Radon transform and multiple attenuation

Radon transform and multiple attenuation

Zhihong (Nancy) Cao, John C. Bancroft, R. James Brown, and Chunyan (Mary)
Xaio

ABSTRACT
Removing reverberations or multiples from reflection seismograms has been a
longstanding problem of exploration geophysics. Multiple reflections often destructively
interfere with the primary making interpretation difficult. The Radon transform, which
integrates some physical property along a particular path, will be evaluated for
attenuating multiple.

Multiples are periodic in the τ-p domain, and predictive deconvolution can be applied
in the τ-p domain to suppress multiples. After NMO correction, moveout errors are
approximately parabolic and tend to map to points in the parabolic Radon domain. The
hyperbolic Radon transform will also map data before and after moveout correction into
points, and multiples can be recognized in the Radon domain. The identified multiple
energy can then subtracted from the data to improve the interpretation process.

INTRODUCTION
Johan Radon (1917) is credited with establishing the Radon transform, a function that
integrates some physical property of a medium along a particular path.

Removing reverberations from reflection seismograms has been a long standing

problem of exploration geophysics. Multiple reflections often destructively interfere with
the primary reflections of interest. The most robust and effective way to suppress
multiples is stacking normal moveout (NMO) corrected seismic gathers (Foster and
Mosher, 1992). Unfortunately, stacking does not eliminate all multiples.

Over the years, many techniques for suppressing multiples have been tried. More
recently, the Radon transform approaches have attracted attention. The generalized Radon
transform integrates the data along any curved surfaces (Chapman, 1981). Particularly,
the slant-stack (or τ -p ) transform integrates the data along planar surfaces (Treitel, et al,
1982). Hampson (1986) applied NMO-correction to common midpoint (CMP) data and
perform a Radon transform along parabolic stacking curve to suppress multiples in
NMO-corrected domain. Yilmaz (1989) applied t2-stretching to the CMP data and then
applied the parabolic Radon transform in the stretched coordinates. Foster and Mosher
(1992) described a hyperbolic Radon transform in stacking NMO-corrected domain.
Oppert (2002) proposed the non-hyperbolic Radon transform by taking the shifted
hyperbola NMO equations. Thorson and Claerbout (1985), and Beylkin (1987) showed
the least-squares solution of the discrete Radon transform. The former also gave the

CREWES Research Report -Volume 15 (2003) 1

Cao et al.

stochastic inverse solution for the problem. Zhou and Greenhalgh (1994) showed the
convolutional operators for the slant-stack transform and the parabolic Radon transform
to increase the resolution. Sacchi and Tadeusz (1995) proposed an improved algorithm
for the parabolic Radon transform to get higher resolution.

The generalized Radon transform u (q,τ ) is defined as

u ( q,τ ) = ∫ d ( x, t = τ + qφ ( x ) ) dx ，
∞
(1)
−∞

where d ( x, t ) is the original seismogram, x is a spatial variable such as offset, φ ( x) is

the curvature based on which the transform curve is defined, q is the slope of the
curvature, and τ is the intercept time (Foster and Mosher, 1992).

Since the seismogram is digitally recorded, a discrete form of equation (1) is:

u ( q,τ ) = ∑ d ( x, t = τ + qφ ( x ) )
x ， (2)
Then the inverse transforms of equation (1) and (2) are:

d ' ( x, t ) = ∫ u ( q,τ = t − qφ ( x ) ) dq ,
∞
(3)
−∞

or

d ' ( x, t ) = ∑ u ( q,τ = t − qφ ( x ) ) ， (4)

q

Also, we can express the Radon transform in velocity domain (Yilmaz, 1989) as follows:

u ( v,τ ) = ∑ d ( x, t = τ + qφ ( v, x ) ) ， (5)
x

d ' ( x, t ) = ∑ u ( v,τ = t − qφ ( v, x ) ) ， (6)

q

THE DISCRETE RADON TRANSFORM

Equation (3) can be written as:

d ' = Lu ， (7)

where L is the linear transformation from the τ -q space to the offset space defined by
equation (3), and given as follows in Fourier domain:

2 CREWES Research Report -Volume 15 (2003)

 Radon transform and multiple attenuation

− iω q jφ ( xk )  j = 1,...,M 
L jk = e  ， (8)
k = 1,...,N 

Assume that a CMP gather d(x,t) is the result of some transform on a function u0(q,τ)
in the τ-q space. In fact, d(x,t) is always contaminated with additive noise n giving:

d = Lu 0 + n ， (9)

