You are on page 1of 5

Downloaded 09/03/13 to 202.68.179.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.

org/

AVO inversion with t-distribution as priori constraint


Qichao Zhou *, Xingyao Yin, Zhaoyun Zong and Hanqing Liu China University of Petroleum(Huadong) Summary Prior information plays an important role in Bayesian AVO inversion. Conventionally, the prior distribution in the inversion focuses on the Gaussian distribution, Huber distribution, Cauchy or the improved Cauchy distribution. However, not all the parameters in field data belong to one of the above distributions. After statistical analysis with wells data of different regions, we found that inverse parameters were generally consistent with t-distribution, while errors were usually shown in Cauchy and Gaussian distribution. In view of this, Bayesian inversion algorithm based on tdistribution as priori constraint was proposed in this paper. The algorithm can adapt to the areas of different parameter distribution by choosing different degrees of freedom. And, improving the matching of the priori information can increase the credibility of the posterior function and so ensure a better inversion. Model test and real data example verify the stability and feasibility of this proposed method. Introduction AVO inversion in Bayesian scheme is helpful to improve the estimation precision of P-wave velocity, S-wave velocity and density. The AVO inversion in Bayesian scheme combines the likelihood function with priori geological information firstly, and then solves the maximum of the posterior to create the objective function. The regularization of priori information can be used to solve "pathological" problem, existing in the three-parameter simultaneous inversion, and the building of likelihood function can reflect the relationship of the inversion target and the seismic data. But when it comes to how to choose priori information and likelihood function, there is no uniform assumption and many scholars have done much research about this. Mallick and Chen (1994) introduced Bayesian framework to AVO data inversion, and realized AVO data waveform inversion using genetic algorithms. Buland and Omre (2003) regarded elastic parameters as Gaussian distribution by statistical analysis, and obtained better results. Downton and Line (2001) tested the AVO sparse pulse waveform inversion on the basis of pre-NMO, found that Cauchy priori has higher resolution than Gaussian priori. Hampson et al. (2005) learned post-stack impedance inversion,and gave the prestack AVO three-parameter inversion formula. Chen and Yin (2007) based on Bayesian theory established constrained inversion about logging data, and improved the inversion stability. Yang and Yin (2008) verified the effect of the improved Cauthy distribution, and proposed a pre-stack inversion method based on nonlinear quadratic programming. Wubshet (2010) studied an inversion method based on multivariate Cauchy distribution as priori function, which not only preserved weak reflection information through the long tail property of the Cauchy distribution but diminished the statistical correlation between the parameters. On the basis of previous studies, we implemented threeparameter inversion in Bayesian scheme by using tdistribution function as priori constraint. Numerical test and real data example demonstrate that it can provide more reliable estimation for elastic parameters. 1 t distribution In order to improve the accuracy of selected priori functions and increase the credibility of the posterior information, we carry on statistical analysis of wells logging data in fields. We find that in most regions the formation parameters agree more with t-distribution, and have a large difference from Gaussian and Cauchy distribution. Therefore, we can say that t-distribution is more reasonable as a priori function (Figure 1). The probability density function of t-distribution is,

f (t ) =

((k + 1) 2) t2 (1 + ) ( k +1) 2 k k (k 2)

(1)

Usually, the value of freedom degree k is larger than 1. However, in order to enhance the resolution of special circumstances, we can also set k under 1. When k is 1, tdistribution becomes Cauchy distribution, and when k is + , it becomes Gaussian distribution. As displayed in Figure 2, the Gaussian and Cauchy distribution are just two special cases of the t-distribution, which means t-distribution has broader applicability.

