Professional Documents
Culture Documents
technical article
Figure 1 Conventional noise attenuation scheme with 50% of original input add-back.
2004 EAGE
39
technical article
because we know the signal estimate is imperfect. Adding back some of the input data, i.e. adding back [some percentage x (signal + noise)], is an attempt to make up for the fact that the modelling did not fully capture all of the signal. Adding back some of the noisy input data is a way of diminishing edge effects and of overcoming spatial smearing and other artifacts created by the assumptions of the mathematical model. Practically speaking, the reason for adding back input data is to make the result look more realistic and less synthetic. It is helpful to consider what is being done to the signal (Fig. 1). The original input data (A) contains all of the signal and all of the noise. In our schematic, we describe this as signal=100 and noise=100. The signal is now modelled using some type of mathematical description, FX deconvolution for instance. We know that the signal model is never perfect. Let us say it is 90% correct, so at this point the modelled signal is an imperfect estimate: it has a signal level of 90 and still contains some percentage (say 5%) of the noise. If we now add back a percentage of the original data, what do we have? After a typical 50% add-back, the output will have a signal level of 140 (90+50=140) and a noise level of 55 (5+50=55), giving us a signal-to-noise ratio of 140/55.
One obvious concern with this technique is that the output signal is 140 instead of close to the original 100. At 140, the signal is not preserved and is actually distorted. Another technique sometimes used is to add back a percentage of the modelled noise (Fig. 2). Remember that even if our signal model is quite accurate again, let us say it is 90% accurate the modelled noise C will still contain some signal. If we add back 50% of the modelled noise, the output will have a signal level of 95 (90 + 50%x10 = 95) and a noise level of 52.5 (5 + 50%x95 = 52.5), giving a signal-to-noise ratio of 95/52.5. This is not as high as 140/55, but at least the technique has not altered the signal. Neither result is satisfactory. A different approach is required that will both preserve the signal and deliver improved signal-to-noise ratios.
A fundamental shift
Most approaches to improving noise attenuation algorithms focus on incremental improvements in the signal model description. However, in view of the fact that the model can never describe the signal perfectly, what is needed is a fundamental shift in the approach to handling the estimated signal and noise models. Rather than simply modelling the signal, or
Figure 2 Conventional noise attenuation scheme with 50% of modelled noise add-back.
Figure 3 The LIFT approach. This is a very simplified schematic to illustrate the method.
40
2004 EAGE
technical article
even adding back a percentage of the input, LIFT estimates the residual signal removed during the signal modelling and incorporates this in the data reconstruction. Figure 3 illustrates the concept. Note that compared to conventional noise attenuation add-back results, LIFT delivers a better signal-tonoise ratio, and better preserves signal. The whole approach can be, and often is, iterative.
Figure 5 LIFT technique for random noise attenuation compared to traditional approach, as illustrated by FXDN example. Note that FX Decon, a popular method for random noise attenuation, actually underperforms and mixes the data. The LIFT result has a higher S/N ratio, does not suffer from smearing, and handles individual noisy traces much more effectively.
2004 EAGE
41
technical article
Attenuating mutiples
The methodology to attenuate multiples (Fig.8) is to perform a de-multiple algorithm of choice usually Radon de-multiple or perhaps even SRME. This is followed by LIFT to attenuate residual multiple energy. In the LIFT sequence, signal can be modelled in a variety of ways, for example, by Zoeppritz-based AVO inversion. Due to the assumptions of Zoeppritzs equations, amplitude has to vary with offset in a smooth well-behaved fashion. However, when a multiple interferes with a primary, the amplitude does not vary smoothly. A Zoeppritz-based AVO inversion will consider multiples and converted waves to be noise and will exclude them in its reconstruction of signal. In this way, primary signal is modelled. From there, the sequence then adds back an estimate of the signal removed during the signal modelling, attenuating residual multiple energy from the original data in an adaptive non-linear fashion. The LIFT technique for multiple attenuation has been found to work well for both land and marine data. Figure 9 shows an example of multiple attenuation in offshore data. From left to right are displays of stacks of the input data, the data after Radon de-multiple (to attenuate the water-bottom multiple), and after LIFT. Because the sea-floor has rugged, rapidly changing topography, the time-offset curve of the water-bottom multiple reverberation is not hyperbolic. A Radon transform typically assumes that a multiple exhibits a parabolic or hyperbolic tx curve, and, in addition, assumes that the apex of the curve is at zero offset. Most standard Radon methods cannot handle situations where the multiple reflections do not honour these assumptions. In addition, Radon methods are aperture-limited. As noted by Wang (2003), the limited spatial aperture causes edge effects, impairing the separation of primary and multiple reflections in the Radon transform domain (the events do not transform as they should to points in p space, rather, they transform to smeared interfering lines). The limited size of the spatial aperture of a seismic gather affects the ability of the Radon transform to separate multiple and primary reflections.