The solution for u in equation (9) is obtained by taking a least-squares approach to

minimize the noise term n, which represents the difference between the actual data u0 and
the modelled data u. The cumulative squared noise term, S = n Tn = (d - Lu 0 )T (d - Lu 0 )

(Oppert and Brown, 2002), is minimized with respect to u0 yielding the desired
least-squares solution (Thorson and Claerbout, 1985):

[ ]
u = LT L LT d ，
−1
(10)

−1
The generalized inverse of L is thus computed to be  LT L  LT .

[
The calculation of LT L ]
−1
is impractical due to the large nature of the matrix and the

instability of the inversion. Although the operator LT L is diagonally dominant, if the

side lobes of the matrix are significant, smearing will occur along the q-axis.
Prewhitening the operator LT L suppresses the side lobes and stabilizes the inversion. A
stable solution for equation (9) is computed by Thorson and Claerbout (1985):

[
u = LT L + µI LT d ， ]
−1
(11)

where the constant µ is a damping factor incorporated to add white noise along the
main diagonal of the inversion matrix, and I is the identity matrix.

CREWES Research Report -Volume 15 (2003) 3

Cao et al.

THE LINEAR RADON TRANSFORM

Theory

∞
u ( p,τ ) = ∫ d ( x, t = τ + px )dx , (12)
−∞

u ( p, τ ) = ∑ d ( x, t = τ + px ) , (13)
x

∞
d ' ( x, t ) = ∫ u ( p,τ = t − px )dp , (14)
−∞

d ' ( x, t ) = ∑ u ( p, τ = t − px ) ， (15)
p

The equation (1) reduces to the linear Radon transform when φ ( x) = x , where x is
offset. Accordingly, equations (1) to (4) are reformatted as equations (12) to (15).

For seismic exploration applications, d(x,t) can be CMP or shot gather. u(p,τ) is its
slant-stack transform with horizontal slowness (or ray parameter) p and intercept time τ.
Here p is used because of its specific meaning of ray parameter and is defined by

sin θ ∆t
p= = , (16)
v ∆x
where θ is the incident angle from the vertical axis, v is the wave propagation velocity.

The rho filter has to be applied before inverse mapping to restore the correct
amplitude and phase. This work was illustrated by Zhou and Greenhalgh (1994).

The Fourier transform of equation (12) becomes

∞
U ( p, ω ) = ∫ D( x, ω )e iωpx dx ， (17)
−∞

Version 1:

To obtain the proper inverse transform D(x,ω) in equation (16), the standard
back-projection D’(x,ω) is used:

∫ ( )
D x ' , ω e iωp (x − x )dx ' dp
∞ ∞ ∞
D ' (x, ω ) = ∫ U ( p, ω )e −iωpx dp = ∫
'

−∞ −∞ −∞
， (18)
∞
(
= ∫ dx ' D x ' , ω
−∞
)∫
−∞
∞
e −iωp (x − x )dp
'

Now the rho filter is introduced:

4 CREWES Research Report -Volume 15 (2003)

 Radon transform and multiple attenuation

∞
ρ ( x, ω ) = ∫ e −iωpx dp ， (19)
−∞

Then equation (17) can be written as:

−∞
∞
( )( )
D ' (x, ω ) = ∫ D x ' , ω ρ x − x ' , ω dx ' = D(x, ω ) ∗ ρ ( x, ω ) ， (20)

Here, “∗” stands for convolution with respect to the spatial variable x. Zhou and
Greenhalgh (1994) derive the form of rho filter in case of infinite p:

2π
ρ ( x, ω ) = δ (x ) ， (21)
ω

Substituting equation (21) to equation (20), we have

2π
D ' ( x, ω ) = D ( x, ω )
ω

or

ω '
D ( x, ω ) = D ( x, ω ) ， (22)
2π

Therefore the slant-stack transform pair in the frequency domain is

U ( p, ω ) = ∞ D( x, ω )e iωpx dx
 ∫−∞
 ， (23)
ω ∞
 D ( x, ω ) = U (ω , p )e
2π ∫−∞
− iωpx
dp


But in practice, variable p has a limited range of [ pmin , pmax ]. In this case, the rho filter

becomes:

ρ ( x, ω ) = ∫
pmax
e −iωpx 
dp =  iωx
e(
 1 −iωpmin x
)
− e −iωpmax x , ωx ≠ 0
， (24)
pmin
 pmax − pmin , ωx = 0

When pmin = pmax , the equation (24) becomes a sinc function:

CREWES Research Report -Volume 15 (2003) 5

Cao et al.

  sin (ωpmax x ) 
2 pmax 
ρ ( x, ω ) =   , ωx ≠ 0 ，
 ωpmax x  (25)
2 p , ωx = 0
 max

Version 2:

Zhou and Greenhalgh (1994) indicate that we can perform the inverse τ-p transform
first and then the forward transform to derive the proper inversion of the slant-stack
transform. The inverse τ-p transform and its Fourier transform are defined as:

∞
d ( x, t ) = ∫ u ( p, t − px )dp ， (26)
−∞

and

∞
D( x, ω ) = ∫ U ( p, ω )e −iωpx dp ， (27)
−∞

Here d(x,t) is the input data and u(p,τ) is the τ-p transform. D(x,ω) and U(p,ω) are the
Fourier transforms of d(x,t) and u(p,τ), respectively. To obtain the forward τ-p transform
U(p,ω) in equation (26), U’(p,ω) is defined as:

∞
U ' ( p, ω ) = ∫ D( x, ω )e iωpx dx
−∞
， (28)
∞
= ∫ dp U p , ω
−∞
'
( '
)∫−∞
∞
e (
iωx p − p ' )dx

The new function g(p,ω) is introduced:

∞
g ( p, ω ) = ∫ e iωpx dx ， (29)
−∞

Then equation (28) can be written as:

U ' ( p, ω ) = U ( p , ω ) ∗ g ( p, ω ) , (30)

In case that the spatial variable x is of infinite range, we have:

2π
g ( p, ω ) = δ ( p) , (31)
ω

The τ-p transform pair in the frequency domain can be written as

6 CREWES Research Report -Volume 15 (2003)

 Radon transform and multiple attenuation

 ω ∞
U ( p, ω ) = ∫ D(h, ω )e iωpx dx
 2π − ∞
， (32)
 D( x, ω ) = U ( p, ω )e
∞
∫−∞
− iω px
dp

In case that the spatial variable x is of finite range [xmin, xmax], we have g(p,ω):

 (
 1 iωxmax p iωxmin p
e −e )
, ωp ≠ 0
g ( p, ω ) =  iωp ， (33)
x − x , ωp = 0
 max min

If xmin=- xmax, g(p,ω) can be written as:

  sin (ωphmax ) 
2hmax 
g (h, ω ) =   , ωh ≠ 0
 ωphmax  ， (34)
2 h , ωh = 0
 max

Comparing the two versions of the linear τ-p transforms, i.e. equation (23) and
equation (32), the only difference is that the deconvolution in version 1 is performed on
the inverse transform in x-direction, and that the deconvolution in version 2 is performed
on the forward transform in p-direction. Deconvolution can improve resolution in the
x-direction for version 1 and in p-direction for version 2. In order to attenuate noise in the
τ-p domain, the transform pair of version 2 is preferred (Zhou and Greenhalgh, 1994).

Slant-stack multiple attenuation

For a one-layer model with velocity v, the traveltime equation in offset domain is:

x2
t =t + 2 ，
2 2
0 (35)
v

where x is the source/receiver offset and t0 is the two-way zero-offset time.

CREWES Research Report -Volume 15 (2003) 7

Cao et al.

0 0

B
0.5
A 0.5

T
A Parabola

Hyperbola
Tau

1
B 1

Straight Line

1.5 1.5
-2000 -1500 -1000 -500 0 500 1000 1500 2000 -4 -3 -2 -1 0 1 2 3 4
-4
Half Offset Ray Parameter p x 10

(a) (b)

FIG. 1. A hyperbola in CMP gather (a) maps onto an ellipse in the slant-stack domain (b); and a
straight line in CMP gather maps to and a point in the slant-stack domain. Energy tangent to the
line B in (a) maps to the point B in (b).

