Sei sulla pagina 1di 5

Potential risks when using seismic attributes as predictors of reservoir properties

CYNTHIA T. KALKOMEY, Mobil E&P Technical Center, Dallas, Texas


he advance of our ability to generate seismic attributes and the growing emphasis on production geophysics has led to the widespread use of seismic attributes as predictors of reservoir properties. In many cases, we can show using seismic modeling or rock physics a physically justifiable relationship between a seismic attribute and the reservoir property of interest. When this is true, we are able to greatly reduce the uncertainty of interwell predictions of reservoir properties. The simple flow diagram in Figure 1 illustrates the basic process. The first critically important step is accurately tying the well and seismic both vertically and areally. We then must choose which seismic attributes we believe to be related to the reservoir property. Using that attribute or set of attributes, we use the dense seismic data to guide the prediction of reservoir properties between sparse well data. A number of methods can be used for the prediction step linear or nonlinear regression, geostatistics, or neural networks. The purpose is estimating in-place hydrocarbon volumes or making reservoir management decisions such as location and number of wells, depletion strategy, or gas and water injection operations. The problem with the process is: How do we identify which seismic attributes to use? The problem. All of the prediction methods (regression, geostatistics, and neural networks) require making inference from seismic data at a small number of wells to a larger population which we assume is rep-

resented by that sample. Enthusiastic practitioners can now generate 10s even 100s of seismic attributes. An all too common practice is identifying which seismic attributes to use solely on the strength of their observed correlations with reservoir properties measured at the wells. Often these correlations are estimated by a very small number of observations. As with any parameter, when the sample size is small, the uncertainty about the value of the true correlation can be quite large. To get an idea of how large this uncertainty can be, consider a sample correlation (r) of 0.80 calculated from 10 data points. The 95% confidence interval estimate of the true correlation is [0.34, 0.95]. The only statement we can make with (95%) confidence is that the true correlation between the reservoir property and seismic attribute is between 0.34 and 0.95. If we have only five data points instead of 10, the 95% confidence interval widens to [-0.28, 0.99]. Since this interval contains zero, we cannot say with confidence that there is any correlation between the reservoir property and the seismic attribute. Most practitioners are already aware that the smaller the sample, the greater the uncertainty about the true value of the correlation. A lesser known phenomena is something statisticians call experiment-wise error rates. This just means that as we generate an ever increasing number of attributes, the greater the chance of observing at least one spuriously large sample correlation. You can get the idea of this if you consider an experiment where you draw a small number of observations from two completely unrelated variables say the annual birthrate of elephants and pounds consumed of a food substance. If we randomly select five years and observe the birthrate of elephants and pounds of sugar consumed, we would probably correctly infer that there is no relationship between the two variables. But if we continue to test this for all possible food types, by chance alone we will likely find at least one food type that appears, on the basis of the sample correlation, to be related to the birthrate of elephants. When this happens, we say weve encountered a spurious correlation. Definition: A spurious correlation is a sample correlation that is large in absolute value purely by chance.

Figure 1. Simple flow diagram showing use of seismic attributes to predict reservoir properties.

For seismic attributes and reservoir properties this means we have happened to sample at locations that yield a large sample correlation when in truth they are

MARCH 1997

THE LEADING EDGE

247

Table 1. Probability of spurious correlation between a reservoir property and a single seismic attribute

Sample size
R 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 5 0.87 0.75 0.62 0.50 0.39 0.28 0.19 0.10 0.04 10 0.78 0.58 0.40 0.25 0.14 0.07 0.02 0.01 0.00 15 0.72 0.47 0.28 0.14 0.06 0.02 0.00 0.00 0.00 20 0.67 0.40 0.20 0.08 0.02 0.01 0.00 0.00 0.00 25 0.63 0.34 0.15 0.05 0.01 0.00 0.00 0.00 0.00 35 0.57 0.25 0.08 0.02 0.00 0.00 0.00 0.00 0.00 50 0.49 0.16 0.03 0.00 0.00 0.00 0.00 0.00 0.00 75 0.39 0.09 0.01 0.00 0.00 0.00 0.00 0.00 0.00 100 0.32 0.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00

