Geophysical Journal International

Geophys. J. Int. (2014) 197, 483–494 doi: 10.1093/gji/ggt519

Advance Access publication 2014 January 16

Time-reverse modelling of acoustic wave propagation

in attenuating media

Tieyuan Zhu
Department of Geophysics, Stanford University, Stanford, CA 94305, USA. E-mail: tyzhu@stanford.edu

Accepted 2013 December 19. Received 2013 November 7; in original form 2013 September 11

SUMMARY

Downloaded from http://gji.oxfordjournals.org/ at University of Texas at Austin on December 29, 2014

Time-reverse modelling (TRM) of acoustic wave propagation has been widely implemented in
seismic migration and time-reversal source imaging. The basic assumption of this modelling
is that the wave equation is time-invariant in non-attenuating media. In the Earth, attenuation
often invalidates this assumption of time-invariance. To overcome this problem, I propose a
TRM approach that compensates for attenuation and dispersion effects during the wave prop-
agation in attenuating media. This approach is based on a viscoacoustic wave equation which
explicitly separates attenuation and dispersion following a constant-Q model. Compensating
for attenuation and dispersion during TRM is achieved by reversing the sign of the attenua-
tion operator coefficient while leaving the counterpart dispersion parameter unchanged in this
viscoacoustic wave equation. A low-pass filter is included to avoid amplifying high-frequency
noise during TRM. I demonstrate the effects of the filter on the attenuation and the phase
velocity by comparing with theoretical solutions in a 1-D Pierre shale homogeneous medium.
Three synthetic examples are used to demonstrate the feasibility of attenuation compensation
during TRM. The first example uses a 1-D homogeneous model to demonstrate the accuracy
of the numerical implementation of the methodology. The second example shows the applica-
bility of source location using a 2-D layering model. The last example uses a 2-D cross-well
synthetic experiment to show that the methodology can also be implemented in conjunction
with reverse-time migration to image subsurface reflectors. When attenuation compensation

GJI Seismology
is included, I find improved estimation of the source location, the excitation timing of the point
source, the magnitude of the focused source wavelet and the reflectivity image of reflectors,
particularly for deep structures underneath strongly attenuating zones.
Key words: Seismic attenuation; Computational seismology; Wave propagation.

cate microseismic sources for reservoir monitoring (e.g. Lu et al.

I N T RO D U C T I O N
In general, time-reverse modelling (TRM) of acoustic wave prop-
agation consists of two steps: (1) reversing recorded seismic data
in time and (2) backpropagating the time-reversed data as sources
from receiver locations through an appropriate earth model. This
method has been successfully implemented in many applications in
seismology and exploration geophysics. For example, time-reversal
source imaging was suggested by McMechan (1982) to backprop-
agate seismograms from receivers to find the focusing energy, that
is, the location of sources. Further applications have since been
reported for the determination of temporal and spatial parameters
of earthquake sources (McMechan et al. 1985; Kao & Shan 2004;
Lu 2007; Kremers et al. 2011), earthquake source patterns (Hu &
McMechan 1987), moment tensor inversion (Kawakatsu & Mon-
tagner 2008; O’Brien et al. 2011), tremors (Larmat et al. 2006; Lok-
mer et al. 2009) and location of an active seismic source (Gajewski
& Tessmer 2005). More recently, TRM has been applied to lo-
2008; Steiner et al. 2008; Larmat et al. 2009). Another example is
reverse-time migration (RTM) which adds two additional steps to
the TRM procedure: propagating the wavefield from the source and
further applying an imaging condition to obtain the structure inter-
face (Baysal et al. 1983; McMechan 1983). These studies, whether
using simple acoustic or even elastic anisotropic wave equations,
did not consider attenuation in their forward and backward (reverse)
modelling.
In a non-attenuating medium, a wave equation is time-invariant
for TRM. In other words, with perfect receiver sampling, the
backward-propagated wavefield mirrors the forward-propagated
wavefield in time. In the 1-D experiment shown in Fig. 1(a), we
can see that the final focused energy is able to reconstruct the ini-
tial source, which is physically supported by the assumption of a
lossless medium (Fink 2006).
However, when intrinsic attenuation is considered in the Earth, the
wave equation is no longer time-invariant under time reversal (Fink