In the slant-stack domain, equation (35) is transformed to (Alam and Austin, 1981):

τ 2 = (1 − p 2 v 2 )τ 02 ， (36)

Comparing equation (35) with (36) we see that the hyperbola in the (x, t) domain
becomes an ellipse in the (τ, p) domain. The linear events in the offset space map to
points in the (τ, p) space (Figure 1). But notice the ellipse in Figure 1b that the ellipse is
incomplete because the offset, in practice, couldn’t be infinite. So when the data maps
from τ-p domain back to the offset domain, amplitude smearing will occur.

Now consider the traveltime for an n-bounce multiple in (τ, p) domain and time delay
between successive bounces for a given trance (Alam and Austin, 1981):

τ 2 = (1 − p 2 v 2 )n 2τ 02 ， (37)

τ n − τ n−1 = (1 − p 2 v 2 )2 τ 0 ，
1
(38)

Multiples are not periodic in the offset domain. Taner (1980) first recognized that the
time separations between the arrivals are equal along a radial direction OR in the offset
space (Figure 2). Then Alam and Austin (1981) showed in equation (38) that all bounces
within a reverberatory layer have the same traveltime for a given p-value, i.e. the
multiples are exactly periodic in the (τ, p) space (Figure 2).

8 CREWES Research Report -Volume 15 (2003)

 Radon transform and multiple attenuation

FIG. 2. The periodicity of multiples along radial trace OR and down p traces from Taner (1980).

Based on the periodic property of the multiples in the (τ, p) domain, Alam and Austin
(1981) and Treitel et al. (1982) investigated the application of predictive deconvolution in
the slant-stack domain for multiple suppression. The shot gather in Figure 3a is
transformed to the slant-stack domain (Figure 3c). Figure 3e is the slant-stack gather after
predictive deconvolution. Figure 3b, d, and f are the autocorrelations of Figure 3a, c and
e, respectively. Figure 3g shows the reconstruction of the shot gather from the slant-stack
gather in Figure 3e. Unlike the autocorrelogram of the shot gather in Figure 3b, the
periodic nature of the multiples in the data is pronounced in the autocorrelogram of the
slant-stack gather (Figure 3d). Note that the periodicity of multiples changes from one p
trace to the next. The largest period occurs along the trace that corresponds to the
minimum p value. The autocorrelogram after predictive deconvolution shows that the
energy in lags less than the specified prediction lag is retained, while the multiple energy
is attenuated (Figure 3f) (Yilmaz, 1989).

CREWES Research Report -Volume 15 (2003) 9

Cao et al.

FIG. 3. Multiple attenuation in the slant-stack domain (a) A shot gather; (b) Autocorrelogram of (a);
(c) The slant-stack gather; (d) The autocorrelogram of (c); (e) The slant-stack gather after
predictive deconvolution, where operator length=240 ms; (f) The autocorrelogram of (e); (g)
Reconstruction of the shot gather from (e) given by Yilmaz (1987).

THE PARABOLIC RADON TRANSFORM

Theory

On CMP or common shotpoint CSP gathers, seismic events are more likely to be
hyperbolic than linear (reflections and diffractions are hyperbolic, but refractions and
direct waves are linear). A hyperbolic Radon transform can be designed so that the
hyperbolic events in the seismic domain map into points in the Radon domain. But the
direct hyperbolic τ-p transform was too expensive to realize and faster method were
sought. Errors in the moveout correction due to velocity approximations are exactly
hyperbolic, but their small amounts of curvature may be approximated by parabolas.
Consequently Hampson (1986) performed the parabolic Radon transform on the
NMO-corrected offset data. Yilmaz (1989) noted that all the hyperbolic events in the
offset domain are transformed to exact parabolas after t2-stretch is applied.

10 CREWES Research Report -Volume 15 (2003)

 Radon transform and multiple attenuation

NMO-corrected input data

A practical approach was presented by Hampson (1986). First, the input CMP gather is
NMO-corrected using the hyperbolic moveout equation

x2
tn = t − 2 ,
2
(39)
vn

where t is the recorded time, tn is the time after NMO-correction, and vn is the
NMO-correction velocity. The resulting moveout of the events, which were originally
hyperbolic, are now approximately parabolic:

x2
t 2 = t 02 + (40)
v2

x2
2
terr = t 02 + 2
(41)
verr

x2 x2
2
t err = t2 − + 2 , error term is exactly hyperbolic (42)
v 2 verr

2
t err x2  1 1
= 1 +  +  (43)
t2 t 2  verr
2
v 2 

terr x2  1 1
= 1 + 2  2 + 2  (44)
t t  verr v 

terr x2  1 1
≈1+ 2  2 + 2  (45)
t 2t  verr v 

x2  1 1
terr ≈ t +  2 + 2  , approximately parabolic (46)
2t  verr v 

t n = τ + qx 2 , (47)

where τ is the two-way zero-offset time, and q defines the curvature of the parabola.

