Seismic trace interpolation using the pyramid transform

Barry Hung*, Carl Notfors and Shuki Ronen, Veritas DGC inc.

Summary frequency λf. Thus, the input data need to have a

reasonable signal-to-noise (S/N) ratio at each frequency.
We present a new method called the pyramid transform for Otherwise, at those frequencies with low S/N ratio,
seismic trace interpolation. The pyramid transform is a inaccurate prediction coefficients will be used for
resampling of data in the f-x-y domain that gives rise to interpolating traces at higher frequencies. For field data,
frequency dependent spatial grids with a relationship that however, it is not uncommon to have a portion of the input
the sampling interval is inversely proportional to the frequencies that has a rather low S/N ratio. In this case, a
frequency. Such transformation is reversible for wave- more robust method than the conventional f-x-y method is
fields. Using prediction filter, we demonstrate that seismic needed.
trace interpolation in the pyramid domain is more robust in
the presence of noise and conflicting dips than interpolation In this paper, we discuss a new transform method called the
in the conventional f-x-y domain. pyramid transform that allows one prediction filter to be
estimated from a range of frequencies, preferably the
Introduction frequencies with high S/N ratio. The pyramid transform has
been proposed by Sun et al. (1996) for the application of
Seismic data are often spatially undersampled in 3-D noise reduction. We extend its application to trace
exploration. Trace interpolation, a well-known solution to interpolation in this work. In the following, we first show
this sampling deficiency, is often used to estimate the features of the pyramid transform and then, through the
unrecorded traces from a spatially undersampled dataset. results of trace interpolation, highlight that the prediction
Various interpolation methods for handling aliased events filtering in the pyramid transform is more robust against
have been proposed over the years and some of their noise and conflicting dips than in the conventional f-x-y
features, such as computational efficiency and data method.
accuracy, have been compared (Abma et al., 2003). One of
the methods that was used as a benchmark in the Theory – Pyramid Transform
comparison studies is called the f-x domain prediction
filtering method, which is used routinely in data processing. To perform the pyramid transform, a 3-D dataset is first
This technique is based on the principle that linear events Fourier transformed in time. This transforms the data to the
are predictable in the f-x domain without any prior f-x-y domain. Within this domain, the data is then spatially
knowledge of the dips of the events (Spitz, 1991). It resampled in such a way that the resulting spatial interval is
operates on 2-D regular data in the temporal frequency (f) inversely proportional to the frequency. That is, the lower
domain in such a way that, for an interpolation order of λ, a the frequency, the smaller the number of data samples. This
spatial prediction filter is designed at f with trace spacing leads to a pyramid shaped sampling (hence the name for the
∆x and then applied to predict data at frequency λf with transform). Moreover, the low frequencies are not over-
trace spacing ∆x/λ. parameterized after the transform which makes frequency
dependent grids more suitable for either inversion or
Based on the Spitz’s method, effort has been made to interpolation. The pyramid transform is based on
enhance the computational efficiency of the method and Shannon’s sampling theorem, so it is invertible. In other
improve the quality of the interpolation results. Porsani words, if one performs the forward and then the inverse
(1999) used the half-step prediction filters in the f-x domain pyramid transform, one may recover the original input data.
to speed up the application process of the filters by a simple This is illustrated in figure 1, where figure 1a shows the
convolution operation. Wang (2002) employed 2-D original data and figure 1b displays the result after the
prediction filters in the f-x-y domain for 3-D data to relax forward and inverse pyramid transform. The difference plot
the requirement that the events must be linear for between the two figures is depicted in figure 1c, which
predictability. With the f-x-y prediction filter, events that shows that the pyramid transform is invertible.
are nonlinear in one direction but linear in another may be
predicted. Hung et al. (2003) addressed the amplitude and Mathematically, the way to spatially resample an input data
the smearing of structural discontinuity issues by using in the pyramid domain is guided by the well-known
noncausal spatial filters. While these work contribute to the equation that relates the maximum non-aliased frequency fm
improvement of the Spitz’s method in one way or another, to a dip angle θ with trace interval δx:
their drawback is that, in the conventional f-x-y method, the v
fm= (1)
prediction filter needs to be estimated at each input 4δx sinθ
frequency f (up to fmax/λ) for interpolating traces at
 Seismic trace interpolation using the pyramid transform