Figure 9 Stacks of input gathers, Radon output, and LIFT output, and the Difference QC display. In this marine example, because the sea-floor has rapidly changing topography, the traveltime curves of the water-bottom multiples are nonhyperbolic and the time apex of the curve varies with offset. LIFT can attenuate residual multiple energy that conventional methods, such as Radon in this case, could not handle. Note also that diffraction energy desirable for subsequent prestack migration is clearly preserved through the LIFT process. with. Historical approaches such as spatial filters perform reasonably well but the results suffer from some spatial smearing, and signal amplitudes are not necessarily preserved. The LIFT approach has reduced the noise without smearing, distorting (or creating) signal. The approach can be iterative and parameters can be fine-tuned for particular data sets.
42
2004 EAGE
technical article
Reliable multiple removal techniques are needed as part of the premigration process. (O'Brien and Gray 1996). Using LIFT to precondition the data prior to PSTM prevents expensive re-runs of prestack migration. Figure 11 illustrates the importance of preconditioning even in cases of more subtle stratigraphic plays. Figure 11 shows data from the Western Canada Sedimentary Basin. In
Figure 9 illustrates that LIFT can attenuate residual multiple energy very well. The rightmost panel is the difference display (the difference between Radon and LIFT), showing that what LIFT has removed is almost entirely noise, there is very little primary leakage. The sequence is very effective in attenuating residual multiple energy that Radon cannot handle. The question sometimes arises whether this multiple attenuation scheme will preserve diffractions for prestack migration. It can be clearly seen that diffraction patterns are still present after the LIFT process. In fact, originally the noise was on top of the diffraction energy, making it difficult to identify the diffractions.
Figure 10 The signal preservation capabilities of LIFT make it the noise attenuation method of choice for data before prestack migration.
Figure 11 Even in subtle stratigraphic plays, preconditioning with LIFT can be worthwhile, in order to prevent migration artifacts which may be misinterpreted as faults. In this example, a 2D line from the WCSB, the effects of LIFT can be seen by comparing the structure stack of CMP gathers without LIFT (top left) and the structure stack after LIFT was run on the gathers to attenuate random noise (top right). LIFT is a prestack process. The gathers without LIFT were input to PSTM (bottom left). The gathers after LIFT were input to a separate PSTM (bottom right).
2004 EAGE
43
technical article
such situations, migration artifacts can easily be misinterpreted as faults. Preconditioning prior to PSTM effectively prevents this pitfall.
Conclusions
LIFT is an approach to noise attenuation that is significantly different from historical approaches. The technique offers a way of attenuating noise while better preserving signal. As well as being used to attenuate random and coherent noise, including difficult multiples that other methods could not handle, it has proved to be an excellent technique for preconditioning data for prestack migration, thus preventing migration artifacts and saving costly re-runs.
Acknowledgements
The authors thank Valery Miroshnikov for his assistance.
Canadian Society of Exploration Geophysicists/Canadian Society of Petroleum Geologists Joint National Convention, Calgary, Canada, Expanded Abstracts. Fatti, J.L., Smith, G.C., Vail, P.J., Strauss, P.J. and Levitt, P.R. [1994] Detection of gas in sandstone reservoirs using AVO analysis: A 3-D seismic case history using the Geostack technique. Geophysics 59, 13621376. Hugonnet, P. and Canada, G. [1995] Aliasing in the parabolic Radon transform. 65th SEG Meeting, Houston, USA, Expanded Abstracts, 13661369. O'Brien, M.J. and Gray, S.J. [1996] Can we image beneath salt? The Leading Edge 15, 1722. Sacchi, M.D. and Ulrych, T.J. [1995] Model re-weighted least-squares Radon operators. 65th SEG Meeting, Houston, USA, Expanded Abstracts, 616618. Spitz, S. [1991] Seismic trace interpolation in the F-X domain. Geophysics 56, 785794. Wang, Y. [2003] Multiple attenuation: coping with the spatial truncation effect in the Radon transform domain. Geophysical Prospecting 51, 7587.
44