actually uncorrelated. How likely is it that this could occur? Probability of observing a spurious correlation. Consider a single seismic attribute as a possible predictor. The probability of observing the absolute value of the sample correlation (r) greater than some constant (R) given the true correlation (r) is zero can be found from
R n-2 p sc = Pr r R = Pr t 1- R2

where n is the sample size or number of locations with

measurements of both the reservoir property and the seismic attribute, and t is distributed as a Students t variate with n-2 degrees of freedom. Note that the probability of a spurious sample correlation depends only on R (the magnitude of the spurious sample correlation) and n (the number of well measurements). An assumption of this relationship is that the n observations are drawn randomly. The reason for a random sample is to insure that the sample is representative of the population under study and that the amount of independent information available to estimate the true correlation is maximized. If in fact the data are spatially correlated, the relationship gives a conservative estimate of the probability of a spurious correlation. This is because the effective sample size will be smaller than the actual sample size; and as n decreases, the probability of a spurious correlation increases. Table 1 shows the probability of observing a spurious correlation for different magnitudes of the sample correlation (R) and different sample sizes (n). Now that weve quantified the probability of a spurious correlation when considering a single seismic attribute, we can calculate the probability of at least one spurious correlation when considering a set of k independent attributes. The probability of observing at least one spurious correlation when considering a set of k independent attributes is simply 1 minus the probability that none of the sample correlations are spurious. This can be expanded to the summation

Tables 2-5. Probability of observing at least one spurious sample correlation from a given number of independent seismic attributes.
Table 2. Five independent seismic attributes Table 3. Ten independent seismic attributes

Sample size
R 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 5 1.00 1.00 0.99 0.97 0.92 0.81 0.65 0.42 0.17 10 1.00 0.99 0.92 0.77 0.53 0.29 0.12 0.03 0.00 15 1.00 0.96 0.80 0.53 0.26 0.09 0.02 0.00 0.00 20 1.00 0.92 0.67 0.34 0.12 0.03 0.00 0.00 0.00 25 0.99 0.87 0.54 0.22 0.05 0.01 0.00 0.00 0.00 35 0.98 0.76 0.34 0.08 0.01 0.00 0.00 0.00 0.00 50 0.97 0.59 0.16 0.02 0.00 0.00 0.00 0.00 0.00 75 0.92 0.36 0.04 0.00 0.00 0.00 0.00 0.00 0.00 100 0.86 0.21 0.01 0.00 0.00 0.00 0.00 0.00 0.00 R 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 5 1.00 1.00 1.00 1.00 0.99 0.96 0.88 0.67 0.32 10 1.00 1.00 0.99 0.95 0.78 0.50 0.22 0.05 0.00 15 1.00 1.00 0.96 0.78 0.45 0.17 0.04 0.00 0.00

Sample size
20 1.00 0.99 0.89 0.57 0.22 0.05 0.01 0.00 0.00 25 1.00 0.98 0.79 0.39 0.10 0.02 0.00 0.00 0.00 35 1.00 0.94 0.57 0.16 0.02 0.00 0.00 0.00 0.00 50 1.00 0.83 0.29 0.04 0.00 0.00 0.00 0.00 0.00 75 0.99 0.59 0.09 0.00 0.00 0.00 0.00 0.00 0.00 100 0.98 0.38 0.02 0.00 0.00 0.00 0.00 0.00 0.00

Table 4. Twenty independent seismic attributes

Table 5. Forty independent seismic attributes

Sample size
R 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 5 1.00 1.00 1.00 1.00 1.00 1.00 0.98 0.89 0.53 10 1.00 1.00 1.00 1.00 0.95 0.75 0.39 0.10 0.01 15 1.00 1.00 1.00 0.95 0.70 0.31 0.07 0.01 0.00 20 1.00 1.00 0.99 0.81 0.39 0.10 0.01 0.00 0.00 25 1.00 1.00 0.96 0.62 0.20 0.03 0.00 0.00 0.00 35 1.00 1.00 0.81 0.29 0.04 0.00 0.00 0.00 0.00 50 1.00 0.97 0.50 0.08 0.00 0.00 0.00 0.00 0.00 75 1.00 0.83 0.16 0.01 0.00 0.00 0.00 0.00 0.00 100 1.00 0.61 0.05 0.00 0.00 0.00 0.00 0.00 0.00 R 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 5 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.99 0.78 10 1.00 1.00 1.00 1.00 1.00 0.94 0.62 0.20 0.02 15 1.00 1.00 1.00 1.00 0.91 0.52 0.14 0.01 0.00

