next up previous print clean
Next: Automatic FEAVO detection Up: Vlad: Focusing-effect AVO Previous: Introduction

Extracting the image perturbation

WEMVA proceeds in three steps. The first one consists of prestack depth-migrating the data with a starting velocity model, then transforming the image to the angle domain. The second step starts by improving the image so that it is closer to the result of a migration with the correct velocity. The improvement may be performed by straightening the gathers or by smoothing the amplitudes along already flat gathers. The improved image is subtracted from the original one to create an image perturbation. The third step is inverting the image perturbation into a slowness perturbation.

The WEMVA flow presented above is actually the one for the procedure named Target Image Fitting (TIF) WEMVA. In a more radical inversion-theory approach, called Differential Semblance Optimization (DSO) WEMVA, steps 2 and 3 are combined into a single step, with the image improvement being performed by a weight operator during the inversion. In theory, DSO WEMVA will not exhibit the same Born-related problems as TIF WEMVA Sava and Symes (2002). However, it assumes that the quality of the image-improvement operator that is embedded as a weight in the inversion can be trusted. I am currently just trying to develop such an image-improvement operator specific to the FEAVO problem, and I need to be able to isolate its effects, so for the time being I will use TIF WEMVA on velocity anomalies inside the Born approximation.

In the context of the FEAVO problem, the image perturbation is the difference between the FEAVO-free image and the original one. I therefore need to eliminate the FEAVO effects from the image. To do so, I have to discriminate between AVO due to focusing and AVO due to actual changes in reflectivity with the incidence angle. I will proceed by first obtaining an estimate of the AVO due to the properties of the reflector.

The variation of amplitudes due to changes in reflectivity with incidence angles lower than $30^\circ$ has been modeled by Shuey (1985) as follows:  
 \begin{displaymath}
R\left(\theta\right)=I+G\sin^2\left(\theta\right),\end{displaymath} (1)
where I and G are scalars depending only on the physical properties of the materials above and below the reflecting interface. If the amplitudes are picked at a single midpoint-depth location with no FEAVO or illumination problems, and they are plotted as a function of the squared sine of the incidence angle, the values will arrange close to a line with intercept I and gradient G. The presence of FEAVO causes the linear dependence to break, as exemplified in Figure 1. The presence or absence of a linear trend in the ($\sin^2$,amplitude) space is therefore a FEAVO-discriminating criterion, which can be used to eliminate FEAVO. I will have to account also for the superposition of FEAVO and reflector-caused AVO.

 
examine_FEAVO
examine_FEAVO
Figure 1
Top panel: Midpoint-angle depth slice from the prestack migrated synthetic dataset shown in Figure 1 of Vlad et al. (2003). Bottom panels: Amplitudes at midpoints marked by vertical thin lines in the upper panel.
view

At each midpoint-depth location, I will compute the most plausible reflector-caused AVO as a linear trend in the ($\sin^2$,amplitude) space, and I will obtain FEAVO by subtracting this trend from the image.

The challenge is now fitting to the data at each midpoint-depth location a linear trend in the ($\sin^2$,amplitude) space. The trend should be as close as possible to that of the reflector-caused AVO, even when focusing effects are present. I will consider only the angles up to $30^\circ$. After that, the accuracy of the angle-gather amplitudes decreases, unless the offsets are extremely densely sampled, which would increase the costs of the migrations in WEMVA too much. Also, after $30^\circ$,the two-term Shuey approximation of reflector-caused AVO stops working. I have experimented with his three-term formula for larger angles. The extra degree of freedom allowed for a curve that fitted everything, including the FEAVO effects at angles lower than $30^\circ$. I have therefore to decided to use only the two-term formula on angles under $30^\circ$.