C The Author 2014. Published by Oxford University Press on behalf of The Royal Astronomical Society. 483
484 T. Zhu
Figure 1. Basic principles of forward propagation and time-reverse propagation in a 1-D experiment in (a) non-attenuating media and (b) attenuating media.
In the non-attenuating medium, the initial source wavelet could be retrieved. In the presence of attenuation, however, time reversibility of the wave equation
was broken.
2006), although it still satisfies source–receiver reciprocity (Snieder
2007). This implies that time-reversed waves that are backpropa-
gated into such a medium will no longer be symmetric in time under
the same condition as for a non-attenuating medium. Fig. 1(b) il-
lustrates how the presence of attenuation breaks the time reversibil-
ity of the wave equation. An initial source wavelet at the source
propagates through an attenuating medium and the resulting data
are recorded with a receiver with reduced amplitude and distorted
phase. In TRM, we inject the recorded wave in reversed time as a
new source from the receiver position. The wave backpropagates
through the medium and is attenuated even further. Despite the per-
fect receiver sampling, the backpropagated wavefield will not be
symmetric with the forward-propagated wavefield in time, and the
refocused source turns out to be different from the initial source.
If, however, an appropriate compensation for the intrinsic attenua-
tion is applied to the backward-propagated wavefields during TRM,
the time-reversal invariance can be (mostly) recovered (Gosselet &
Singh 2007). By compensating in this way for attenuation and dis-
persion effects during TRM, we should expect to recover the correct
amplitude and phase of the backward-propagated wavefield as well
as the final reconstructed source.
Some recent works propose to compensate attenuation and dis-
persion in the time-reversal process in acoustic bioimaging (Treeby
et al. 2010; Ammari et al. 2011). However, not much attention has
been received in geophysical community. Limited studies have at-
tempted to perform TRM with attenuation compensation in RTM
algorithms. These studies used the damping viscoscalar wave equa-
tion (Deng & McMechan 2007) and the viscoelastic wave equation
based on the standard linear solid (Deng & McMechan 2008). Nu-
merical results showed a general improvement in the RTM image
after attenuation compensation. However, this modelling approach
is limited by the coupled attenuation and dispersion operators. As
can be seen, when attenuation compensation is implemented by
applying a negative sign to the attenuation term, the dispersion
will not be correctly compensated. Zhang et al. (2010) derived a
viscoacoustic wave equation based on constant-Q model and then
modified it to extrapolate the source and receiver wavefields. How-
ever, their derivation for the wave equation is not clearly described
in the abstract. It is more or less ad hoc to introduce the normalized
operators for attenuation compensation during the backward mod-
elling. Moreover, the methods described above were intentionally
designed and evaluated for RTM, and it is not clear how to apply
these approaches to time-reversal imaging.
Downloaded from http://gji.oxfordjournals.org/ at University of Texas at Austin on December 29, 2014
The purpose of this paper is to develop a rigorous and general
TRM approach for attenuating media and to test this TRM ap-
proach using synthetic seismic experiments. In our previous study
(Zhu & Harris 2014), we derive a time domain, nearly constant-
Q viscoacoustic wave equation for modelling wave propagation in
attenuating media. This viscoacoustic wave equation can model
approximate constant-Q attenuation and dispersion behaviour of
seismic waves in the relevant seismic frequency band. More im-
portantly, it decouples attenuation and dispersion effects, a definite
advantage for attenuation compensation during TRM (Treeby et al.
2010). Based on this equation, I present an adjoint viscoacoustic
wave equation to compensate for attenuation and dispersion effects
by simply changing the sign of attenuation operator. The implemen-
tation flow for TRM with attenuation compensation is presented in
the following section. Next, I show that this TRM procedure will
be stabilized by applying a low-pass filter to attenuation and disper-
sion operators in the spatial frequency domain. Finally, I validate the
TRM approach in a 1-D homogeneous model, and show two applica-
tions to illustrate the effectiveness of compensating for attenuation
and dispersion effects for time-reversal imaging and RTM.
METHODOLOGY OF TRM IN
AT T E N UAT I N G M E D I A
Constant-Q model
For seismic modelling and imaging, attenuation (proportional to
1/Q) is considered to be approximately linear with frequency in
many observational frequency bands (McDonal et al. 1958; Kjar-
tansson 1979; Aki & Richards 1980), that is, Q is constant over
these frequency ranges. Kjartansson (1979) explicitly gave a lin-
ear description of attenuation that exhibits the exact constant-Q
characteristic. The dispersive phase velocity and attenuation of the
constant-Q model are given by Kjartansson (1979) as follow:
 γ
ω
c P = c0 , (1)
ω0
 πγ  ω
α = tan , (2)
2 cP
where the velocity c0 is given at a reference frequency ω0 , the
parameter γ = 1/π tan−1 (1/Q) is dimensionless, and we know 0 <
γ < 0.5 for any positive value of Q. Hence, the quality factor Q

