147

F-x linear prediction filtering of seismic images

Mark P. Harrison

ABSTRACT

The f-x linear prediction filtering algorithm is reviewed and tested on several
synthetic images. It is found that the f-x filter, when applied to noise-free synthetics,
produces little or no attenuation of continuous layers, but does laterally smear sharp
discontinuities. On noisy synthetic images, numerical measurements indicate that the f-x
filter performs better at attenuating random noise than does the f-k filter. The f-x filter,
however, produces greater lateral smearing of discontinuities than does the f-k filter. The
residual noise after f-x filtering still appears fairly random, and the filter does not give rise
to the same type of coherent "streaks" that a severe f-k filter is seen to create. In addition,
the f-x filter is able to extract the signal without any guidance from the user, whereas an f-k
dip reject filter must be manually selected, usually after inspection of a f-k spectrum plot.
The f-x filter, however, is not able to discriminate between coherent noise with large dip
and true events. Application of the f-x filter to an actual seismic image produces good
results, and no attenuation of coherent signal is seen to occur.

INTRODUCTION

The signal-to-noise ratio for converted-wave stack images is often poor. This
usually necessitates the application of some sort of image noise attenuation process, often
in the form of f-k (frequency-wavenumber) or Karhunen-Loeve filtering (Jones and Levy,
1987). This paper looks at a different approach to signal-enhancement, f-x filtering, in
which the seismic image is modelled as being composed of a number of linearly-coherent
reflections. This justifies the use of a prediction method in the spatial direction of the f-x
domain to optimally extract linear features and suppress random noise. This method was
first proposed for seismic data by Canales (1984), and has since been elaborated upon by
others (e.g. Gulunay, 1986). In the following sections the theory behind the method will
be reviewed and a comparison of the relative performance of f-x and f-k filtering will be
made.

THEORY

A seismic image represents a collection of zero-mean amplitude values that are

functions of time t and horizontal location x (trace number). The image can usually be
modelled at some scale (Canales, 1984) as being composed of a number N of continuous
dipping reflectors, each with slope si, i.e.,
N

a (t,x) = _ wi * _(t-ti)
i:l , (1)
 148

where w I is the temporal wavelet associated with the i'th reflector, convolved with a Dirac
spike located at time ti, and

ti = '_i + six.

Xl is the intercept time at some reference x-location, and sl is the slope of the reflector.
Taking the fourier transform of equation 1 w.r.t time gives
N
a(co,x) = _ Wi((o) e-J°(_'_)
i=l

where Wl(a)) is the fourier transform of the wavelet wl(t). This can be rewritten as
N
a(c0,x) = _ Ci(co)e-J_x
i=l

where Cl(o)) is a complex function of co only. This shows that each frequency, when
viewed in the x-direction, is just a sum of weighed sinusoids of varying amplitudes and
periods, implying that changes in a frequency component in the x-direction are predictable.
Given this predictable nature, it is possible to design a unit-distance prediction
operator for each frequency that gives, in the least-squares sense, the most likely value for
the next sample based on previous samples. This leads to the design of a complex least-
squared-error prediction filter. The theory behind complex prediction filtering can be found
in Treitel (1974), and is reviewed here. Letting the vector f be the prediction operator of
length m+l and the vector _ be the predicted values of a, then, for each frequency, the filter
equations can be written as

a00 ... 0
al a0
• al

• = 0 an ao
-, 0
_am+n/

an.1

or
_= Af,

where the ai are samples in the x-direction and the co subscript has been dropped. Defining
the desired output as the vector d, which, in this case, is just the sequence advanced by one
sample, then the prediction error e will be

e=d-_
=d- Af,
and the error energy will be
I = erie

where eII is the transposed complex conjugate ofe. The error energy becomes
 149

I = (d-Af)H(d-Af)
= dHd. ftfArl d _ dHAf + fHAHAf.

