first break volume 26, June 2008 special topic

Leveraging Technology

Extending seismic bandwidth using the

continuous wavelet transform
Michael Smith1*, Gary Perry1, Jaime Stein2, Alexandre Bertrand3 and Gary Yu2 describe the
realization of a Bandwidth Extension technique with synthetic and real examples.

R esolution is the ability to identify individual features

or details in a given image. By the 3D nature of
seismic data, seismic resolution involves both verti-
cal (temporal) and horizontal (spatial) resolution.
Temporal resolution is a function of the frequency content of
acquisition and processing for a long time. Spectral broadening,
primarily high frequency extension, has been a controversial
subject because of the belief that, due to the earth absorption
of preferentially high frequencies, these are lost and it is impos-
sible to restore them to the spectrum. Many techniques, like
a given signal. Achieving optimal thin bed resolution requires Q-compensation (e.g., Wang, 2006) have proven this assump-
a broadband spectrum. Since most seismic is very band-limited tion incorrect.
it is desirable to extend the bandwidth of the data in a manner Given a temporal sampling rate Δts the maximum theoret-
that is both verifiable and consistent with the geology. ical frequency, called the Nyquist frequency Fny, at which the
In seismic processing, many methods are employed to data will remained un-aliased can be defined as:
extend the bandwidth of the signal while maintaining an accept-
able signal-to-noise ratio (SNR). Many deconvolution processes, [1]
such as spiking, attempt to sharpen the wavelet. Spectral whit-
ening is also often employed to boost frequencies, most often on
the high end (but sometimes on the low end as well). Generally This limit represents the maximum frequency at which the sig-
speaking, the two problems most often encountered with any of nal contains energy still compatible with the sampling theorem.
these methods are that they tend to increase the noise level more The classical resolution limit λR derived from the Rayleigh crite-
than the signal and the amount of improvement expected is usu- ria states that top and bottom reflectivities of a thin bed can be
ally substantially less than an octave. distinguished for thickness of up to a quarter of the dominant
However, new methods have been introduced that produce wavelength λD (λR ~1/4 λD). However, this does not mean that
results that compare favourably to both well log synthetic seis- it is impossible to detect thinner beds than this: Widess (1973)
mograms and to geologic conditions (e.g., high frequency imag- proved conclusively that the amplitude of a reflection coming
ing or HFI, Hamarbatan et al., 2006). Most of these methods from such a thin bed varies with the thickness of the bed. He
have concentrated on extending the upper end of the spectrum showed that although completely resolving top and base of a
but, especially when one is going to invert the seismic data, thin bed may be difficult, this thickness still has a measurable
extending the lower end is also very desirable. The low frequen- effect on the signal. His work pushed the limit of resolution to
cy earth model should, ideally, extend no further than to about a new minimum of about half the Rayleigh criterion. The new
4-5 Hz. Any information lower than this point tends to improp- thickness of resolution λW is given by λW ~1/2 λR ~1/8 λD.
erly influence the resulting inversion between control points. This evidence of reflectivity below resolution implies that
With these criteria in mind a proprietary method called BE the seismic wavelet contains information about reflectors
(Bandwidth Extension) was developed utilizing the continuous beyond the dominant frequency of the seismic wavelet and that
wavelet transform (CWT) to extend the signal frequencies both high frequencies could potentially be restored through digital
upward and downward. We first introduce the background the- processing – the motivation of the new bandwidth extension
ory on resolution issues. Then we introduce the BE technique, development.
which is finally applied to both synthetic and real examples.
Bandwidth extension (BE) using CWT
Theoretical background on resolution The traditional method of analyzing a seismic trace in the fre-
Extending the bandwidth of a signal, both in the high end and quency domain uses the Fourier transform or the fast Fourier
the low end of the spectrum, has been a goal of seismic data transform (FFT). The problem with using a FFT is that it

1
Geotrace, Dallas, TX, USA.
2
Geotrace, Houston, TX, USA.
3
Geotrace Norge, Stavanger, Norway.
* Corresponding author, email: [email protected].

