Geostatistica

© All Rights Reserved

62 views

Geostatistica

© All Rights Reserved

- MULTI-ECHELON SUPPLY CHAIN SIMULATION USING METAMODEL
- Refference for mineral reserves and resources
- UNDERGROUND MINING
- Manual Mine Sight Geoestadistica
- Normal Calculations Worksheet
- Formula Sheet
- exercises.pdf
- Chapter 7 Notes(1)
- 3360n
- SolidsHandouts3.pdf
- 12 Probability and Statistics
- Statistics Formula
- Mtech CS PCB 2015
- 5990-8882EN
- THE NEUTROSOPHIC STATISTICAL DISTRIBUTION, MORE PROBLEMS, MORE SOLUTIONS
- Lecture 4 Disc r Prob f 2016
- Explaining Sigma Shift v2
- Tutorial_3Q.pdf
- Quality Control
- Student 1 PPT MAT Manuscript Science Achievement in Trece Martinez Final Edited

You are on page 1of 237

and

Advanced Property Modelling

Ashley Francis

Managing Director

Earthworks Environment & Resources Ltd

http://www.sorviodvnvm.co.uk

2003 Earthworks Environment & Resources Ltd. All rights reserved.

INSTRUCTOR BIOGRAPHY

ASHLEY FRANCIS

Ashley is a geophysicist and

geostatistician whose career has

encompassed over 20 years worldwide oil industry experience of

exploration, development and

production geophysics. He has also

consulted to the nuclear and

engineering sectors on subsurface

definition and uncertainty. Since

1993 he has specialised in

geostatistics in addition to

geophysics.

Ashley has worked in or on behalf

of service companies,

consultancies and oil companies in

North and South America, Europe,

Africa, Middle East, Far East and

Australia. He spent 5 years with

LASMO plc in Technical Services

assisting and advising asset teams

worldwide in geophysics

(particularly inversion), geostatistics, risk and uncertainty.

After leaving LASMO in 2001, Ashley founded Earthworks, a consultancy

specialising in subsurface geoscience, based in Salisbury, Wiltshire, UK. His

innovative ideas are now being developed in new and ultra-fast stochastic

seismic inversion software at Earthworks.

He lectured in Borehole Geophysics to Honours Graduates at the University of

the Witwatersrand, South Africa 1989-90 and was a Visiting Research Fellow

at the Post Graduate Institute in Sedimentology, University of Reading, UK

1995-97. He continues to teach geostatistics to MSc Petroleum Geoscience

students at Imperial College, London, a role he began in 1999. His current

geostatistical course has been running successfully since 1996 and the newer

inversion course since 2000. Both courses are also run as part of the

Geoscience Training Alliance consortium in London and in Houston and are

widely acclaimed by participants. Ashley is a committee member and regular

attendee at the SEG Development and Production Forum. He was Chairman

of the 2000 meeting and is Chairman of the forthcoming 2003 meeting to be

held in July in Big Sky, Montana. He presents widely at conferences on the

subjects of geophysics and geostatistics.

Ashley is a member of SEG, EAGE, IAMG, BSSS, IPSS, IAS, SPE and

PESGB. He is a Fellow of the Royal Astronomical Society.

All rights reserved

LIST OF CONTENTS

Theory Manual

1. Statistics

2. Stochastic Theory

3. Stationarity

4. Estimation and Kriging

5. Stochastic Simulation

6. Deterministic, Stochastic and Best Estimate

7. Support

8. Entropy and Bayes Theorem

9. Cokriging

10. Introduction and Framework

11. Stratton Field Introduction

12. Scale Up of Well Logs

13. Data Analysis

14. Facies Modelling

15. Variograms in Petrel

16. Stratton Field Facies Variogram Modelling

17. Seismic Variograms

18. Facies simulation

19. Petrophysical Modelling

20. Seismic Inversion Constraints

All rights reserved

EARTH

WORKS

STATISTICS

History of Geostatistics

Geostatistics arrived with the publication of two papers by Matheron in 1962 and 1963.

