
Campillo, M., Feigner, B., Bouchon, M., and Bethoux, N. (1993). Attenuation of
crustal waves across the Alpine range. J. Geophys. Res., 98(B2):19871996.
Campillo, M. and Paul, A. (2003). Long-range correlations in the diuse seismic
coda. Science, 299:547 549. doi:10.1126/science.1078551.
Chandrasekhar, S. (1960). Radiative Transfer. Dover, New York.
Chazalon, A., Campillo, M., Gibson, R., and Carreno, E. (1993). Crustal wave
propagation anomaly across the pyrenean range. comparison between observations
and numerical simulations. Geophys. J. Int., 115:829838.
Choukroune, P. (1992). Tectonic Evolution of the Pyrenees. Ann. Rev. Earth Planet.
Sci., 20:143158.
Daigni`eres, M., Gallart, J., and Banda, E. (1981). Lateral variations of the crust in
the North Pyrenean Zone. Ann. Geophys.
Dainty, A. M. and Toks oz, M. N. (1981). Seismic codas on the earth and the moon:
a comparison. Phys. Earth Planet. Int., 26:250260.
Derode, A., Larose, E., Campillo, M., and Fink, M. (2003). How to esti-
mate the Greens function of a heterogeneous medium between two passive
sensors? Application to acoustic waves. Appl. Phys. Lett., 83(15):30543056.
doi:10.1063/1.1617373.
BIBLIOGRAPHY 105
Dewberry, R. S. and Crosson, R. S. (1995). Source scaling and moment estimation
for the Pacic northwest network using S-coda amplitudes. Bull. Seismol. Soc.
Am., 85(5):13091326.
Duvall, T. L., Jeferies, S. M., Harvey, J. W., and Pomerantz, M. A. (1993). Time-
distance helioseismology. Nature, 362:430 432. doi:10.1038/362430a0.
Dziewonski, A. M. and Anderson, D. L. (1981). Preliminary reference earth model.
Phys. Earth Planet. Int., 25:297356.
Fehler, M., Hoshiba, M., Sato, H., and Obara, K. (1992). Separation of scattering
and intrinsic attenuation for the Kanto-Tokai region, Japan, using measurements
of S-wave energy versus hypocentral distance. Geophys. J. Int., 108(3):787800.
Gret, A., Snieder, R., and Scales, J. (2006). Time-lapse monitoring of rock
properties with coda wave interferometry. J. Geophys. Res., 111(B03305).
doi:10.1029/2004JB003354.
Gusev, A. A. and Abubakirov, I. R. (1987). Monte-carlo simulation of record enve-
lope of a near earthquake. Phys. Earth Planet. Int., 49:3036.
Gusev, A. A. and Abubakirov, I. R. (1996). Simulated envelopes of non-isotropically
scattered body waves as compared to observed ones: another manifestation of
fractal heterogeneity. Geophys. J. Int., 127(1):4960.
Hastings, W. K. (1970). Monte Carlo sampling methods using Markov chains and
their applications. Biometrika, 57:97109.
Hennino, R.and Tr
v
/
v
0
30
60
p
r
e
c
i
p
i
t
a
t
i
o
n
[
m
m
/
d
]
A
B
Abbildung 2:
Anderungen der seismischen Geschwindigkeit am Vulkan Merapi (A) und
deren hydrologische Modellierung (B)
In einer nichtlinearen Inversion werden die Parameter eines hier entwickelten hy-
drologischen Modells angepasst, das in der Lage ist, die gemessenen Geschwin-
digkeitsanderungen durch Schwankungen im Grundwasserspiegel zu erklaren. Die
Vorwartsrechung dieser Inversion nutzt ortsabhangige Sensitivitatsfunktionen und
beschreibt neben dem Verlauf auch die Laufzeitabhangigen Amplituden der Ge-
schwindigkeitsvariationen. Die Anpassung des hydrologischen Modells an die Ge-
schwindigkeitsvariationen wird in Abbildung 2B verdeutlicht. Im unteren Teil der
Abbildung ist die tagliche Niederschlagsmenge in Blau und der daraus bestimm-
te Grundwasserspiegel in Schwarz dargestellt. Im oberen Teil sind die gemessenen
(Punkte) und vom Modell vorhergesagten (durchgezogene Linien) Geschwindigkeits-
variationen f ur zwei Laufzeitfenster dargestellt. Rot entspricht Laufzeiten zwischen
2-4 s und Blau zwischen 6-8 s. Die gute
Ubereinstimmung dieser Kurven belegt,
dass das hydrologische Modell den korrekten physikalischen Zusammenhang zwi-
schen Niederschlag und
Anderungen der seismischen Geschwindigkeit beschreibt.
Dies wiederum zeigt, dass die hier vorgestellte Methode tatsachlich in der Lage ist,
geringf ugige
Anderungen der seismischen Geschwindigkeit zu uberwachen. Die Me-
thode ist einfach, auf existierende Daten anwendbar und hat ein groes Potential
f ur Anwendungen in anderen Bereichen (Snieder and Page, 2007).