© 2008 EAGE www.firstbreak.org 97

 special topic first break volume 26, June 2008

Leveraging Technology

transforms local information into global information. The domain, it is limited by the uncertainty principle that states
solution of using a short Fourier transform alleviates some we cannot know time and frequency with the same accura-
of the problem in that it preserves some local information. cy simultaneously. This obviously limits the amount of infor-
However, the short window restricts the frequencies that can mation usable to predict the higher and lower order harmon-
be analyzed and the extent of localization is limited by the ics at any given time. This is why the redundant nature of the
single window length chosen. The wavelet transform does not CWT was chosen as our transform in contrast with the dis-
have these same limitations and permits the analysis of both crete wavelet transform, (DWT) which uses an orthonormal
local information and global information simultaneously. basis and thus has no redundancy. The fuzzy nature of the
Here, we use the CWT to perform a time-series analysis CWT allows us to compute harmonics and sub-harmonics in
of a seismic trace that decomposes the trace into its respective a fuzzy way, that is, with incomplete information in both the
amplitude and phase components in both frequency and time time and frequency domains.
domains. The CWT is defined as the convolution of a time The time and frequency resolutions vary with scale(s) by
series f(t) with a scaled (s) and translated (τ) wavelet ψ(t). the standard deviations of the mother wavelet in the time
domain (Δt) and the frequency domain (Δω) as sΔt and Δω/s.
[2] So the time resolution increases and the frequency resolution
decreases with smaller scales (higher frequencies), and the time
resolution decreases and the frequency resolution increases
where (*) indicates the complex conjugate. with larger scales (lower frequencies). In other words, in the
low end of the spectrum frequency resolution is more impor-
The scaled wavelets are called daughter wavelets as they are tant than time resolution, and at the high end of the spectrum
scaled from the mother wavelet ψ. just the opposite is true. The result is that the product of the
The CWT ‘is in theory infinitely redundant - the extreme standard deviations of time and frequency are constant for the
redundancy is less of a problem than one might imagine; a entire frequency spectrum analyzed (Qian, 2002). This gives
number of researchers have found ways of rapidly extracting us the most optimal picture of the time-frequency domain.
the essential information from these redundant transforms’ By using the time-series analysis of the CWT, we can com-
(Hubbard, 1998). Because the implementation of the CWT is a pute harmonics and sub-harmonics from the available band-
discrete operator and not a truly continuous operator, a choice width in the seismic wavelet. Harmonics are an integer mul-
needs to be made as to how many daughter wavelets will be tiple of a fundamental frequency, and sub-harmonics a ratio
used, thus how much redundancy (infinite redundancy is just of one over an integer multiple (e.g., ½). Since we are using a
not practical). A minimum of 10 scales (voices) per octave complex wavelet, amplitude and phase information is avail-
is sufficient to recreate the input time series from the trans- able for this computation. By convolving harmonic and sub-
form by computing its reconstruction. Furthermore, the moth- harmonic frequencies onto the seismic trace, we can recover
er wavelet ψ must also meet the admissibility condition (Qian, this information and better reveal reflectivity that otherwise
2002) as the analyzing wavelet is going to be used to recon- is difficult to detect. This unmasking of reflectivity is accom-
struct the original time series after bandwidth is extended. plished by increasing the bandwidth of the seismic wave-
We chose to use the Morlet wavelet as the mother wavelet. let with harmonic and/or sub-harmonic frequencies added
The Morlet wavelet is a complex function representing a plane to the seismic trace. Since this is a convolution-like process,
wave modulated by a Gaussian function. The complex nature harmonic or sub-harmonic frequencies, which correspond to
of the wavelet allows the calculation of amplitude and phase reflectivity in the seismic trace, will be added to the seismic
for each scale at distinct times. The choice of this wavelet was wavelet, but harmonics that do not correspond to reflectivi-
driven by the nature of the problem we are analyzing. In our ty will effectively fall out of the transform.
case, we are analyzing a seismic trace, and a Gaussian mod- The final step is reconstruction of the modified time
ulated plane wave would be a good match for seismic time series from the CWT domain of time and scale to a func-
series analysis. The Morlet wavelet is given by: tion of time only, our new seismic trace with extended band-
width. The reconstruction formula is given by:

[3]
[4]
The CWT ‘provides a very redundant, but also a very finely
detailed description of a signal in terms of both time and fre- where Cψ is given by the admissibility condition.
quency’ (Walker, 1999). These properties are utilized to pre-
dict the harmonics and sub-harmonics used for bandwidth In implementation, the seismic trace is transformed by the
extension. Because the CWT operates in the time-frequency CWT. The fundamental frequencies are chosen based on