where v is medium velocity. Rearranging equation (1) and efficient, because the process needs not be repeated for
setting θ=90o give rise to the following equation: each frequency, but also more stable and robust against
v noise, since a frequency range with high S/N ratio can
sin θ = =1 (2) always be chosen for the estimation. To illustrate this point,
4 f ' δx
consider the equation of the propagator z which, for a
where f’ is the critical frequency. Equation (2) indicates
single event, predicts the wavelet from x to x+∆x (Sun,
that for frequencies below f’, one may resample the spatial
1996):
intervals in such a way that the new intervals ∆xf (a
function of f) fulfill the relationship of ∆xf = v/4f and yet z = e id 2πf∆x (3)
retain dip information up to 90o; whereas, for frequencies where d is the dip of the event. In the conventional f-x-y
above f’, one can only obtain dip information of domain, z is a function of f. However, in the pyramid
θmax = sin-1 (v/4f’δx) with the original trace interval δx. The domain, z is frequency-invariant when ∆x is set to v/4f as
use of the critical frequency f’ for the pyramid transform is mentioned above. For multiple events, says three events,
schematically illustrated in figure 2 where figure 2a shows one may use the Van Der Monde matrix equation to relate
the input data cube in the f-x-y domain and figure 2b the the prediction coefficients p’s and the propagators z’s:
data volume after the pyramid transform. One may observe  z 2 z1 1  p   z 3 
that the volume of the data has shrunk significantly after  1   1  1 
 z 2 z 2 1  p  =  z 3  (4)
the pyramid transform. Hence, the pyramid is equivalent to  2 2   2   23 
the cube in the conventional f-x-y domain in terms of  z 3 z 3 1  p3   z 3 
information amount but with smaller data volume. In    
addition, the base of the pyramid occurs at frequency f’; Since the matrix equation is a linear system, the prediction
above f’ the data arrangement is same as the corresponding filter obtained from the frequency-invariant z’s is also
section in the f-x-y domain. independent of frequency in the pyramid domain. In
practice, the estimation of the prediction filter is performed
in the pyramid domain. From the filter, the frequency
invariant z’s are obtained. These z’s are then mapped back
to the original f-x-y domain to obtain the frequency
dependent prediction filter pf(f) which can then be used for
interpolating traces at frequency λf. Thus, the application of
the prediction filter is done in the original f-x-y domain.
y

x
(a) (b) f

(a)

Forward pyramid Inverse pyramid

transform transform
y
(c)
Figure 1. (a) The input data. (b) The result after the forward x
f
and inverse pyramid transform. (c) The difference plot
between (a) and (b). (b)

The feature of frequency dependent grids in the pyramid

domain has an important application for spatial prediction
f’
filtering. That is, the prediction filter is independent of
temporal frequency and hence a single prediction filter can
be estimated from a range of frequencies below the critical Figure 2. (a) Input data cube in the original f-x-y domain.
frequency. This makes the estimation of the filter not only (b) Data volume after the pyramid transform.
 Seismic trace interpolation using the pyramid transform

In the following section, we demonstrate how the pyramid (a)

transform can be applied to trace interpolation. We use only
the f-x domain in the following examples. However, the
theory which has been discussed above is applicable to
both the f-x and f-x-y domains.