is frequency-independent, that is, constant Q. Note that constant
Q is not exactly equivalent to assuming that α is proportional to
frequency since c P is slightly dependent on frequency.
Forward modelling
Based on the constant-Q model (Kjartansson 1979), the time do-
main nearly constant-Q (visco)acoustic wave equation in attenuat-
ing media was first introduced by Zhu & Harris (2014). It is written
as
1 ∂ 2 pF ∂
= ηL pF + τ H pF ,
c0 ∂t
2 2 ∂t
where the source wavelets emit at source positions (xs , z s ),
pF (xs , z s , t) = p0 (xs , z s , t),
and the two fractional Laplacian operators are L =
(−∇ 2 )γ +1 , H = (−∇ 2 )γ +1/2 ; pF is the spatial pressure field of
forward modelling in time t from 0 to T, and c0 is the acoustic
velocity at the reference frequency ω0 . The attenuation and
dispersion operator coefficients are given by
2γ −2γ 2γ −1 −2γ
η = −c0 ω0 cos πγ and τ = −c0 ω0 sin πγ ,
−1
where the variable γ is defined as γ = 1/π tan (1/Q). The range
of γ is 0 < γ < 1/2 with any positive Q. The first term in the right-
hand side of eq. (3) is related to dispersion effects, and the second
is related to attenuation effects (Zhu & Harris 2014). The most
attractive feature of this viscoacoustic wave equation is the explicit
separation of attenuation and dispersion (eq. 3). This is in contrast
to other wave equations in which the attenuation and dispersion are
encapsulated by a single term (Deng & McMechan 2007, 2008;
Carcione 2010). In particular, in the following section, I show the
viscoacoustic wave equation to be advantageous for compensating
for attenuation and dispersion effects during TRM.
Time-reverse modelling
Mathematically, time reverse involves replacing time t by −t. There-
fore, I write the constant-Q wave equation (eq. 3) in reversed time
as follows:
1 ∂ 2 pB ∂ within the filter passband. Frequencies above our cut-off curves are
= ηL p B + τ H pB ,
c02 ∂(−t)2 ∂(−t)
where the boundary condition is
p B (xr , zr , t) = pF (xr , zr , T − t).
Here, p B is the spatial pressure wavefields of TRM in time t, and
pF (xr , zr , T − t) is the data recorded at time T − t at the receiver
positions.
This TRM wave equation is not identical to the forward mod-
elling wave equation (eq. 3), because the first-order time derivative,
representing attenuation effects breaks the temporal symmetry of
equation (eq. 3). However, if I simply change the sign of the atten-
uation term in eq. (6), I have
1 ∂ pB2
∂
= ηL p B − τ H pB ,
c02 ∂(−t)2 ∂(−t)
∂
which becomes a time-invariant wave equation for TRM since it is
the same as the forward modelling in eq. (3). Thus, with pF (x, t) as
the solution of the forward modelling equation (eq. 3), p B (x, −t) is
a solution of eq. (3) with the sign of the attenuation term reversed.
Time-reverse modelling in attenuating media 485
The negative sign functions to compensate for attenuation during
the propagation of the reversed wavefields. I emphasize that the
first, dispersion-related term on the right-hand side of eq. (8) is time-
independent and does not reverse sign (i.e. the frequency-dependent
phase velocity remains unchanged in time). As explicated by Treeby
et al. (2010), higher frequencies travel to the receiver faster than
lower frequencies in the forward propagation; when waves back-
propagate in reversed time, higher frequencies will again need to
travel faster than the lower frequencies to arrive simultaneously at
the original source. As a result, using this wave equation implicitly
compensates for dispersion in wavefields. We will see this in the
(3) first numerical experiment.
In practice, the higher frequencies in recorded data are invariably
contaminated with noise. Attenuation compensation during TRM
might amplify such unwanted frequency content. To prevent high-
(4) frequency noise from growing exponentially, I apply a low-pass filter
to the attenuation and dispersion operators in eq. (8) in the spatial
Downloaded from http://gji.oxfordjournals.org/ at University of Texas at Austin on December 29, 2014
frequency domain when calculating the time-reversed wavefields.
The cut-off wavenumber is calculated by the cut-off frequency over
the maximum velocity of media.
To filter out the high-frequency noise caused by attenuation com-
pensation, I rewrite eq. (8) by substituting t̃ = −t
(5)  
1 ∂ 2 pB ∂
= ∇ 2 p B + ηL p B − ∇ 2 p B − τ H p B , (9)
c0 ∂ t̃
2 2 ∂ t̃
 
where ∇ 2 p B denotes non-attenuated dispersion, ηL p B − ∇ 2 p B
denotes attenuation associated dispersion and τ H ∂∂t̃ p B is the atten-
uation operator. By this means, I avoid filtering the first Laplacian
operator ∇ 2 p B , which is independent on attenuation. Indeed, eq. (9)
is an adjoint constant-Q viscoacoustic wave equation for TRM. If
we substitute −t = t̃ into eq. (9), it becomes eq. (3).
Below, I show the effect of the filter on attenuation and phase
dispersion. A Tukey window shaped filter was illustrated for high-
frequency noise control. The attenuation and dispersion values are
calculated using signals recorded at 20 and 100 m from the source
(Treeby & Cox 2010; Zhu & Harris 2014). The medium parameters
are given in the first example in the following section. I applied two
different Tukey filters in eq. (9). The cut-off frequency was set to
250 Hz, and the taper ratios were 0.5 and 0.1, respectively. Figs 2(a)
and (b) show that the calculated attenuation and dispersion curves
agree with theoretical curves of Kjartansson’s constant-Q model
(6)
damped as desired.
Numerical implementation
(7)
To incorporate a perfectly matched layer (PML) absorbing boundary
condition in the numerical simulation, I chose to use a coupled first-
order constitutive equation instead of a second-order equation (eqs
3 and 9). Eq. (9) is thus rewritten as
∂υ 1
= − ∇ p, (10)
∂ t̃ ρ0
∂ρ
= −ρ0 ∇ · υ, (11)
∂ t̃
(8)
 