120 SUMMARY
Tagliche Variationen und Instrumentenfehler
Hier werden zwei weitere Anwendungen der Methode zur Gewinnung von Green-
schen Funktionen aus seismischem Umgebungsrauschen f ur Veranderungsmessung-
en vorgestellt. Die erste Untersuchung dieses Kapitels stellt eine Verbesserung der
zeitlichen Auosung der Interferometrie rauschgenerierter Signale vor. Sie beschreibt
anhand der Daten vom Vulkan Merapi tageszeitliche
Anderungen der seismischen
Geschwindigkeit und deren spektrale Zusammensetzung. Der Vergleich mit den Am-
plitudenspektren verschiedener meteorologischer Groen und deren Phasen deutet
darauf hin, dass die tageszeitlichen Schwankungen mit
Anderungen der Lufttempe-
ratur zusammenhangen.
Eine andere Art von Veranderungen wird in der zweiten Studie mittels rauschge-
nerierter Seismogramme untersucht. Hier wird der Eekt von Zeitfehlern in den
Seismometern betrachtet. Es werden die Zeitdierenzen zwischen den paarweise un-
tersuchten Seismometern bestimmt. Eine Matrixinversion f uhrt anschlieend zu den
Zeitfehlern der einzelnen Seismometer. Diese Methode bietet die Moglichkeit, die
Genauigkeit der Zeithaltung unabhangig von Zusatzinformationen direkt anhand
der Daten zu uberwachen. Sie ist damit ideal geeignet, um die Zeithaltung einzelner
Seismometer auch im Nachhinein zu korrigieren.
Ausblick
Mit der Energietransfertheorie und der Nutzung von Phasenkorrelationen im seismi-
schen Rauschen werden zwei Methoden verwendet, die in der Seismologie ein groes
Potential f ur weitere Anwendungen und neue Entwicklungen haben.
Die Simulation der Energieausbreitung in einem kombinierten Modell aus statistisch
beschriebenen kleinskaligen Heterogenitaten und deterministisch beschriebener ma-
kroskopischer Struktur erlaubt weitergehende Untersuchungen von Problemen, in
denen Diskontinuitaten eine Rolle spielen. Unter anderem liee sich die Ausbreitung
teleseismisch einfallender Energie in der Erdkruste studieren, die sehr empndlich
f ur die Eigenschaften des Wellenleiters ist. Der nachste Schritt bei der Entwick-
lung von Algorithmen zur Simulation der Energieausbreitung in der Erde sollte die
Behandlung von Oberachenwellen und deren Kopplung mit Raumwellen sein.
Die Interferometrie rauschgenerierter Seismogramme kann in verschiedensten, weit
uber die Seismologie hinausgehenden, Bereichen angewendet werden, bei denen keine
aktiven Quellen eingesetzt werden konnen. Die Kombination dieser Methode mit
Lokalisierungstechniken f ur die Quelle der Geschwindigkeitsvariationen scheint hier
eine sinnvolle Entwicklung zu sein.
Literaturverzeichnis
Aki, K. and Richards, P. G. (1980). Quantitative Seismology, Theory and Methods,
volume I. W. H. Freeman and Company, San Francisco.
Campillo, M. and Paul, A. (2003). Long-range correlations in the diuse seismic
coda. Science, 299:547 549. doi:10.1126/science.1078551.
Ryzhik, L., Papanicolaou, G., and Keller, J. B. (1996). Transport equations for
elastic and other waves in random media. Wave Motion, 24:327370.
Sens-Schonfelder, C. and Wegler, U. (2006a). Passive image interferomety and sea-
sonal variations of seismic velocities at Merapi volcano, Indonesia. Geophys. Res.
Lett., 33:L21302. doi: 10.1029/2006GL027797.
Sens-Schonfelder, C. and Wegler, U. (2006b). Radiative transfer theory for estimati-
on of the seismic moment. Geophys. J. Int., 167(3):13631372. doi: 10.1111/j.1365-
246X.2006.03139.x.
Snieder, R. and Page, J. (2007). Multiple scattering in evolving media. Phys. Today,
60(4):4955.
121
Scientic developement
Christoph Sens-Sch onfelder
10/1997 09/2003 Studies in geophysics at the Universit at Leipzig, (Leipzig,
Germany)
05/2000 09/2003 Fellow of the German National Merit Foundation (Studien-
stiftung des deutschen Volkes)
08/2000 06/2001 Studies in geophysics at the H ask oli
Islands (Reykjavk,
Iceland)
09/2003 Diploma
01/2004 07/2004 Scientic co-worker at the Universit at Kiel (Kiel,
Germany)
since 08/2004 Scientic co-worker at the Universit
1
2
Leipzig (Leipzig,
Germany)
11/2006 07/2007 Research visit at the Laboratoire de Geophysique Interne
et Tectonophysique (Grenoble, France) with support of the
German Academic Exchange Service (DAAD)
03/2007 Granting of the Gnter Bock award by the Deutsche Geo-
physikalische Gesellschaft (German Geophysical Society)
for an outstanding publication