The error energy can be minimized by taking the derivative w.r.t, the filter vector f and
setting to zero, i.e.,

_)f- dHA+fHAHA=0,
or
fHAHA = dHA. (2)

Takingthecomplexconjugate
andtransposing,
this
becomes

AHAf = AHd.

Def'ming the complex autocorrelation matrix R as

R = AHA

and the complex cross-correlation matrix g as

g= And,
then equation 2 becomes
Rf = g. (3)

These are the complex-valued normal equations that must be solved for the complex filter f.
The complex matrix R can be written in the form

R=P+jQ

where P is real and symmetric, and Q is real and skew-symmetric, i.e.,

Q_ = -Q

where Qt is the transpose of Q. Treitel (1974) shows that by breaking equation 3 into it's
real and imaginary parts, it can be rewritten as a second matrix equation;

[g"l (4)

where Re and Im designate the real and imaginary parts of a function. This real-valued
matrix equation can be solved to give the filter coefficients f. For the process being studied
here, the g vector is just the first sub-diagonal column of the R matrix, plus one additional
lag. The left matrix in equation 4 is block-Toeplitz (Treitel, 1974), and can be inverted
using a Levinson-like recursion given by Robinson (1967).
Solving Equation 4 will result in the prediction operator to be applied in the +x
direction. Prediction can also be done in the -x direction, giving left and right prediction
operators. For prediction in the -x direction, reversing the sample order and going through
a similar derivation leads to the folowing;
 150

R*L = g*.

Taking the complex conjugate, this becomes

Rf.* = g,

which shows that the -x prediction operator is just the complex-conjugate of the +x
prediction operator. A single filter, incorporating both +x and -x prediction, is then given
by

I"2 ....'2'v'2
f*mfm'l
" f_o_ f 1,
(5)
where the coefficients have been divided by 2 to give proper normalization. The
computation and application of this filter to the complex series a(x) for each frequency then
gives the prediction-filtered set _(t.o,x), which is then inverse-transformed to give the f-x
filtered output.
A difference image can be constructed by taking the point-by-point difference
between the input image and the filtered image;

d(t,x)
= a(t,x)
- _(t,x) (6)

and is often useful in evaluating the performance of the filter.

METHOD AND RESULTS

The flow followed in implementing the f-x algorithm is outlined in Figure 1. To

illustrate the method, a simple synthetic image (Figure 2) was created with only a single
linear event. This event was generated by convolving a 8-75 hz bandpass filter operator
with a spike placed at the time position appropriate for each trace. The input image is first
fourier-transformed in time to give a complex frequency series at each trace (x) location.
These traces are then reordered to give for each frequency a sequence of complex samples,
one sample from each of the transformed traces in the x-direction. The complex
autocorrelation for each of the frequency sequences is then generated, and the first m+l
lags are used to generate the prediction operator. For the examples given in this paper, 7
lags were used, giving a total operator length of 15 samples (7 in each direction). The
resulting operator is then convolved with the x-ordered sequence for that frequency, and
the process is repeated until all frequencies in the transformed data set have been done.
These Faltered sequences are reordered back into their respective x-trace positions, and the
inverse fourier transform is applied, giving the filtered result shown in Figure 3. To assess
the result of applying the filter, the sample-by-sample differences between the input image
and the filtered output image were computed using equation 6, and are plotted in Figure 4.
As expected for this simple example, the prediction worked very well, and Figure 4 shows
that there is no visible difference between the input and output images.

To assess the filter's ability to suppress random noise, the synthetic image of
Figure 2 was corrupted with random noise with a bandwidth of 6-120 hz, which
approximates the bandwidth of a seismic field recording system. The variance of the noise
was made equal to the variance of the 200 ms bandpass wavelet used to construct the
 151

original synthetic. The resulting image is shown in Figure 5, and is seen to have a severe
noise level, close to the upper limit of that normally found on a seismic image. In order to
roughly assess the filter's noise-suppression capabilities, a portion of the image, outlined in
Figure 5, was selected over which the noise variance levels before and after filtering were
computed. In this case, it was found that the filter attenuated the dipping reflector by about
35%. To account for this, the filtered image was scaled by a constant factor to bring the
amplitudes of the dipping event closer to it's original level (Figure 6), and a difference
image was computed (Figure 7). From the difference image, it is seen that a substantial
amount of noise has been rejected. Measurements within the control portion of the image
indicate a reduction in the noise variance of 16.6 db.