p = c02 1 + ηL̃ − 1 − τ H̃ ρ, (12)
∂ t̃
where
L̃ = (−∇ 2 )γ , H̃ = (−∇ 2 )γ −1/2 , (13)
486 T. Zhu
Figure 2. The filtered attenuation (a) and dispersion (b) curves compared
with the unfiltered theoretical curves of the Kjartansson constant-Q model
(solid line). The star and circle symbols represent filtered values using a
Tukey window with taper ratios of 0.5 and 0.1, respectively.
and υ is the particle velocity, p is the spatial pressure field at time
t̃, c0 is the acoustic velocity, ρ and ρ0 are the acoustic and ambient
density, respectively.
I calculated the first-order spatial derivatives by the staggered-
grid pseudospectral method. The temporal derivative is solved with
the staggered-grid finite-difference approach. The fractional Lapla-
cian operators are implemented with the fractional Fourier pseu-
dospectral method as shown by Carcione (2010). Wavefield simu-
lations using the first-order conservation equations (eqs 10–12) can
be found in Zhu & Harris (2014).
TRM workflow
My complete procedure for implementing TRM in an attenuating
medium consists of three steps:
(1) Reverse recorded seismic data in time pF (xr , yr , zr , T − t)
and then enforce that data as a boundary condition at those original
receiver positions, mathematically expressed by p B (xr , yr , zr , t̃) =
pF (xr , yr , zr , T − t), where xr , yr , zr are the coordinates of re-
ceivers.
(2) Solve eqs (10)–(12) with the boundary condition for propa-
gating waves in reversed time. Attenuation and dispersion effects are
automatically compensated as time progresses. The source–receiver
geometry is usually known in the data. The medium parameters (ve-
locity, density and attenuation) are assumed to be known prior to
the TRM process.
(3) Finally, search for the maximum value of the pressure field
p B (x, y, z, t̃) at each time step and store the spatial position. The
location of the maximum pressure amplitude throughout entire time
of TRM is the source position (x0 , y0 , z 0 ) and its time t0 is the
excitation time of the source (Gajewski & Tessmer 2005; Saenger
2011), where t̃ runs from 0 to T and t0 = T − t̃. The pressure field
p B (x, y, z, T − t0 ) is shown as the time-reverse imaging.
TRM IN A 1-D HOMOGENOUS MODEL
To demonstrate the effect of attenuation compensation on the recon-
structed images in the spatial domain using TRM, I implemented
the simulation and reconstruction of a spatial source in a 1-D atten-
uating medium. I chose Pierre Shale rock to provide the viscoelastic
Downloaded from http://gji.oxfordjournals.org/ at University of Texas at Austin on December 29, 2014
properties of a homogeneous attenuating medium with quality fac-
tor Q = 32 and set the P-wave velocity to 2164 m s–1 at the reference
frequency 100 Hz (Carcione 2009). I posited a typical shale den-
sity of 2.2 g cm–3 . The initial condition pF (xs , t = 0) was a Ricker
wavelet S(xs ) in space, shown in Fig. 3(d). The two receivers are
located as shown by the dashed line in Fig. 3(a). The forward simu-
lation was performed using 512 gridpoints spaced 0.5 m apart. The
acoustic pressure time-series are recorded at receivers from times
t = 0 to t = T. In this example, T = 1.0 s. The cut-off frequency for
a cosine low-pass filter was 200 Hz.
For reverse-time propagation, the simulated traces were first
flipped in time and used as boundary conditions at the receiver
positions. The reversed time t̃ = T − t runs from 0 to T, where t
is the time of the forward propagation. For example, the snapshot
at reversed time t̃ = 0.8 s is equivalent to that at the forward time
t = 0.2 s. Thus, the snapshots of the forward propagation are used
as references. To illustrate the effectiveness of the attenuation com-
pensation, I ran two scenarios: time-reversal imaging without and
with attenuation compensation. As expected, the amplitude of the
signal without attenuation compensation (black line) at t̃ = 0.8 s in
Fig. 3(a) is smaller than that in the forward propagation at t = 0.2
s (red line). I then scanned for the maximum pressure amplitude to
give the location and the origin time of the reconstructed source.
The final comparison of the reconstructed source and the initial
source is shown in Fig. 3(c). The reconstructed source image gives
the approximate location but the wrong focusing time (t̃ = 0.98 s).
In contrast, with attenuation compensation, amplitudes were ampli-
fied during wave propagation as shown in Fig. 3(b). Fig. 3(d) shows
the reconstructed source that gives the approximate location as well
as the correct focusing time (t̃ = 1.0 s).
I deliberately reversed the sign of the dispersion operator and
reran TRM as above. The results are shown in Fig. 4. Clearly,
Fig. 4(b) shows the source peak detected at t̃ = 0.985 s and not
t̃ = 1.0 s, the original source excitation time. It is evident that
reversing the sign of the dispersion operator in eq. (9) results in
overcompensated dispersion (arriving earlier).
A P P L I C AT I O N S
Two applications, including 2-D time-reversal imaging and RTM in
attenuating media, are discussed in this section.
Time-reversal imaging in a 2-D model
In the first example, I applied my time-reversal imaging approach
to characterize a point source in a 2-D model. The conventional
 Time-reverse modelling in attenuating media 487

Downloaded from http://gji.oxfordjournals.org/ at University of Texas at Austin on December 29, 2014

Figure 3. Time-reversal snapshots at 0.8 s (top panels) and at the final focused time (bottom panels) in the same medium as forward modelling. The black lines
represent the reconstructed wavefields without attenuation compensation (a and c) and with attenuation compensation (b and d). The red lines are reference
signals at forward modelling times t = 0 s and t = 0.2 s, respectively, in a 1-D homogeneous medium. The two vertical dashed lines indicate the receiver
locations. Here, the reconstructed source with maximum amplitude appears too early at t̃ = 0.98 s in (c). With attenuation compensation, not only is the exact
source excitation time recovered but also the magnitude of the reconstructed source (black line) almost perfectly matches the initial source (red line).