CREWES Research Report -Volume 15 (2003) 11

Cao et al.

Say d(x,tn) is the NMO-corrected gather. The forward and inverse Radon transform in
the NMO-corrected coordinates take the forms

(
u (q, τ ) = ∑ d x, t n = τ + qx 2 , ) (48)
x

and

( )
d ' ( x, t ) = ∑ u q, τ = t n − qx 2 . (49)
q

Stretching input data

Events on the CMP gather have hyperbolic traveltimes defined by

t 2 = τ 2 + x2 v2 ， (50)

where t is the two-way traveltime, τ is the two-way zero-offset traveltime, x is the offset,
and v is the stacking velocity.

Apply stretching in the time direction by setting t ' = t 2 and τ ' = τ 2 . Equation (50)
takes the form

t ' = τ ' + x2 v2 ， (51)

From equation (51) we can see that the hyperbolae are transformed exactly to
parabolae in the stretched coordinates. Now say d(x,t’) is the t2-stretched input data. The
mapping from the x-t’ domain to the τ’-q space is achieved by summing over offset:

u (q,τ ' ) = ∫ d (x, t ' = τ ' + qx 2 )dx ，

∞

−∞

or

u (q,τ ' ) = ∑ d (x, t ' = τ ' + qx 2 ) ， (52)

x

And the inverse transforms are

d ' (x, t ' ) = ∫ u (q,τ ' = t ' − qx 2 )dq ，

∞

−∞

or

12 CREWES Research Report -Volume 15 (2003)

 Radon transform and multiple attenuation

( )
d ' x, t ' = ∑ u q,τ ' = t ' − qx 2 ，( ) (53)
q

From equation (51), we know that the physical meaning of q is taken as the square of
the horizontal slowness, or the inverse of the square of stacking velocity (Yilmaz, 1989).
But, in practice, equation (53) doesn’t give the exact inversion of the Radon transform.

To obtain superior image resolution in the transform domain, a similar approach as

was used with the linear Radon transform, version 2, is taken in the following derivation
of the parabolic τ-q transform. Now we take the Radon transform in the t2-stretching
domain as an example, i.e. equations (52) and (53).

Say D(x,ω’) and U(p,ω’) are the Fourier transforms of d(x,t’) and u(q,τ’), respectively.
D’(x,ω’) is the standard Fourier transform of the reverse parabolic Radon transform.
Then, in the frequency domain, we have (Zhou and Greenhalgh, 1994):

( ) ∞
D x, ω ' = ∫ U q, ω ' e −iω qx dq ，
−∞
( ) ' 2
(54)

The forward projection function U’(x,ω’) of the parabolic τ-q transform is:

( ) ∞
U ' q, ω ' = ∫ D q, ω ' e iω qx dx ，
−∞
( ) ' 2
(55)

Substituting equation (54) to (55), we can get:

( )
U ' q, ω ' = ∫
∞

−∞ −∞
( )
∫
∞
U q ' , ω ' e iω x (q − q )dq ' dx
' 2 ')

， (56)
= ∫ dq U (q , ω )∫ ( )dx
∞ ∞
' ' ' iω ' x 2 q − q ')
e
−∞ −∞

The function σ (q,ω’) defined by:

σ (q, ω ' ) = ∫ e iω qx dx ，
∞ ' 2
(57)
−∞

is introduced. Then equation (56) can be written as:

U ' (q, ω ' ) = U (q, ω ' ) ∗ σ (q, ω ' )， (58)

In case of infinite spatial input data, Zhou and Greenhalgh (1994) derived σ (q,ω’) as:

CREWES Research Report -Volume 15 (2003) 13

Cao et al.