Examples

Figure 3a shows a linear event that is contaminated with

noise (5–70 Hz) but has the highest S/N ratio in the
frequency range of 10-15 Hz. The even-numbered traces of (b)
this input were decimated. After the decimation, the event
is aliased above 50 Hz. The decimated traces were then
reproduced by the conventional f-x method and the result is
depicted in figure 3b. The difference between the original
input and the f-x interpolation result is displayed in figure
3c. One may observe from the difference plot that the
conventional f-x method does not perform well for this
example even for one linear event. It is because it uses
frequencies with low S/N ratio to obtain the prediction
filter for interpolation, i.e., for instance, the aliased (c)
frequency of 60 Hz uses the prediction filter estimated at
30 Hz, which has low S/N ratio, for interpolation. On the
contrary, when only the frequency range with the highest
S/N ratio (in this case 10-15 Hz) is used for the estimation
of the prediction filter in the pyramid domain, a greatly
improved interpolation result, as shown in figure 3d, is
obtained. From the difference plot shown in figure 3e, it is
apparent that the pyramid method performs better than the
conventional f-x method in the presence of noise.
(d)
In cases where conflicting dips as well as noise pose a
problem to the conventional f-x method, the pyramid
method may be used for generating better results. Figure 4a
shows the migrated image of subline 242 of the
SEG/EAGE model. Its highest S/N ratio is in the frequency
range of 10-20 Hz for most parts of the area. Again, the
even-numbered traces were removed and then reproduced
by interpolation. Figure 4b displays the result obtained by
the conventional f-x method. Using the frequency range of
10-20 Hz for the estimation of the prediction filter, the (e)
interpolation result obtained by the pyramid method is
plotted in figure 4c. It can be seen that the pyramid method
reproduces the data much better than the conventional f-x
method. In particular, the area on top of the salt dome and
the area around the salt in the timegate of 1.6-2.2 s, where
conflicting dips of the events make it difficult for the
conventional f-x method to estimate the prediction filters
accurately. Moreover, the computational time needed to
generate figures 4b and 4c was almost the same because, in
the pyramid method, the extra time needed to transform the Figure 3. (a) Original input before its even-numbered traces
data to the pyramid domain is offset by the time that is are removed. (b) Interpolation result using the conventional
saved in estimating only one prediction filter for a f-x method. (c) The difference plot between (a) and (b). (d)
frequency range. Interpolation result using the pyramid method. (e) The
difference plot between (a) and (d).
 Seismic trace interpolation using the pyramid transform

Conclusions

We have presented a new method for seismic trace

interpolation based on the pyramid transform. The pyramid
transform resamples the data in the f-x-y domain in such a
way that the new spatial interval is inversely proportional
to the frequency. This gives rise to an important feature for
prediction filtering that the prediction filters in the pyramid
domain are frequency independent and therefore can be
estimated from a range of frequencies. Using this feature, it
has been shown that the pyramid method is more robust
against noise and conflicting dips than the conventional
f-x-y method for the applications of trace interpolation.

References

(a) Abma, R., and Kabir, N., 2003, Comparisons of

interpolation methods in the presence of aliased events:
73rd Ann. Internat. Mtg., Soc. Expl. Geophys., Expanded
Abstracts, 1909-1912.

Hung, B., and Notfors, C., 2003, Seismic trace

interpolation using noncausal spatial filters in the f-x-y
domain: 65th Mtg.: Eur. Assn. Geosci. Eng., D21.

Porsani, M. J., 1999, Seismic trace interpolation using half-

step prediction filters: Geophysics 64, 1461-1467.

Spitz, S., 1991, Seismic trace interpolation in the F-X

domain: Geophysics, 56, 785-794.

Sun Y. and Ronen S., 1996. The pyramid transform and its
application to signal/noise separation: Stanford Exploration
(b) Project Report (SEP-93), 161-176.

Wang, Y., 2002, Seismic trace interpolation in the f-x-y

domain: Geophysics 67, 1232-1239.

Acknowledgments

We thank Veritas DGC for the permission to publish this

work.

(c)
Figure 4. (a) Original migrated subline of the SEG/EAGE
model before its even-numbered traces are removed.
(b) Interpolation result using the conventional f-x method.
(c) Interpolation result using the pyramid method.