time-reversal imaging technique has previously been used to suc-
cessfully locate seismic sources (e.g. McMechan 1982; Gajewski
& Tessmer 2005; Steiner et al. 2008). Here, I further investigate
whether I can improve the characterization of the seismic source
(e.g. the spatiotemporal history of the source) with my method. For
clear illustration, I use only acoustic data.
Let us consider a cross-well geometry for microseismic moni-
toring, which has been employed in practice to monitor hydraulic
fracturing (Warpinski 2009; Song & Toksöz 2011). The velocity
and Q models are shown in Fig. 5. The injection well is on the left-
hand side, and the monitoring well is on the right-hand side. The
lowest velocity around depth 100 m models the effects of fluid in-
jection. The model is discretized with 841 × 641 gridpoints. I use
a time step of 11 μs. The grid spacing of the horizontal and vertical
axes are x = z = 0.2 m. A time-varying Ricker wavelet with
centre frequency of 1000 Hz is injected as a point source at position
(x S = 11.8 m,z S = 90 m). 40 receivers are deployed at depths rang-
ing from 12 to 168 m with a spacing of 4 m. The source–receiver
geometry is plotted in Fig. 5(a). Synthetic data were generated by
a forward modelling viscoacoustic code, with Q based on the stan-
dard linear solid model (Zhu et al. 2013). The resulting acoustic
and viscoacoustic seismograms are shown in Fig. 6.
The focused energy is used to determine the location and time
where the source was excited, thus providing information about the
original source. Fig. 7 shows the time-reversal imaging results of
the viscoacoustic data without (c) and with (e) attenuation compen-
sation. For comparison, I also reconstructed the source location and
magnitude using the acoustic time-reversal imaging of the acoustic
data with the results appearing in Figs 7(a) and (b). It is notewor-
thy that, without attenuation compensation, the magnitude of the
source is underestimated. Picking the maximum pressure ampli-
tude yields an estimated source location at (7.6 m, 89 m), which is
shifted about one dominant wavelength from the true source location
(11.8 m, 90 m). The corresponding excitation time of the estimated
source is −0.56 ms (the negative sign indicating that the estimated
source excitation time is late compared to the original time). When
attenuation compensation is included into TRM, the estimated
source location is found at (12.0 m, 90.2 m). The sharpness and mag-
nitude of the reconstructed source is also notably improved. These
observations can also be easily identified from the cross-sections
through estimated x S and z S in Figs 8(a) and (b), respectively. The
spatial resolution has been clearly enhanced. Meanwhile, I show
comparisons of the recorded time-series at the source without and
with compensating attenuation in Fig. 8(c). Observe that the tem-
poral history of the source wavelet with attenuation compensation
(blue line) approximates the reference one (red line) fairly well. The
estimation without attenuation compensation is further from the true
location, and therefore it fails to provide source time information.
In order to examine how this modelling approach performs with
noisy data, I added Gaussian noise to the two data sets. The Gaussian
noise is generated by creating the signal-to-noise ratio (S/N = 1 dB)
on a logarithmic decibel scale (the viscoacoustic data are used as
the reference signal). Fig. 9 shows noisy acoustic and viscoacous-
tic data sets. The seismic events cannot be identified in the noisy
viscoacoustic seismogram. Fig. 10 shows the time-reversal imaging
results for the noisy viscoacoustic data set without (c) and with
(e) attenuation compensation. Again, for comparison I obtained the
reference image by backpropagating noisy acoustic data using the
488 T. Zhu
Figure 4. Time-reversal experiments conducted at 0.8 s (top panel), 0.985
(middle panel) and 1.0 s (bottom panel) in the same medium as forward
modelling. In addition to attenuation compensation, I reversed the sign of
the dispersion operator. The source focusing occurs at t̃ = 0.985 s with
amplitude being somewhat low.
acoustic time-reversal imaging technique. The results are shown in
Figs 10(a) and (b). In the non-compensated case in Fig. 10(c), the
scan for maximum amplitude yields the estimated source location
at (6.6 m, 88.8 m). The excitation time of the estimated source is
−0.8 ms. It is unsurprising that this result is worse than the noise-
free case. This is because the scan for maximum amplitude is ham-
pered by the strong noise. In contrast, in the compensated case, the
estimated source location is found at (11.6 m, 90.0 m). The excita-
tion time of the estimated source is correctly 0 ms. The focal area
is clearly visible despite a higher noise level in the image than in
the previous examples. These observations can also be easily iden-
tified from the cross-sections at position x S and z S in Figs 11(a) and
(b), respectively. Again, I found that the temporal history of source
wavelet with attenuation compensation (blue line) is quite close
to the reference one (red line) shown in Fig. 11(c). In summary,
time-reversal imaging with attenuation compensation enhances the
spatiotemporal resolution of the original source even though the
data are contaminated with noise.
Figure 5. Seismic models—(a) P-wave velocity and (b) Q. A point source
is located at xs = 11.8 m,z s = 90 m, shown as a black star. The inverted
triangles represent 40 receivers. The receiver depths range from 12 to 168 m
with a spacing of 4 m.
Downloaded from http://gji.oxfordjournals.org/ at University of Texas at Austin on December 29, 2014
Figure 6. Recorded seismograms at the receiver well, as computed by (a)
the acoustic solver and (b) the viscoacoustic solver.
Reverse-time migration
The second example illustrates the feasibility of TRM with attenu-
ation compensation in an RTM algorithm.
Typically, acoustic RTM includes a three-step procedure of (i)
forward propagation of a wavefield from a source through an ap-
propriate velocity model, (ii) backpropagation of the measured data
in reversed time through the same model and (iii) superposition
of both propagations, using an imaging condition. No attenuation
is assumed in this implementation. Mittet et al. (1995) presented
a complete procedure to compensate for attenuation in both the
source and the receiver wavefields in an attenuating medium. Here,
rather than compensating for both the source and receiver wave-
fields simultaneously, I apply attenuation compensation for the re-
ceiver wavefields in step (ii) and the attenuation loss for the source
 Time-reverse modelling in attenuating media 489