2013 SEG SEG Houston 2013 Annual Meeting

DOI http://dx.doi.org/10.1190/segam2013-0497.1 Page 3252

AVO inversion with t-distribution as priori constraint


where p (d | m) is the likelihood function, n is the noise variance, and G is wavelet matrix. Supposing parameters m to be t distribution:
2

Downloaded 09/03/13 to 202.68.179.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

p (m ) =

where m is the standard deviation of model parameters. Posterior probability distribution can be expressed as follows,
p= 1
2 (2n )

k + 1 ) 2 1 1 k ( ) ( )k 2 2 2 (

(1 +
m

(m )
2 m

k +1 2

(4)

N2

k +1 ) k +1 mT m (d Gm)T (d Gm) 2 (1+ 2 ) 2 exp[ ] 1 2 2 n k m 1 k 2 ( )( )k m 2 2 (

(5)

1 ( 2
2 n

)N

Figure 1: inverse parameter maps in wells

Where, Maximizing equation (5), yields,

k +1 ) 2 1 1 k ( ) ( )k 2 2 2 (

2 2 1 T m= GT G + (k +1)n /(km + mi2 )I Gd

(6)

Likelihood function t distribution Posterior objective Model constraint New objective parameters inversion Figure 2: Maps of the different degrees of freedom of the t distribution and Gaussian distribution, Cauchy distribution 2 Methods and principles The proposed inversion workflow is shown in Figure 3. The posterior probability distribution of the parameters m is: Elastic parameters Figure 3: The proposed Bayesian inversion workflow with t distribution as prior information Model constraints are added to initial target function to enhance the stability of inversion. Model constraint matrix can be formulated as (7) = Pm Where is well data related to model parameter, and P is integral operator matrix. Model constraint is added to the objective function, and the final objective function becomes, F(m) = FG (m) + Ft (m) + FEI (m)

p(m| x) =

p(m) p(x | m)

p(x | m)dm

p(m) p(x | m)

(2)

Where, m is the parameters to be estimated, x is the observed data, p (m | x) is a posteriori probability, p (m) is a priori probability, and p ( x | m) is the likelihood function. The Bayesian estimation aims to obtain model parameters m from observing data with noise. Background noise of seismic data is assumed to be Gaussian noise. And the likelihood function is: 1 ( d Gm )T ( d Gm ) exp[ ] (3) p (d m) = 2 N 2 2 ( 2 n ) 2 n

= (d Gm )T (d Gm ) + a ln( 1+ + b ( Pm)T ( Pm)

mT m ) 2 km

(8)

2013 SEG SEG Houston 2013 Annual Meeting

DOI http://dx.doi.org/10.1190/segam2013-0497.1 Page 3253

AVO inversion with t-distribution as priori constraint


2 a = (k + 1) n , b

Downloaded 09/03/13 to 202.68.179.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Where,

is the constraint coefficient,

which is used to control the accuracy and stability of the inversion. Solving equation (8), yield,
2 2 m = (GT G + (k + 1) n /(k m + mi2 ) I + PT P) 1 (GT d + PT )

(9)

3 The choice of freedom degree From equation (9), we can see that the freedom degree k is the main factor in the objective function. For the general area, the value of freedom degree can be obtained via extracting parameter from well data with curve fitting. For the regions without well data, we can extract a seismic trace and to determine the degree of freedom in an inversion scheme. 4 Model test With a Ricker wavelet and reflection coefficients, the synthetic traces are displayed in Figure 4.
0.09 0.08 0.07 0.06 0.05 100 0.04 80 0.03 0.02 0.01 0 -1 60 40 20 0 -2 80 60 40 20 0 -2 80 60 40 20 0 -2 100 100 200 180 160 140 120 200 180 160 140 120 200 180 160 140 120

Figure 5: The inversion results without noise with Cauchy, Gaussian and t distribution as prior information

Figure 6: The inversion results with S/N=2 with Cauchy, Gaussian and t distribution as prior information From Figures 5 and figure 6, we can see that appropriately reducing freedom degree can enhance the resolution, and increasing freedom degree can improve the stability. By choosing different freedom degrees, t distribution can find a balance between the resolution and stability. 5 Real data example The common mid-point traces near a well in field are utilized to test the feasibility of the proposed method in this paper. Figure 7 to Figure 9 display the comparison of the inversion results of P-wave velocity, S-wave velocity and density with Cauchy, t distribution and Gaussian distribution as prior constraints, respectively. It can be seen that, the inversion results from t distribution as prior information are more reliable and accurate.