98 www.firstbreak.org © 2008 EAGE

first break volume 26, June 2008 special topic
Leveraging Technology

user input as to the frequency range to use for computing

harmonics and sub-harmonics. The frequency points repre-
senting the range of frequencies to use are called the pivot
frequencies. Each pivot frequency is chosen from analyz-
ing the seismic spectrum, and an octave below this pivot
is used for computing harmonic frequencies or scales. For
sub-harmonics, a different pivot frequency is chosen from
the seismic spectrum, and an octave above this pivot is used
for computing sub-harmonic frequencies or scales. The har-
monics and/or sub-harmonics can extend upward/down-
ward 1-2 octaves, of course honouring Nyquist and 0 Hz
boundaries.
Energy density adjustments are made so that the har-
monic and sub-harmonic scales produce a better shaped
amplitude spectrum. These adjustments are made prior to
adding back the harmonic scales to the spectrum of the seis-
mic wavelet. The transformed seismic trace, with added har-
monics and/or sub-harmonics, is then reconstructed to give
the new bandwidth extended seismic trace.
Figure 1 Bandwidth Extension applied to the wedge model. (a) Low-
Wedge model example frequency synthetic (input to BE). (b) High-frequency extension of (a) (output
of BE). (c) High frequency synthetic to compare to BE results. Blue and green
We have tested the high frequency extension of the BE algo- arrows indicate the limit of resolution of the top and bottom reflectors of
rithm on a simple wedge model to analyze its performance the wedge respectively.

Figure 2 Input data (left) and bandwidth-extended data (right) from an onshore 3D survey. Red traces are synthetic seismograms produced from well log reflectiv-
ity using an extracted wavelet from the respective data. Horizontal timelines are separated by 100 ms.

© 2008 EAGE www.firstbreak.org 99

 special topic first break volume 26, June 2008

Leveraging Technology

in a classic resolution problem. The model is made of a flat fully resolved, according to Rayleigh’s criteria, at trace 40
top with positive reflectivity and a dipping bottom event (thickness = 11.7 ms) where both top and bottom reflectors
of equal positive reflectivity to the top. The bottom reflec- can be distinguished.
tor has a positive dip of 0.3 ms/trace, which means that the The new BE approach is used in an attempt to improve the
wedge is thickening with increasing trace numbers towards resolution of the wedge. After BE (Figure 1b), the top and bot-
the right of the section. Realistic frequency content (7-55 Hz tom reflectors of the wedge can be distinguished at trace 28,
at half points) is used to model the low frequency response which correspond to the same resolution as the same wedge
of the wedge (Figure 1a). At this frequency range the bed is modeled using a frequency range of 7-85 Hz (Figure 1c).

Figure 3 Amplitude and phase spectra of the extracted wavelets (Input on the left and BE on the right). These are very diagnostic of the process as the input
data is shown to have a good fit with the well at 17-55 Hz (less than two octaves) with near zero phase while the bandwidth extended data match the well from
10-120 Hz (3.5 octaves), still maintaining the near-zero phase fit.

Figure 4 The input data on the left exhibits useable frequencies up to about 55 Hz while the bandwidth-extended data (right) has high signal data well past 120
Hz with good correlation with the well.

100 www.firstbreak.org © 2008 EAGE

first break volume 26, June 2008 special topic
Leveraging Technology

Figure 5 At the top of this figure are vertical sections through the 3D at normal bandwidth (left) and after bandwidth extension (right). At bottom are two
horizon slices 20 ms below the top of the carbonate marker. The top of the carbonate reservoir is displayed as a black-dash horizon in the vertical sections and
the vertical section location is marked with the bold black line.