The best way to fit the linear trend would be to formulate the fit as a 2-D inverse problem, penalizing nonphysical variations of AVO along the midpoint or along the reflector. While I intend to do so in the future, I experimented for the time being with a 1-D approach mapped in Figure 2.

 
extract1D_flow
extract1D_flow
Figure 2
1-D FEAVO extraction flowchart.
view

At each (midpoint, depth) location I fitted by least-squares a straight line through the [reflectivity, $\sin^2$(angle)] pairs. I then subtracted the linear trend and applied a low-cut filter along angle. A depth slice from the results of the FEAVO extraction, together with the corresponding true FEAVO anomalies, is shown in Figure 3.

 
extractFEAVO
extractFEAVO
Figure 3
Top panel: Depth slice from the optimal angle-domain image perturbation in the top panel of Figure 6 of Vlad et al. (2003)). Bottom panel: FEAVO anomalies extracted by applying the workflow in Figure 2 to the image built by migrating with a constant velocity the dataset in Figure 1 of Vlad et al. (2003).
view

The extracted anomalies are closer in morphology to the true FEAVO in the optimal image perturbation. Also, the extraction worked up to $30^\circ$, in contrast with the previous image-processing-based technique proposed by Vlad et al. (2003), which only worked up to $21^\circ$.

A 2-D extraction procedure in which the intercept and gradient would vary more smoothly along the midpoint would make it possible to eliminate in the future the low-pass filter step, which is responsible for the reverse-polarity ``borders'' around the extracted anomalies. However, the results of the 1-D extracting procedure may still provide some valuable insight into the effectiveness of FEAVO extracting procedure, given that the ``unfair'' advantage of the lack of noise is compensated by the simplicity of the 1-D extraction.

The slowness update produced by inverting the extracted image perturbation is displayed in Figure 4.

 
slow_proj
slow_proj
Figure 4
Slowness backprojection of the image perturbation using one nonlinear WEMVA iteration, with an unregularized solver.
view

The slowness update produced this way puts anomalies with the correct polarities in the correct places. The shape of the anomalies could nevertheless be improved with the use of a 2-D FEAVO extraction procedure.

The method presented in this paper discriminates FEAVO on the basis of a local property--nonlinear variation with the squared sine of the angle. FEAVO anomalies, however, are correlated in tent-like shapes throughout the prestack volume, as shown in Figures 3, 4 and 9 in Vlad (2002). That paper proposed a ``focus-filter-spread'' approach to discriminate FEAVO from uncorrelated amplitude effects. The shapes of the anomaly surfaces in the prestack volume can be precomputed as functions of the known velocity. Summation along them would highlight the anomalies; only the ``bright stars'' would be kept. Then, the de-noised Radon-space image would be spread along the surfaces, to keep only the FEAVO anomalies in the image perturbation. Vlad et al. (2003) identified a major difficulty with this approach: the alternating polarities in the wavelet would lead to summed values canceling each other. The obvious solution--taking the unsigned values--would, however, lead to the inability to discriminate between positive and negative velocity anomalies, and to the inversion process never converging if the data contains FEAVO-generating velocity anomalies of more than one sign.

The conundrum of FEAVO extraction based on its global properties can however be broken by using a procedure that I present below. Instead of a ``focus-filter-spread'' approach, I will use a ``discriminate-focus-filter-mask'' one. I will first process the prestack volume through a FEAVO discriminator based on the local properties of the anomalies. Then, as in the ``focus-filter-spread'' approach, I will square or take the modulus of the values in the prestack volume, sum along precomputed velocity-dependent surfaces, filter to keep only the ``bright stars'' in the Radon domain, then spread back. But instead of feeding the result of spreading into the inversion, I will use it as a mask that I will apply to the result of local property-based FEAVO discrimination. That will filter out all areas with amplitude variations that are nonlinear, but are not FEAVO (because they do not correlate vertically and laterally in the known tent-shape). Since the synthetic dataset used in this paper does not have non-FEAVO amplitude departures from the Shuey (1985) model, it represents a good benchmark for the accuracy of the inversion, even in the absence of the de-noising ``focus-filter-mask'' step.


next up previous print clean
Next: Automatic FEAVO detection Up: Vlad: Focusing-effect AVO Previous: Introduction
Stanford Exploration Project
5/23/2004