(a) (b) (c) (d) Figure 4: (a) Richer wavelet, (b)reflection coefficient, (c) the synthetic data without noise and (d) the synthetic data with S/N=2 Figure 5 displays the inversion results without noise with Cauchy, Gaussian and t distribution as prior information, respectively. We can see that, the three methods can identify reflective layers. However, the inversion results with t distribution and Cauchy distribution as prior information own higher resolution. Therefore its better to choose Cauchy distribution or t-distribution whose freedom degree is small. Figure 6 displays the inversion results with S/N=2 with Cauchy, Gaussian and t distribution as prior information, respectively. It can be seen that the inversion result of Gaussian distribution can roughly identify reflective layers with greater volatility and lower accuracy. The result of Cauchy distribution is improved compared with Gaussian. However, it will produce false reflections. And the result of t distribution shows higher resolution and better stability.

2013 SEG SEG Houston 2013 Annual Meeting

DOI http://dx.doi.org/10.1190/segam2013-0497.1 Page 3254

AVO inversion with t-distribution as priori constraint

Downloaded 09/03/13 to 202.68.179.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

(a)

(b)

(c) (a) (b) (c)

Figure 7: Comparison of true p-wave velocity (in red) and estimated p-wave velocity (in black) with (a) Cauthy distribution (b) t distribution (c) Gaussian distribution

Figure 9: Comparison of true density (in red) and estimated density (in black) with (a) Cauthy distribution (b) t distribution (c) Gaussian distribution 6 Conclusion T distribution as a priori function shows more accordance with the distribution of actual formation parameters, and it has better inversion results. Curve fitting of logging data should be done to obtain the initial degree of freedom. If logging data is unavailable the degree of freedom could be selected according to the complexity of the region. When the initial inversion results are quite different from the model, the value of the freedom degree should be modified significantly. When the results slightly differ from the model, we should decrease the step length of modifying freedom degree.

(a)

(b)

(c) Acknowledgments We would like to acknowledge the sponsorship of the National 973 Program of China (2013CB228604), the National Grand Project for Science and Technology (2011ZX05030-004-002, 2011ZX05019-003 and 2011ZX05006-002) and Foundation from geophysical key lab of Sinopec for funding this research.

Figure 8: Comparison of true s-wave velocity (in red) and estimated s-wave velocity (in black) with (a) Cauthy distribution (b) t distribution (c) Gaussian distribution

2013 SEG SEG Houston 2013 Annual Meeting

DOI http://dx.doi.org/10.1190/segam2013-0497.1 Page 3255

http://dx.doi.org/10.1190/segam2013-0497.1 EDITED REFERENCES Note: This reference list is a copy-edited version of the reference list submitted by the author. Reference lists for the 2013 SEG Technical Program Expanded Abstracts have been copy edited so that references provided with the online metadata for each paper will achieve a high degree of linking to cited sources that appear on the Web. REFERENCES

Downloaded 09/03/13 to 202.68.179.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Chen, J., and X. Yin, 2006, Seismic parameter estimation from AVO waveform inversion: SPG/SEG International Geophysical Conference. Downton, J. E., 2001, Constrained three parameter AVO inversion and uncertainty analysis: CSEG. Hampson, D. P., B. H. Russell, and B. P. Bankhead, 2005, Simultaneous inversion of prestack seismic data: Presented at the 75th Annual International Meeting, SEG. Mallick, S., 1995, A comparison of poststack and prestack inversion of seismic data: 65th Annual International Meeting, Expanded Abstracts, SEG, 651654. Smith, G. C., and P. M. Gidlow, 1987, Weighted stacking for rock property estimation and detection of GAS: Geophysical Prospecting, 35 , no. 9, 9931014, http://dx.doi.org/10.1111/j.13652478.1987.tb00856.x. Tarantola A, 2005, Inverse problem theory and methods for model parameter estimation: SIAM. Wubshet, M. A., 2010, Regularization of the AVO inversion problem by means of a multivariate cauchy probability distribution, Alberta: University of Alberta. Yang, P., and X. Yin, 2008, Nonlinear quadratic programming Bayesian prestack inversion: Chinese Journal of Geophysics, 51 , 18761882 (in Chinese).

2013 SEG SEG Houston 2013 Annual Meeting

DOI http://dx.doi.org/10.1190/segam2013-0497.1 Page 3256

You might also like