Sample size
20 1.00 1.00 1.00 0.97 0.63 0.19 0.02 0.00 0.00 25 1.00 1.00 1.00 0.86 0.36 0.06 0.00 0.00 0.00 35 1.00 1.00 0.95 0.50 0.09 0.01 0.00 0.00 0.00 50 1.00 1.00 0.75 0.15 0.01 0.00 0.00 0.00 0.00 75 1.00 0.97 0.30 0.01 0.00 0.00 0.00 0.00 0.00 100 1.00 0.85 0.09 0.00 0.00 0.00 0.00 0.00 0.00

248

THE LEADING EDGE

MARCH 1997

a)

b)

Figure 2. Probability of observing at least one spurious sample correlation of magnitude (a) 0.6, and (b) 0.4 when actually the well data and seismic attributes are uncorrelated.

Figure 3. Possible outcomes when selecting a seismic attribute as a predictor of a reservoir property.

1 - (1 - p sc ) = p sc (1 - p sc )
k i =1

( i -1)

where psc is the probability of a spurious correlation for a single attribute and k is the number of independent seismic attributes considered. From the summation we can see that the penalty for increasing the number of attributes considered from k-1 to k is
p sc (1 - p sc )
( k -1)

spurious correlation decreases. To avoid this, one should use the table where k equals the number of independent linear combinations of the set of seismic attributes considered. These tables can be used to assess the risk of selecting a seismic attribute that is actually uncorrelated with the reservoir property being predicted. For example, look at the column headed 5 and row labeled 0.60 of Table 3. We see that, given we have only five well measurements, there is a 96% chance of observing a sample correlation coefficient of 0.60 for at least one of the 10 seismic attributes considered, when no correlation actually exists between the attributes and the property. If we double the number of wells to 10, the probability of at least one spurious correlation of magnitude 0.6 or more drops to 0.50. What happens if we increase the number of independent seismic attributes considered from 10 to 20? Table 4 shows that the chance of observing at least one sample correlation of 0.60 or more in absolute value from a sample of size five, when no correlation actually exists between the attributes and the reservoir property, is almost 100% (its actually 0.999). Even if we double the number of wells to 10, in this case there is still a 75% chance of at least one spurious correlation. Factors influencing the probability of a spurious correlation. Figure 2 reveals the factors that influence the probability of observing spurious correlations. Figure 2a shows the probability of observing at least one sample correlation greater than or equal to 0.6 when the actual correlation is zero. Figure 2b presents the same results for an observed sample correlation of 0.4. From Figure 2, we see that the probability of observing one or more spurious correlation increases as the number of independent well measurements decreases, or the number of independent seismic attributes considered as potential predictors increases, or the absolute value of the observed sample correlation decreases. Associated risks. Statistical decision theory defines risk as the expected loss due to our uncertainty about the true state of nature under the possible decision alternatives or actions. Figure 3 illustrates the possible outcomes which must be considered in order to assess the risk

Tables 2-5 give the probability of observing a spurious sample correlation for one or more seismic attributes from a set of k = 5, 10, 20, and 40 independent attributes. Each table gives the probability of observing a sample correlation greater than or equal to 0.1 to 0.9 when the reservoir property is actually uncorrelated with the attributes, given a sample size of 5, 10, 15, 20, 25, 35, 50, 75, or 100. Note, we are assuming that the seismic attributes are independent. If in fact the attributes considered are correlated with each other, then the probabilities of observing at least one spurious correlation shown in the tables will be too large. This is because the effective number of independent seismic attributes will be smaller than the actual number of attributes considered and as k decreases, the probability of at least one

MARCH 1997

THE LEADING EDGE

249

Figure 4. Scatter plot of porosity versus seismic amplitude at the 10 well locations.

a)

b)

p o r o s i t y

Figure 5. True spatial distributions. (a) True distribution of porosity ranging from 0 (dark blue) to 20% (red). Well locations are indicated are indicated by black dots. (b) Seismic attribute over the same area.

a) seismic attribute
Figure 6. (above) Scatter plot of porosity versus seismic amplitude at all locations.

b)

Figure 8. Estimation errors. (a) Areas overestimated by more than five porosity units. (b) Areas underestimated by more than five porosity units.