Matheron was a researcher at the Centre de Geostatistique, Fontainebleau, Paris and

is regarded as the founding father of geostatistics. His papers were a mathematical

formalism of earlier empirical work by Danie Krige. Krige, who published his Masters

thesis in 1951, worked on the problem of ore grade estimation in South African gold and

uranium mining. He became interested in the discrepancy between reserve estimation

reporting and the actual recovery of minerals. He was concerned about the magnitude

of the economic decisions that were based on the reserve estimation studies. His

work began a revolution in mining reserve evaluation methods that has continued to the

present day and includes our own industry.

Pierre Carlier coined the word krigeage to honour Kriges original work, a word that was

later anglicised to kriging by Matheron in a paper published in 1963 for the benefit of

the English-speaking scientific community. Kriging is the name given in geostatistics

to a collection of generalised linear regression techniques for the estimation of spatial

phenomena.

For many years France dominated the world of geostatistics. Although still an important

research centre, other centres such as the Center for Reservoir Forecasting, Stanford,

USA and the Norwegian Computing Centre, Oslo, Norway are now achieving eminence,

particularly in hydrocarbon applications of geostatistics.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

What is Geostatistics?

Geostatistics is to statistics as geophysics is to physics. If a theoretical geophysicist

were to read a treatise on geophysics he would risk getting the disappointing

impression that it is but a collection of particular cases concerning the theories of

field potential and of wave propagation. In the same way, a specialist in mathematical

statistics would not find much more in geostatistics than variances, covariances and

optimal linear estimators.

Geostatistics is a collection of generalised linear regression techniques for the

estimation of spatial phenomena. It is not statistics applied to geological data, nor is it

seismic attribute mapping, just another gridding algorithm or building of 3D reservoir

models. Geostatistics will have application in all these tasks and more, but they do

not define it.

Geostatistics is based on fundamental assumptions concerning the behaviour of

random variables in a spatial coordinate system. At its heart is a theory of regionalised

variables that postulates that the real world is a single realisation of a random

function. By making observations we can deduce some of the properties of the random

function model and use these properties to constrain our estimates of the realisation

at unmeasured locations. Unlike traditional spatial estimation methods, geostatistics

is not ad hoc : it is founded on a clear hypotheses and assumptions about spatial

variables.

Basic Definitions

quantity which can take any number xi or a range of values with a given

probability distribution.

Cumulative Distribution Function (CDF) F(x) is defined as the function

which shows the probability of the random variable being less than (or

alternatively greater than) a specified value x. In practice we approximate

this function with the cumulative frequency.

Probability Density Function (PDF) f(x) is defined as the function which

shows the number of times one particular value xi or range of values f(x) of

the random variable occurs. In practice we approximate this function with

the histogram, a measure of frequency of occurrence within class.

Examples of a random variable are

Multiple birthdays in a company of N people

Energy

Temperature

In fact we can extend this to include any physical system which we can measure. An

alternative expression would be a random function. The independent variable is a point

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

in the sample space which describes the outcome of an experiment. In the context

of subsurface description description, the earth is the outcome of the experiment

comprising all of the natural (and man-made) processes operating on it up to that time

such as erosion, deposition, compaction and so forth.

Porosity-Thickness

Porosity-Thickness

0.80

Number of Data55

mean 5.84

std. dev. 1.36

coef. of var0.23

maximum10.23

upper quartile6.43

median 5.61

lower quartile5.10

minimum3.72

0.300

Number of Data55

mean 5.84

std. dev. 1.36

coef. of var0.23

maximum10.23

upper quartile6.43

median 5.61

lower quartile5.10

minimum3.72

0.60

0.40

0.20

0.00

0.0

4.0

8.0

phiH

12.0

ft

16.0

Frequency

Cumulative Frequency

1.00

0.200

0.100

0.000

0.0

4.0

8.0

phiH

12.0

16.0

ft

Assumptions in statistics

McBratney and Webster (1983) make the following observations concerning statistics

and geostatistics

to the population variation (a variogram will be explained below). In other

words, classical statistics assumes that the samples are independent of

each other.