Downloaded from http://gji.oxfordjournals.org/ at University of Texas at Austin on December 29, 2014

Figure 7. Time-reverse source imaging using acoustic and viscoacoustic data without noise. The imaging is the maximum amplitude of the pressure field
throughout entire time modelling. (a) Reference image and (b) zoom over the source area. (c) Image without attenuation compensation and (d) zoom over the
source area. (e) Image with attenuation compensation and (f) zoom over source area. The colour scale indicates the magnitude of the source. The black star
represents the true point source location (xs = 11.8 m, z s = 90 m). Without attenuation compensation, the estimated source location is shifted. The inverted
triangles represent receivers.
490 T. Zhu
Figure 8. Comparisons among the reconstructed sources by the proposed
(blue) and conventional time-reversal technique (black) and the reference
source (red). (a) Horizontal cross-line at depth z s = 90 m. (b) Vertical cross-
line at distance xs = 11.8 m. (c) Time-series recorded at source location
(xs = 11.8 m, z s = 90 m).
wavefields in step (i) (Deng & McMechan 2007). The last procedure
to produce a RTM image is to apply a suitable imaging condition
to render the reflectors of the medium. I used a source normalized
cross-correlation of the forward-propagated source wavefield and
backward-propagated receiver wavefield (Claerbout 1971) given by
the following formula:
Downloaded from http://gji.oxfordjournals.org/ at University of Texas at Austin on December 29, 2014
Figure 9. Noisy (a) acoustic and (b) viscoacoustic data. For these, I added
the same Gaussian noise to the two data sets of Fig. 6. The Gaussian noise is
generated by creating the signal-to-noise ratio (S/N = 1 dB) on a logarithmic
decibel scale (the viscoacoustic data are used as the reference signal).
at 150.5 m depth. Apparently, the reflections are attenuated in the
viscoacoustic data. Before backpropagation, the direct wave was
muted. We used smoothed velocity and Q models for RTM input.
For RTM with and without attenuation compensation, the source
and receiver wavefields were saved only at every fourth time step to
reduce the volume of data storage. The Tukey filter was chosen with
a cut-off frequency of 1200 Hz and a taper ratio of 0.5 to suppress
the noise growth during attenuation compensation. After applying
the normalized cross-correlation imaging condition to the source
T
PF (x, t)PB (x, T − t)dt and receiver wavefields for all 101 shots, I then stacked them and
I (x) =
0
. (14) filtered the stacked RTM image to remove low wavenumber noise.
T The reference migration image in Fig. 13(b) shows acoustic data
PF (x, t)PF (x, t)dt
0
As a result, in the imaging position, the image consists of cor-
related wavefields, which represents the reflectivity of model. The
attenuation compensation tends to recover the amplitude loss and
approximate reflectivity energy as in the completely non-attenuating
case.
The reflectivity of the cross-well test model is shown in Fig. 13(a).
The corresponding velocity and Q values of each zone are listed in
Table 1. The constant density is 2.2 g cm–3 . The most highly at-
tenuating zone is zone 4, with Q = 30. The model is discretized
in a 201 × 601 grid with a grid spacing of x = z = 0.5m.
Receivers are in the left-hand well. I deployed 101 sources with
depths from 5 to 300 m in the right-hand well, located at x =100 m.
The source is a Ricker wavelet with a centre frequency of 400 Hz.
I implemented the forward modelling approach (eq. 3) with a PML
boundary of 20 gridpoints. Fig. 12 shows the acoustic and viscoa-
coustic common-shot-gather data, respectively. The shot is located
migrated with acoustic RTM. Fig. 13(c) displays the RTM output
without attenuation compensation. The image lacks illumination at
the boundaries of the highly attenuative zone 4 and at the interfaces
below. This is because the diffracted or reflected waves travelling
through this zone will be attenuated to such a degree that they are
hardly identifiable in the recorded data. When attenuation compen-
sation is included, the reflectors in target zones 4 and 5 appear to be
better illuminated as shown in Fig. 13(d).
D I S C U S S I O N A N D C O N C LU S I O N S
Whereas the traditional TRM approach typically neglects attenua-
tion effects on wavefields, I have presented a new TRM approach
in attenuating media to compensate for those effects of wavefields.
This TRM approach is based on the adjoint viscoacoustic wave
equation. The most attractive feature of this equation is the explicit
separation of amplitude attenuation and phase dispersion. Thus,
 Time-reverse modelling in attenuating media 491

Downloaded from http://gji.oxfordjournals.org/ at University of Texas at Austin on December 29, 2014

Figure 10. Time-reversal imaging using the noisy viscoacoustic data in Fig. 9(b). The imaging is the maximum amplitude of the pressure field throughout
entire time modelling. (a) Reference image and (b) the magnified source area. (c) Image without attenuation compensation and (d) the magnified source area.
(e) Image with attenuation compensation and (f) the magnified source area. The colour scale indicates the magnitude of the source. Black stars represent the
point source location. The estimated source location is shifted when attenuation has not been compensated.
492 T. Zhu

Downloaded from http://gji.oxfordjournals.org/ at University of Texas at Austin on December 29, 2014

Figure 12. Common shot gather at 150.5 m depth. (a) Acoustic data.
(b) Viscoacoustic data.