Figure 7. (right) Estimate of porosity generated using collocated cokriging with the seismic attribute as the covariate.

requires additional assumptions about the true magnitude of the correlation between the well property and the seismic attribute.) To quantitatively assess risk, we would have to assign the cost or economic consequence of each of the four possible outcomes. The costs will obviously be situation dependent; however, we can generalize the cost in a qualitative sense. The cost of a Type I Error (using a seismic attribute to predict a reservoir property when actually uncorrelated) is: Inaccurate prediction biased by the attribute. Inflated confidence in the inaccurate prediction apparent prediction errors are small. The cost of a Type II Error (rejecting a seismic attribute for use in predicting a reservoir property when in fact they are truly correlated) is: Less accurate prediction than if wed used the seismic attribute. Larger prediction errors than if wed used the attribute. I believe that for most cases the economic consequences of making highly confident but inaccurate predictions are more severe than the consequences of a Type II Error. A simple example. To illustrate the consequences of a Type I Error, consider a simple example. Suppose you have measurements of porosity and a seismic attribute

associated with choosing seismic attributes to predict a reservoir properties. A Type I Error occurs if no relationship truly exists between the seismic attribute and the reservoir property of interest, yet we select the seismic attribute as a predictor. A Type II Error occurs if a physical relationship does exist between the seismic attribute and the reservoir property of interest, but we fail to use the seismic attribute as a predictor. The previous section quantified the probability of a Type I Error occurring when the selection criteria is based solely on the magnitude of the sample correlation between a seismic attribute and a reservoir property. (Calculating the probability of a Type II Error occurring

250

THE LEADING EDGE

MARCH 1997

extracted at 10 wells. Figure 4 shows a scatter plot of porosity versus the seismic attribute with a sample correlation of 0.82. It could be very tempting to proceed with the prediction using this relationship. The true spatial distribution of porosity over the area is shown in Figure 5a; Figure 5b shows a seismic attribute over the same area. If we look at the complete scatter plot over the entire area shown in Figure 6 we see a shotgun pattern indicating the true correlation of zero. By happenstance, the field has been drilled in locations which makes it appear that there is a strong relationship between this attribute and porosity when in fact there is none. We have observed a spurious sample correlation. If we use this attribute for prediction, our resulting estimate of porosity will bear a strong resemblance to the seismic attribute regardless of whether we used regression, geostatistics or neural networks. Figure 7 shows such an estimate, generated using collocated cokriging. This prediction results in large actual estimation errors. Figure 8a highlights areas where the porosity is overestimated by 5-12 porosity units. Figure 8b shows areas where the porosity is underestimated by 5 -12 porosity units. As you can imagine, drilling locations chosen on the basis of this prediction will be less than optimal. This could also be disastrous for other reservoir management decisions. For example, if an injection strategy was designed on the basis of this predic-

tion, it would likely be chosen about 90 from the optimal orientation. Conclusions. There are two main points to remember from this work. First, the probability of observing spurious sample correlations between a seismic attribute and well data can be quite large if the number of independent well measurements is small or the number of independent attributes considered is large. Secondly, when the probability of a spurious correlation is large, then selection of seismic attributes based solely on empirical evidence is risky it can lead to highly confident, but highly inaccurate predictions and thus poor business decisions. Therefore, it is strongly recommended, especially when the number of wells is small, that only those seismic attributes that have a physically justifiable relationship with the reservoir property be considered as candidates for E predictors. L
Cynthia Kalkomey received a doctorate in statistics from Southern Methodist University (1991). She joined Mobil in 1979 and is currently manager of Reservoir Characterization, Mobil Exploration and Producing Technical Center. Corresponding author: Cynthia Kalkomey, Mgr., Reservoir Characterization, Mobil Exploration and Producing Technical Center, 13777 Midway Rd., Dallas, TX 75244; phone 972-851-8598; fax 972-851-8703.

MARCH 1997

THE LEADING EDGE

251

Potrebbero piacerti anche