π
σ (q, ω ' ) = [1 + isign(q )] ， (59)
2ω ' q

Equation (59) indicates that q-direction deconvolution is required even if the input data
have an infinite spatial extent. This is different from the linear Radon transform, version
2. In the Fourier transform domain, equation (58) becomes:

( ) (
U ' kq ,ω ' = U kq ,ω ' σ kq ,ω ' ， )( ) (60)

or

U ' (k q , ω ' ) ω ' U ' (k q , ω ' )

U (k q , ω ) =
'
= ， (61)
σ (k q , ω ' ) σ ' (k q )

Here the new function σ’(kq) is given by Zhou and Greenhalgh (1994):

π
σ ' (k q ) = ∫ [1 + isign(q )]
∞ −ik q q
e dq
−∞ 2q
， (62)
π
= [1 + isign(q )]
kq

In fact, the Fourier transform variable kq corresponding to variable q is positive.

Therefore, equation (62) can be written as:

2π
σ ' (k q ) = ， (63)
kq

and equation (61) becomes:

ωk p
(
U kq ,ω ' = ) 2π
(
U ' kq ,ω ' ， ) (64)

which shows that the q-direction deconvolution enhances the Fourier transform
components. Therefore the resolution in the parabolic τ-p transform domain is increased.

In case that the variable x has a limited range, [xmin, xmax], there is no analytical solution
for the equation (57), unless ω’=0, in which case σ (q,ω’)= xmin - xmax. This integral has to
be approximated by numerical quadrature methods such as the rectangular, trapezoidal or
Simpson rules.

14 CREWES Research Report -Volume 15 (2003)

 Radon transform and multiple attenuation

Velocity-stack

We can do the parabolic Radon transform in the velocity domain. Then equation (52)
and equation (53) can be written (Yilmaz, 1989) as:

( ) (
u v, τ ' = ∑ d x, t ' = τ ' + 4 x 2 v 2 ， ) (65)
x

( ) (
d ' x , t ' = ∑ u v, τ ' = t ' − 4 x 2 v 2 ， ) (66)
q

We have seen that equation (66) can’t give the exact inversion of equation (65). So a
rho filter is convoluted to u(v,τ’) before integration over velocity (Beylkin, 1987):

( ) ∞
(
d ' x, t ' = ∫ ρ (τ ) ∗ u v,τ ' = t ' − 4 x 2 v 2 dv ，
−∞
) (67)

where ρ (τ ) has a Fourier transform of the form ω exp(iπ 4) .

In this case, L, the linear transform [Equation (8)] from the velocity space to the offset
space is modified as:

− iωx 2j vk2  j = 1,...,M 

L jk = e  ， (68)
k = 1,...,N 

Multiple Suppression

Since the mapping function is summed along parabolic curve, a parabola in the
t -space domain, such as a primary or multiple, ideally maps onto a point in the τ-q or
2

Radon domain. Figure 4 (Yilmaz, 1989) shows primaries and multiples in the offset
domain and the velocity domain. Those multiples have same velocity with their related
primaries. Hence, we are able to distinguish multiples from primaries in the velocity
domain based on velocity discrimination and attenuate multiples.

Now we take the velocity-stack transform into practice (Yilmaz, 1989). Figure 5a is a
synthetic CMP gather with three primary reflections; 5b is a synthetic CMP gather with
one primary reflection (arrival time at 0.2 s at zero-offset trace) and its multiples; 5c
integrates 5a and 5b; 5d is the t2-stretching section of 5c. In Figure 6a is the velocity stack
of Figure 5d; 6b is the velocity-stack after undoing t2-streching, here we can separate
multiples and primaries based on velocity differences as indicated; 6c is the full CMP
gather reconstruction from 6b. Comparing Figure 6b with Figure 5d, Figure 6b is a
reasonably good reconstruction. On Figure 7c, only primaries are reconstructed from

CREWES Research Report -Volume 15 (2003) 15

Cao et al.

Figure 6b and there is some residual multiple energy. In practice, non-hyperbolic events,
such as direct and refracted wave, will be lost during the parabolic Radon transform. So
on Figure 7b, only multiples are constructed from Figure 6b. Then these multiples are
subtracted from the original data Figure 5c, and all of those events aside from multiples
are left on the output. On Figure 6b, we find some distortion of the wavelet in shallow
events (less than 1 s). This is caused by t2-stretching. This distortion will produce blurry
events in the offset domain, correspondingly, such as Figure 6c, 7a, 7b and 7c.