An f-k filter was also applied to the noisy image for comparison, giving the result
shown in Figure 8. A comparison of the f-k filtered image and the f-x filtered image
(Figure 6) shows that both have done a comparable job of attenuating the noise. The noise
left by the f-k f'dter appears very coherent, whereas the noise left by the f-x is more random
and lower frequency. Measurements within the control portion indicate a reduction in the
noise variance of 14.8 db, compared to 16.6 db for the f-x filter.

A more complicated synthetic was constructed, having reflectors with

discontinuities and conflicting dips, as well as two large-amplitude noise glitches. Random
noise identical in amplitude and frequency to that used in the previous example was added,
giving the results shown in Figure 9. A control portion for attenuation comparison was
also selected for this image, and is outlined in the figure. The result of f-x filtering the
image is shown in Figure 10. The two large-amplitude glitches are largely removed, and
there is little smearing of the glitches into adjacent traces. The discontinuity on the top flat
horizon is seen to have been spread horizontally over a distance of seven traces (the width
of the prediction filter). The difference image, which is not shown here, shows about a
15% loss of amplitude on the two steepest events, relative to the other reflectors.
Measurements made within the control portion give a reduction in noise variance of 8.4 db.

An f-k filter was also applied to the image of Figure 9, giving the result shown in
Figure 11. Comparison of the f-x filtered image (Figure 10) and the f-k filtered image
(Figure 11) indicates that both methods have achieved roughly the same amount of noise
attenuation, with the noise remaining in the f-k filtered image again appearing higher
frequency and less random than in the f-x filtered image. The f-k filter is seen to produce
less lateral tapering of the discontinuity on the top flat event than does the f-x filter.
Measurements made within the control portion indicate a 4.7 db reduction in the noise
variance, compared to 8.4 db for the f-x filter.

Plotted in Figure 12 is an f-k power spectrum of the noisy input image, which
shows the noise to be evenly distributed over the entire f-k spectrum. The most steeply-
dipping event is seen to alias at frequencies greater than about 65 hz. An f-k power plot of
the f-k f'fltered section of Figure 11 is displayed in Figure 14, and shows that the noise has
been removed from the spectrum everywhere except within the wedge enclosing the
dipping events. The filter has also removed the aliased frequencies of the most steeply-
dipping event, which produces a change in waveform shape for that event. Figure 13 is an
f-k power plot of the f-x filtered image of Figure 10, from which it is seen that the noise
has been uniformly attenuated throughout the spectrum, including within and beneath the
signal band where f-k filtering has had no effect.

As a final example, the f-x filter was run on the radial-component section of line
FS90-1 in the Springbank, Alberta area (Lawton and Harrison, 1990) The original and f-x
filtered sections are displayed in Figures 15 and 16 respectively. For comparison, an f-k
filter was also applied, giving the section shown in Figure 17. The f-x filter is seen to give
 152

better overall continuity than does the f-k filter. The very high-dip noise trains have not
been attenuated by the f-x filter, but have been removed by the f-k filter.

DISCUSSION

The main parameters that the user of an f-x filter has to decide upon are the length
of the prediction filter and the size of the window in which it is designed. In the
complicated synthetic, a longer operator and/or a smaller window size probably would have
given better preservation of the most steeply-dipping events, but no work has been done to
confirm this. The choice of seven lags for the operator length appears to work well in the
examples presented here, but it is possible that a longer operator might have produced
better results. A major constraint on the length of the filter operator, however, is the
computational cost of having to design a filter for each individual frequency. In order that
the method be practical, it is desirable to keep the number of lags used as small as possible.