Close observations of a random function will duplicate neighbouring

information to some extent due to spatial correlation

This means the effective number of samples is smaller than the actual count

of samples

Neighbouring samples minimise duplication if they are separated by a

distance greater than the range of the variogram

Univariate Statistics

Statistics is about summary description of data. Univariate statistics summarise a

single variable. It is important to distinguish between the true population parameters:

And the statistics, which are estimates of the true population parameters based on data

samples we have collected:

We call the summary parameters moments. The first moment is the mean or arithmetic

average of the data which describes the central tendency (what statisticians call the

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

location) of the distribution. The most common way to calculate the arithmetic mean

is to add up all the sample values and divide by the number of samples. We can

write this as:

The terms mean, average and mathematical expectation are synonymous. The

expectation can be denoted in a number of ways including.

The analogy between frequency of occurrence and probability allows us to make the

following definition. Let X be a random variable assuming the values x1, x2, with

corresponding probabilities p(x1), p(x2),. The mean or expected value of X is defined

by

provided that the series converges absolutely. If X is continuous rather than discrete

then we would have an integral rather than a summation

The mean is the average of all the possible outcomes or measurements. It is our Best

Guess at an unmeasured sample. For a closed data set, using the mean as a predictor

will minimise the average squared difference between each value and the mean least

squares. If we extend this idea to estimating the mean from a subset of a larger

population then the the mean estimated from the subset will minimise the squared

prediction error if the population has a Gaussian (Normal) distribution.

The arithmetic mean is not the only way of describing the central tendency of a set of

data or a distribution. Some variables such as resistivity or permeabilty do not average

arithmetically and the geometric or harmonic mean may be more appropriate.

Other measures of location include the median. The median is the value of the middle

sample if the samples are arranged (ranked) in order. Half the values are above the

median and half below the median. In the case of an even number of samples the

median is conventionally defined as the arithmetic average of the two middle values.

The median is important in terms of probability density functions (pdfs) in that it defines

the value as likely to be exceeded as not (the P50). The median is often the estimator

for central tendency used in robust statistical methods because it is not sensitive to the

magnitude of the values, only to their magnitude order or rank.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

The mode is the value that occurs most frequently. The tallest bar on a histogram

indicates where the mode is located. A data set may contain more than one mode.

However, for continuous sampled data the mode has no strict definition as its value

varies with the precision of the data values. In the case where no two samples

are exactly the same value, any sample could represent the mode. For categorical

variables, such as geological descriptors (sand, shale etc) the mode would be the most

frequently occurring class.

Permeability is the most important example in our industry of a non-arithmetic

averaging property. For effective permeability (keff) the average permeability will range

between the arithmetic mean and the harmonic mean (mh)

For angled flow, the effective permeability is close to the geometric mean (mg) :

if the probability density function (pdf) is log-normal. These examples are theoretical

end members and in practice the effective permeability usually ranges between the

arithmetic and geometric averages: for continuous layers keff is closer to the arithmetic

mean and for random permeable blocks it tends towards to the geometric mean.

Marsily (1986) gives a good description of permeability estimation.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

Declustering

The arithmetic mean is a sample estimate of the expected value. However, this is

really only true if x1, x2, is a declustered sampling of a distribution, in which case the

arithmetic mean is used as an estimate of the expected value. If we define the mean

as a weighted sum then we can decluster it by appropriate assignment of weights.

In order to understand the assumptions in calculating the mean, a preferred way to

write it is as

by an appropriate weight

An unbiasedness condition

This definition is elegant because it clearly shows how the mean is a weighted linear

estimator obtained by adding a weight contribution from each sample (first equation).

The second equation illustrates that each sample is equally weighted. As we shall see,

most algorithms for mapping work in the same way, except the weights applied to each

sample to calculate the estimate are not equal but dependent on some other criteria.

Equation two also shows that the weights are inversely proportional to the number of

samples. Because of this it follows that the sum of the weights must be 1. This is the

unbiasedness condition and is summarised in the final equation.

By continuing the analogy between frequency of occurrence and probability we can

consider the weights as probabilities. So if we have different groupings of samples we

may not want to give all samples equal weight in estimating the mean.