FIG. 4. Mapping from offset domain to velocity domain (Yilmaz, 1987)

16 CREWES Research Report -Volume 15 (2003)

 Radon transform and multiple attenuation

(a) (b) (c) (d)

FIG. 5. (a) A synthetic CMP gather with three primary reflections; (b) A synthetic CMP gather with
one primary reflection (arrival at 0.2 s at zero-offset) and its multiples; (c) Integration of (a) and (b);
(d) t2-stretching section of (c) from Yilmaz (1987).

(a) (b) (c)

FIG. 6. (a) The velocity-stack of the CMP gather in Figure 5(d); (b) The velocity-stack after undo
t2-stretching; (c) The CMP gather reconstructed from (b) given by Yilmaz (1989).

CREWES Research Report -Volume 15 (2003) 17

Cao et al.

(a) (b) (c) (d)

FIG. 7. (a) The reconstructed CMP gather from Figure 6b; (b) Only multiples constructed from
Figure 6b; (c) Only primaries constructed from Figure 6b; (d) Subtraction of (b) from Figure 5c from
Yilmaz (1989).

THE HYPERBOLIC RADON TRANSFORM

The Hyperbolic Radon Transform

A hyperbolic Radon transform performed on NMO-corrected gather, which is more

correct than a parabola, was introduced by Foster and Mosher (1992). Foster and Mosher
(1992) mentioned that there are two important conditions for keeping low cost of these
computations. The first is that the stacking surface should be time-invariant so that
computations may be performed in the frequency-space domain, and the second is that
the matrix operators should have Toeplitz form so that fast solvers may be used.

Now we copy equation (1), the generalized Radon transform, as:

∞
u (q,τ ) = ∫ d ( x, t = τ + qφ ( x ))dx ， (69)
−∞

where d is the original seismic data, x is the offset, t is the two-way traveltime, u is the
transform in model space, τ is the intercept time, and q is the curvature parameter.

Equation (69) is transformed to the frequency domain as:

18 CREWES Research Report -Volume 15 (2003)

 Radon transform and multiple attenuation

∞
U (q, ω ) = ∫ D( x, ω )e iωqφ ( x ) dx ， (70)
−∞

The summation along the curvature defined by parameter q, now becomes an integration
of phase shifts in the frequency domain. The discrete form of equation (70) is:

U (q j , ω ) = ∑ D(x k , ω )e
N
iωq jφ ( xk )
∆x k ， (71)
k =1

Since multiples have hyperbolic moveout curves relative to traveltime and offset
distance, Foster and Mosher (1992) gave the factor of the time-delay function (phase
shift) as follows:

φ ( x k ) = x k2 + z ref
2
− z ref ， (72)

where xk is the offset receiver position and zref is a constant parameter defined as the
reference depth. The choice of zref is not entirely arbitrary because the difference between
these hyperbolae and those of reflected waves is controlled by this parameter. The
smallerφ( xk), the more compact the events will appear in the transform domain. With the
value of zref, events reflected from this depth are optimally resolved.

The basic assumption of this method when applied to multiples suppression is that the
moveout of multiples is different from that of the primaries.

The Shifted-Hyperbolic Radon Transform

Malovichko (1978) and Castle (1994) derived the shifted-hyperbolic NMO equation
for a horizontally layered model as:

h2
t = τ s + τ 02 + ， (73)
v2

where

CREWES Research Report -Volume 15 (2003) 19

Cao et al.

τ s = τ 0 + (S − 1) , (74)

t0
τ0 = , (75)
S

v 2 = SVrms
2
, (76)

µ4
S= , (77)
µ 22

N N
µ n = ∑ ∆t iVi n ∑ ∆t i ， (78)
i =1 i =1

The shifted-hyperbolic curve represents a Dix NMO equation shifted by the time τs and
exact through fourth order in offset. Eequation (73)can be written as:

t 02 µ 24 x 2 µ 22 t 0 µ 22
t = t0 + + − ， (79)
µ 42 µ 4Vrms
2
µ4

or more simple:

2
t t  x2
t = t0 − 0 +  0  + 2
， (80)
S S SVrms

With the summation curve of equation (80), the function φ(x) in equation (4) becomes
(Oppert 2002):

4z 2 x 2 2z
φ (x ) = + − ， (81)
S2 S S

CONCLUSIONS
In the slant-stack domain, multiples are periodic for every p value. Predictive
deconvolution can be applied in the slant-stack domain to suppress multiples. The
parabolic Radon transform and the hyperbolic Radon transform can be performed to
attenuating multiple based on the velocity discrimination. One of the problems associated
with the Radon transform is that a CMP gather only includes a cable-length portion of a
hyperbolic traveltime trajectory. The finite cable length will cause smearing of the

20 CREWES Research Report -Volume 15 (2003)

 Radon transform and multiple attenuation

stacked amplitudes along the velocity axis. This can be overcome by a solution given by
Thorson and Claerbout (1985).