An area for further testing of the f-x filter is in cases where events are curved, and
have amplitude variations. If these events are altered or discarded by the filter, then the
method may have limited use in areas of large sub-surface structure. Also, Gulunay (1986)
gives a proof that f-x filtering does not work correctly if events with conflicting dips are
present, as could occur in structured areas. From the synthetic images shown here, as well
as other tests, it appears, however, that the filter still performs well when this happens.

From the synthetic examples, it is seen that the f-x filter is better at attenuating
random noise than is the f-k f'flter. The residual noise after f-x filtering still appears fairly
random, and it does not give rise to the same type of coherent "streaks" that a severe f-k
filter is known to produce. In addition, the f-x filter is able to extract the signal without any
guidance from the user, whereas an f-k dip reject filter must be manually selected, usually
after inspection of an f-k spectrum plot. The f-x filter therefore appears to have some
important advantages over the f-k filtering method. It is seen from Figure 16 that the f-x
filter is not able to distinguish coherent linear noise from true reflections, which can be a
disadvantage. It is possible that better results could be obtained in some cases by using
both an f-x filter to remove random noise, and a mild f-x to remove high-dip coherent
noise.

CONCLUSIONS

The f-x linear prediction filter was reviewed and tested on several synthetic images.
It was found that the filter, when applied to noise-free synthetics, produces little or no
attenuation of continuous layers, but does laterally smear sharp discontinuities. On noisy
synthetic images, numerical measurements indicate the f-x filter performs better at
attenuating random noise than the f-k filter. The residual noise after f-x filtering still
appears fairly random, and the filter does not give rise to the same type of coherent
"streaks" that a severe f-k filter was seen to produce. In addition, the f-x filter is able to
extract the signal without any guidance from the user, whereas an f-k dip reject filter must
be manually selected, usually after inspection of a f-k spectrum plot. The f-x filter is not
able to attenuate coherent dipping noise, which appears to the algorithm as valid signal.
Application of the f-x filter to an actual seismic image produced results which compared
favorably to those obtained by f-k filtering.
 153

REFERENCES

Canales, L.L., 1984, Random noise attenuation: Presented at the 54th Ann. Mtg., Soc. Explor. Geoph.
Gulunay, N., 1986, FXDECON and complex Wiener prediction filter: Presented at the 56th Ann. Mtg.,
Soc. Explor. Geoph.
Jones, I.F., and Levy, S., 1987, Signal-to-noise ratio enhancement in multi-channel seismic data via the
Karhunen-Loeve transform: Geophysical Prospecting, v. 35, 12-32.
Lawton, D., and Harrison, M., 1990, A two-component reflection seismic survey, Springbank, Alberta: in
this volume.
Robinson, E.A., 1967, Multichannel time series analysis with digital computer programs: San Francisco,
Holden-Day.
Treitel, S., 1974, The complex wiener f'flter: Geophysics, v. 39, 169-173.
 = 51]

a(t,x)
Input image da|a Set

Fourier transform in time

a(t,x)_ a(_,x)

$ |

Reorder by frequency [
a(x) for each co I
Generate the autocorrelationfunction
and prediction filter for each frequency

I
Filter the offset sequences
a(x) _(x) foreach co

Reorder by offset

Inverse fouriertransform
_(_,x) _ a(t,x)

+
Compute the difference image

d(t,x) = a(t,x) - _(t,x)

Fig. 1. Process flowchart for the fix filtering algorithm.

 155

TraceNumber
101 91 81 71 61 51 41 31 21 11 0.0

I__H__i__H__H[__i__________I_]
!i____H_iH_i__H_i_______i____i °'
°._[_H_i_i_I_H_I_L_]_j_ °'
o,lllllllllll]lll
IIIIIII tlll_llllllllllllllll
°_
Zo.ollllll]lllillllllllllHiiiilllillliilllilltllll/_llllllJi
o.,
io_lllllllllllllllll[llll[llllllllllll_llllllllEl[llllltil[ll
illltl
°'
o, I111_1 IIII °°
o8_llllllllllllllllllllllllllllllllHIIII]lllllllllllll]HI
°"
o°_l IIIIIIIIII]ll IJIJI]llllJlll]2i
,ioi__i__i______i_____H___________i____
Fig. 2. A simple model with a single linear retlection. The wavelet bandwidth used is
8-75 hz.