© 2008 EAGE www.firstbreak.org 101

 special topic first break volume 26, June 2008

Leveraging Technology

The application of BE to the wedge model has proved fy the porosity trend - a great time saver and productivity
successful at increasing seismic resolution according to tool because we can see and interpret the geology quickly
Rayleigh’s criteria. In other words, BE can significantly with good confidence.
improve the temporal resolution by extending optimally
the original bandwidth (Figure 1a) to a broader bandwidth Conclusions
(Figure 1b) that matches the targeted bandwidth (Figure 1c) Extension of bandwidth using harmonics and sub-har­
and ultimately increases the thin bed detectability. monics predicted and computed with the CWT will enhance
the seismic resolution allowing a more refine and detail
Real data example 1 interpretation. There are limitations: the sampling theorem
This first example is from an onshore 3D survey with thin limits the maximum recoverable reflectivity to Nyquist and
siliclastic reservoirs. The comparison before-and-after anti-aliasing filters will set this limit below Nyquist. Also,
BE is shown in Figure 2. The input frequency content is frequencies available for harmonic prediction and limits
approximately 17-55 Hz, and a relatively good match is introduced by a hard cut-off field filter and notch filter, as
obtained with the synthetic (correlation of 70%). With well as the uncertainty principle, affect the ability to extend
the application of BE, harmonics and sub-harmonics are bandwidth.
used to model one additional octave of data on both the The Widess model suggests that there is seismic reflec-
low and high side of the spectrum. The output frequency tivity available below one fourth of the dominant frequen-
content is 10-120 Hz. With this broader spectrum, we cy wavelength. This information can be extracted, result-
can observe much richer low frequencies, as well as more ing in an increase in resolution by adding harmonic and sub-
high frequency details. One disadvantage of most spectral harmonic frequencies back to the data. In addition, it is well
broadening techniques is that they also reduce SNR. Here, known that it is desirable to have at least one and a half
the SNR stays at the same level as indicated by good octaves, if not two octaves, of bandwidth to decrease side
match between the BE result and the high frequency syn- lobes and remove the ‘ringy’ character of our seismic data.
thetic (correlation of 68%). Once this is done, many features such as minor faults, on-
The amplitude and phase spectra of the extracted laps, pinch outs, and other stratigraphic subtle features come
wavelets (Input and BE) also are presented in Figure 3. to light. All of these features can have a significant impact on
They are another diagnosis of the confidence that lays in interpretation of seismic data and field discovery.
the bandwidth-extended data compared to the true Earth Extension of bandwidth to the low end of the spectrum
reflectivity measured by the well data. It shows that the is also possible with the use of sub-harmonics predicted
match between earth and seismic is good over a much from the available data. This capability is especially help-
broader frequency range (from less than two octaves to ful for broadband inversion work when the input seismic is
3.5 octaves). It also shows that BE produces a very stable missing a large part of the low end of the spectrum.
zero-phase spectrum over the entire frequency range. Three examples presented here (both synthetic and real
data) have proven the benefits of the new methodology.
Real data example 2 The thin bed resolution of the wedge model was significant-
This second real example is a carbonate reef play from an ly improved. In the real case studies not only were broader
onshore 3D dataset. The geological objective is to identify spectrums attained but the results were confirmed by well-
porosity with higher resolution data. Figure 4 illustrates based diagnostics.
the before and after BE comparison of the pre-stack time
migration (PSTM) section with its respective amplitude References
spectrum and the synthetics using extracted wavelet. The Hamarbatan, N., Ajaily, A. and Hassab, A. [2006] High frequency imag-
input frequency is up to 55 Hz and the BE-enhanced fre- ing in an exploitation production environment - the Murzuq Basin,
quency is up to 140 Hz. The seismic and well synthetic case study. 68th EAGE Annual Conference and Exhibition, Extended
correlations before and after BE were quite good and Abstract, F005.
comparable at 70%. Hubbard, B. B. [1998] The world according to wavelets - the story of a
To further understand the impact of broader band- mathematical technique in the making. 2nd Edition, A K Peters.
width, a before-and-after comparison is presented using Qian, S. [2002] Introduction to time-frequency and wavelet transforms.
a horizon slice at the top of reservoir and shown in Fig- Prentice-Hall.
ure 5. The improvement of resolution after BE for inter- Walker, J.S. [1999] A primer on wavelets and their scientific applications.
pretation in terms of subtle geological features and trends CRC Press.
is very evident. The colour scale was calibrated with wells Wang, Y. [2006] Inverse Q-filter for seismic resolution enhancement. Geo-
and dark blue represents good porosity. With higher band- physics, 71, 51-60.
width, it is now much easier and more reliable to identi- Widess, M. B. [1973] How thin is a thin bed? Geophysics, 38, 1176-1180.

102 www.firstbreak.org © 2008 EAGE