COMMENTS
In Figure 6b, the CMP gather in the Radon domain doesn’t have very good resolution.
Another sample is shown in Figure 8, which is a common-scatter point (CSP) gather (real
data from Alberta) and its velocity stack, which is a hyperbolic Radon transform that
shows high resolution. The above principle of multiple attenuation will be applied to
these types of CSP gathers formed by the equivalent offset method (EOM), Bancroft
(1998). It is proposed that the improved focussing of CSP gathers, both spatially and
temporally, will produce an improvement in the attenuation of multiples in prestack
migrations.

FIG. 8. Velocity stack or the hyperbolic Radon transform of a CSP gather (real data from Alberta)
from Bancroft (2003).

CREWES Research Report -Volume 15 (2003) 21

Cao et al.

ACKNOWLEDGEMENTS
Thanks to Dr. Bancroft and Dr. Brown for instruction, discussion, and suggestions.
Thanks to the CREWES Project for funding this research.

REFERENCE
Alam, A.. and Austin, J., 1981, Suppression of multiples using slant stacks: 51st Ann. Internat. Mtg., Soc.
Expl. Geophys., Expanded Abstracts, 3225-3257.
Bancroft, J.C., Geiger, H.D., and Margrave, G.F., 1998, The equivalent offset method of prestack time
migration: Geophysics, 63, 2042-2053.
Bancroft, J.C., 2003, A practical understanding of pre- and post-stack migration: SEG Course Notes.
Beylkin, G., 1987, Discrete Radon transform: IEEE Trans. Acoust., Speech, and Signal Proc., ASSP-35,
No.2, 162-172.
Castle, R. J., 1994, A theory of normal moveout: Geophysics, 20, 68-86.
Chapman, C.H., 1981, Generalized Radon transforms and slant stacks: Geophys. J. R. Astr. Soc, 66,
445-453.
Claerbout, J.F., 1985, Imaging the Earth’s Interior: Blackwell Scientific Publications.
Foster, D.J. and Mosher, C.C., 1992, Suppression of multiple reflections using the Radon transform:
Geophysics, 57, 386-395.
Hampson, D., 1986, Inverse velocity stacking for multiple elimination: J. Can. Soc. Expl. Geophys., 22,
44-55.
Malovichko, A.A., 1978, A new representation of the traveltime curve of reflected waves in horizontally
layered media: Applied Geophysics, 91, 47-53, (in Russian).
Oppert, S.K., 2002, Radon methods for improved imaging of high-velocity layers using mode-converted
energy: M.Sc. Thesis, Univ. of Calgary.
Oppert, S.K. and Brown, R.J., 2002, Improved Radon transforms for filtering of coherent noise, CREWES
Research Report, 14.
Taner, M.T., 1980, Long-period sea-floor multiples and their attenuation: Geophys. Prosp., 28, 30-48.
Thorson, J.R. and Claerbout, J.F., 1985, Velocity-stack and slant-stack stochastic inversion: Geophysics,
50, 2727-2741.
Treitel, S., Gutowski, P.R., and Wagner, D.E., 1982, Plane-wave decomposition of seismograms:
Geophysics, 47, 1375-1401.
Sacchi, M.D. and Tadeusz, J.U, 1995, High-resolution velocity gathers and offset space reconstruction:
Geophysics, 60, 1169-1177.
Yilmaz, Ö., 1987, Seismic Data Processing: Soc. Expl. Geophys.
Yilmaz, Ö., 1989, Velocity-stack processing: Geophys. Prosp., 37, 357-382.
Yilmaz, Ö., 2001, Seismic Data Processing: Soc. Expl. Geophys.
Zhou, B. and Greenhalgh, S.A., 1994, Linear and parabolic τ-p transforms revisited: Geophysics, 59,
1133-1149.

22 CREWES Research Report -Volume 15 (2003)