Trace Number
101 91 81 71 61 51 41 31 21 11 1
O0 O0

0.1 0.1

02 02

03 O0

0.4 04

z
• 05 O6 fD
_
E
Og 0.8

0.7 0.7

0.8 0.8

0.9 09

l.O 1.0

Fig. 3. The f-x filtered version of thc imagc in Figure 2.

 156.

Trace Number
81 71 61 51 41 31 21 11 !
D.O 0.0

Ol O.l

02 0.2

0.3 0.:3

0.4 0.4

oJ 05 0.6

06 08

0.7 0.7

08 0.8

0_9 0.9

I0 |.0

Fig. 4. The difference between the input image (Figane 2) and the f-x filtered image
(Figure 3).

Trace Number
81 71 81 51 41 31 21 11 I
0.0

OA

0.2

03

0.4

z
_ 0.5

• 0.6

0.7

0.8

0,9

1.0

Fig. 5. Band-limited random noise added to the synthetic image in Figure 2. The box
indicates the area over which noise variances were calculated.
 157

Trace Number
101 91 81 71 61 51 41 31 21 11 1
0,0

Ol

0,2

0.3

0.4

05 •

0g

0.7

0.8

0.9

1.0

Fig. 6. The f-x filtered version of the image in Figure 5.

Trace Number
I01 91 81 71 61 51 41 31
O0 0.0

01 0,1

0.2 0.2

0 3 0,3

0.4 0.4

05 05 (D

06 0.6

0 7 0.7

0.@ 08

0.9 09

I0 I0

Fig. 7. The difference between the input image (Figure 5) and the f-x filtered image
(Figure 6).
 158

Trace Number

91 81 71 61 51 4] 31 21 Ii l

Fig. 8. The result of applying an f-k filter to the noisy image of Figure 5.

Trace Number

71 61 51 41 31 21 II I
0.0 O0

0.I 0.I

02 0.2

0.3 0.3

0.4 04

z
_1 05 05 _

0 B 0.6

0.7 O.7

0.15 08

0.9 0.9

1.0 1.0

Fig. 9. A more complicated synthetic image with various dipping reflections, a

discontinuity, a pair of noise glitches, and band-limited noise. The box indicates
the area over which noise variances were calculated.
 159

Trace Number
1ol 9t 61 71 61 51 4z 31 zl zt 1
OO O0

01 01

0.2 02

0.3 0.3

04 0.4

05 0.5 f_
E
06 06

0.7 07

08 08

09 09

I0 10

Pig. 10. The f-x f'dtered version of the image in Figure 9.

Trace Number
101 91 fll 71 61 61 41 31 21 11 1

Fig. 11. The f-k f'fltered version of the image in Figure 9.

NYQUIST 0.0 NTQUIST

-300.00 -150.00 0.0 150.00 300.00

W A V E N U M B E R ( C Y C L E S / 1 000 Ii

Fig. 12. An f-k spectrum power plot of the input image shown in Figure 9.
 161

o.o

300.00 450.00
WAVENUMBER ( C Y C L E S / 1 000

Fig. 13. An f-k spectrum power plot of the f-x filtered image shown in Figure 10.
 62

Nrouisr

300.00
WAVENUMBER tCYCLES/1000

Fig. 14. Anf-kspectrumpowerplotof the f-k filtered image shown in Figure 11.
 TIME (S)
o. m. o. m. o. m. o. m
 165

121 13"f 141 151 161 171 181 191 201

1.5

I'l

2.O

Fig. 17. The f-k fihered version ofthe section showninFigure 15.