Figure 11. Comparisons among the reconstructed sources by the proposed propagation. As a result, this approach becomes stable as illustrated
(blue) and conventional time-reversal technique (black) and the reference in various numerical experiments.
source (red). (a) Horizontal cross-line at depth z s = 90 m. (b) Vertical cross- In the time-reversal imaging experiments with attenuation com-
line at distance xs = 11.8 m. (c) Time-series recorded at source location pensation, we found the time-reversed field focused in the vicinity
(xs = 11.8 m, z s = 90 m). of the original point source location and at the original focal time.
The magnitude of the reconstructed source is amplified and approx-
imates the reference source well, even using noisy data. The method
Table 1. P-wave velocity, is, therefore, especially attractive for weak onsets which are not de-
density and quality fac-
tectable on single traces of the network, such as microseismic data
tors of cross-well model in
Fig. 13(a).
sets. In practice, however, it may be challenging to apply attenuation
compensation when the noise frequency band overlaps the signal
Zone Cp (km s–1 ) Q frequency band. In this case, denoising processing must be applied
1 2.6 50 before running a TRM with attenuation compensation.
2 2.7 100 This approach could also be quite valuable for pre-stack RTM to
3 3.1 150 improve the resolution of images, particularly beneath very strong
4 2.6 30 attenuation areas. The reflectors in the conventional RTM might be
5 2.8 50 dimmed due to high-attenuation geological environments but may
6 3.5 300 be clearly imaged in the attenuation compensated RTM.
The primary contribution of this work is to present a novel TRM
approach in attenuating media. I have demonstrated the potentials of
attenuation compensation can be done by simply reversing the sign this TRM approach for the time-reversal imaging of seismic sources
of the amplitude attenuation operator and leaving the sign of the and RTM image in attenuating media. Experiments with these syn-
phase dispersion operator unchanged. This is superior to using other thetic data clearly demonstrate the ability of the TRM method to
wave equations for attenuation compensation. For example, revers- improve four aspects: (i) the estimation of the source location, (ii)
ing the sign of the attenuation operator will not correctly compensate the excitation timing of the point source, (iii) the magnitude of
for dispersion because the amplitude attenuation and phase disper- focused source wavelet and (iv) the reflectivity image of structure
sion are encapsulated by a single operator (Deng & McMechan beneath high-attenuation zones.
2007, 2008; Carcione 2010). Another equation proposed by Zhang Future work should focus on two aspects. First, we can extend
et al. (2010) introduces the normalization exponential operator for the extension to the 3-D viscoacoustic TR modelling and the vis-
attenuation compensation, which might compensate for amplitude coelastic TR modelling based on the viscoelastic wave equation
attenuation and phase dispersion but is more or less ad hoc. Com- (Zhu & Carcione 2013). In addition, we need to test this approach
pensating for attenuation is further stabilized by applying a low-pass with real data for locating (micro)earthquake sources and imaging
filter, which avoids amplifying the high-frequency signals during the subsurface reflectors by RTM.

Figure 13. (a) Reflectivity model with block numbers referenced in Table 1. For illustration, the inverted triangles denote receivers in the left well. The black
stars represent sources in the right well; (b) acoustic RTM of acoustic data; (c) acoustic RTM of viscoacoustic data; (d) Q-RTM with attenuation compensation
of viscoacoustic data.
AC K N OW L E D G E M E N T S
Fruitful discussions with Prof Jerry Harris on this work are highly
acknowledged. I would like to express my gratitude to Dr Stew-
art A. Levin for thorough comments and suggestions on the first
draft, and Dr Phil Hubbard for mostly correcting language errors
in the manuscript. I also thank the editor Randy Keller, and two re-
viewers, Erik Saenger and an anonymous reviewer for constructive
comments. This work was funded by the Stanford Wave Physics
laboratory.
REFERENCES
Aki, K. & Richards, P.G., 1980. Quantitative Seismology—Vol. 1: Theory
and Methods (2nd edn 2002), W. H. Freeman.
Ammari, H., Bretin, E., Garnier, J. & Wahab, A., 2011. Time reversal in
attenuating acoustic media, Contemp. Math., 548, 151–163.
Baysal, E., Kosloff, D.D. & Sherwood, J.W.C., 1983. Reverse-time migra-
tion, Geophysics, 48, 1514–1524.
Carcione, J.M., 2009. Theory and modeling of constant-Q P- and S-waves
using fractional time derivatives, Geophysics, 74, T1–T11.
Carcione, J.M., 2010. A generalization of the Fourier pseudospectral method,
Geophysics, 75, A53–56.
Claerbout, J.F., 1971. Toward a uniffied theory of reffiector mapping,
Geophysics, 36, 467–481.
Deng, F. & McMechan, G.A., 2007. True-amplitude prestack depth migra-
tion, Geophysics, 72, S155–S166.
Deng, F. & McMechan, G.A., 2008. Viscoelastic true-amplitude prestack
reverse-time depth migration, Geophysics, 73, S143–S155.
Fink, M., 2006. Time-reversal acoustics in complex environments,
Geophysics, 71(4), SI151–SI164.
Gajewski, D. & Tessmer, E., 2005. Reverse modelling for seismic event
characterization, Geophys. J. Int., 163, 276–284.
Gosselet, A. & Singh, S.C., 2007. Using symmetry breaking in time-reversal
mirror for attenuation determination, in Proceedings of the SEG Expanded
Abstracts, 26, pp. 1639–1643.
Time-reverse modelling in attenuating media 493
Downloaded from http://gji.oxfordjournals.org/ at University of Texas at Austin on December 29, 2014
Hu, L.Z. & McMechan, G.A., 1988. Elastic finite difference modelling and
imaging for earthqake sources, Geophys. J. Int., 95, 303–313.
Kao, H. & Shan, S.-J., 2004. The source-scanning algorithm: mapping the
distribution of seismic sources in time and space, Geophys. J. Int., 157,
589–594.
Kawakatsu, H & Montagner, J-P., 2008. Time-reversal seismic-source imag-
ing and moment-tensor inversion, Geophys. J. Int., 175, 686–688.
Kjartansson, E., 1979. Constant-Q wave propagation and attenuation,
J. geophys. Res., 84, 4737–4748.
Kremers, S., Fichtner, A., Brietzke, G., Igel, H., Larmat, C., Huang, L.
& Käser, M., 2011. Exploring the potentials and limitations of the
time-reversal imaging of finite seismic sources, Solid Earth, 2, 95–
105.
Larmat, C., Montagner, J-P., Fink, M., Capdeville, Y., Tourin, A. & Clévé
dé, E., 2006. Time-reversal imaging of seismic sources and applica-
tion to the great Sumatra earthquake, Geophys. Res. Lett., 33, L19312,
doi:10.1029/2006GL026336.
Larmat, C.S., Guyer, R.A. & Johnson, P.A., 2009. Tremor source location
using time reversal: selecting the appropriate imaging field, Geophys. Res.
Lett., 36, L22304, doi:10.1029/2009GL040099.
Lokmer, I., O’Brien, G.S., Stich, D. & Bean, C.J., 2009. Time reversal
imaging of synthetic volcanic tremor sources, Geophys. Res. Lett., 36,
L12308, doi:10.1029/2009GL038178.
Lu, R., 2007. Time reversed acoustics and applications to earthquake loca-
tion and salt dome flank imaging, PhD thesis, Massachusetts Institute of
Technology.
Lu, R., Toksöz, M.N. & Willis, M.E., 2008. Locating microseismic events
with time reversed acoustics: a synthetic case study, in Proceedings of the
78th Annual International Meeting, SEG, Expanded Abstracts, pp. 1342–
1346.
McDonal, F.J., Angona, F.A., Mills, R.L., Sengbush, R.L., Van Nostrand,
R.G. & White, J.E., 1958. Attenuation of shear and compressional waves
in Pierre shale, Geophysics, 23, 421–439.
McMechan, G., Luetgert, J. & Mooney, W., 1985. Imaging of earthquake
source in long valley caldera, California 1983, Bull. seism. Soc. Am.,
75(4), 1005–1020.
McMechan, G.A., 1982. Determination of source parameters by wavefield
extrapolation, Geophys. J. R. astr. Soc., 71, 613–628.
494 T. Zhu
McMechan, G.A., 1983. Migration by extrapolation of time-dependent
boundary values, Geophys. Prospect., 31, 413–420.
Mittet, R., Sollie, R. & Hokstad, K., 1995. Prestack depth migration
with compensation for absorption and dispersion, Geophysics, 60, 1485,
doi:10.1190/1.1443882.
O’Brien, G., Lokmer, I., Barros, L.D., Bean, C., Saccorotti, G., Metaxian,
J.-P. & Patane, D., 2011. Time reverse location of seismic long-period
events recorded on Mt. Etna, Geophys. J. Int., 184, 452–462.
Saenger, E. H., 2011. Time reverse characterization of sources in heteroge-
neous media, NDT & E Int., 44(8), 751–759.
Snieder, R. K., 2007. Extracting the Green’s function of attenuating acoustic
media from uncorrelated waves, J. acoust. Soc. Am., 121, 2637–2643.
Song, F. & Toksöz, M.N., 2011. Full-waveform based complete moment
tensor inversion and source parameter estimation from downhole micro-
seismic data for hydrofracture monitoring, Geophysics, 76(6), WC103–
WC116.
Steiner, B., Saenger, E.H. & Schmalholz, S.M., 2008. Time re-
verse modelling of low-frequency microtremors: application to hy-
drocarbon reservoir localization, Geophys. Res. Lett., 35, L03307,
doi:10.1029/2007GL032097.
Treeby, B.E. & Cox, B.T., 2010. Modeling power law absorption and dis-
persion for acoustic propagation using the fractional Laplacian, J. acoust.
Soc. Am., 127, 2741–2748.
Treeby, B.E., Zhang, E.Z. & Cox, B.T., 2010. Photoacoustic tomography in
absorbing acoustic media using time reversal, Inverse Probl., 26, 115003,
doi:10.1088/0266-5611/26/11/115003.
Warpinski, N.R., 2009. Microseismic monitoring: inside and out, J. Petrol.
Tech., 61, 80–85.
Zhang, Y., Zhang, P. & Zhang, H., 2010. Compensating for visco-acoustic
effects in reverse-time migration, in Proceedings of the 80th Annual In-
ternational Meeting, SEG Expanded Abstracts, 29, pp. 3160–3164.
Zhu, T., Carcione, J.M. & Harris, J.M., 2013. Approximating constant-Q
seismic propagation in the time domain, Geophys. Prospect., 61, 931–
940.
Zhu, T. & Carcione, J.M., 2013. Theory and modelling of constant-Q
P- and S-waves using fractional spatial derivatives, Geophys. J. Int.,
doi:10.1093/gji/ggt483.
Zhu, T. & Harris, J.M., 2014. Modeling acoustic wave propagation in
Downloaded from http://gji.oxfordjournals.org/ at University of Texas at Austin on December 29, 2014
heterogeneous attenuating media using decoupled fractional Laplacians,
Geophysics, in press.