The second moment of a distribution is the variance and it summarises the spread of

the values about the central tendency (expected value). The variance is the average

squared difference of the observed values from the mean:

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

Since it involves squared differences, a single extreme value in the data can have a

strong influence on the calculation of the variance. The variance is one of the most

important parameters used in geostatistics and we shall return to it in many guises

during the course.

The standard deviation is simply the square root of the variance. For a Normal

(Gaussian) distribution, the variance (and therefore standard deviation) have specific

meaning in terms of probabilities. A Normal distribution is completely specified by the

first and second moments (mean and variance).

Note that in geostatistics we use the 1/n definition of variance, which is the observed

variation for a data set. If we require an unbiased estimate of the population variance

then it can be obtained from

The inter-quartile range (IQR) is another useful measure of the spread of observed

values. It is the difference between the upper and lower quartiles. Like the median,

and other percentile-based descriptors, the IQR is robust against the effect of extreme

values within the data. Some estimation techniques, such as fractal measures and

esoteric distributions such as Levy-Stable (of which the Normal is a special case), use

approximations based on discarding certain proportions of the tails of the distribution

in order to improve the estimation.

A measure of spread such as standard deviation is also an estimate of the confidence

we have on a prediction. We have already seen that the mean is a best estimator

for an unmeasured value. The standard deviation is the confidence we have that a

true value will be close to the prediction. For a variable with little variation and thus

a small standard deviation our confidence will be high, but for a variable with a large

variation our confidence will be low that the true value is close to the prediction. The

standard deviation (or variance) have a specific meaning in terms of probability when

we consider the Gaussian distribution (see below).

The coefficient of skewness captures the degree of symmetry of a distribution:

The numerator is the cubed average difference between data values and their mean

and is the third moment. As such, coefficient of skewness is very sensitive to erratic or

outlying values. Consequently, sometimes the magnitude of skewness is ignored and

only the sign of the skewness is considered - positive skew indicating a long tail of high

values, negative skew a long tail of low values.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

The central limit theorem is very important in statistics, but somewhat of a counterintuitive concept. The basic theorem provides the mathematical formal conditions

that support the experimental evidence that the distribution of sample means usually

approximates to a normal distribution, regardless of the nature of the distribution

function of the parent population.

What this means is that even if the distribution function of the parent data is distinctly

non-normal, if we take groups of samples and compute their means then the

distribution of means will tend towards a normal distribution. The consequence is that

we can formulate statistical tests to compare the sample mean against the population

or against means calculated for different sample sets. This allows us to establish

confidence intervals, for example to state the probability that two separate datasets are

from the same population.

The normal distribution is usually justified in representing the distribution of a sum of

additive effects possessing stochastic independence, such as instrumental errors,

regardless of the probability distribution of the individual effects. The independence

may be violated in many natural phenomena, thus accounting for the failure of many

spatial processes to follow normal distributions (Olea, 1991). For processes where the

effects are multiplicative rather than additive, a skewed distribution such as the lognormal may be more appropriate.

The statisticians infatuation with the normal distribution tends to focus interest away

from the fact that, for real data, the normal distribution is often rather poorly realised,

if it is realised at all. The main criticisms of assuming a normal distribution in real

data cases is it is a symmetrical distribution for additive processes (whereas the earth

science processes tend to be multiplicative) and that it is too light in the tails of the

distribution. In other words, extreme values occur more frequently than predicted by a

normal distribution of the same mean and variance as the data. Both of these issues

can be tackled, by normal score transforms (anamorphosis) and by richer families of

distributions such as the Levy-stable functions, of which the normal distribution is an

end member case.

Normal distribution

The normal or Gaussian distribution is often observed in the natural world, either for

data as they come or for transformed data. The heights of a large group of people

selected randomly will look normal in general, for example. The normal distribution

is symmetric about its mean and has a standard distribution , which is such that

practically all of the distribution (99.73%) lies inside the range - 3 x + 3. The

frequency function is:

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

A standard normal distribution (written N(0,1)) is one for which the mean is 0 and the

standard deviation is 1. To convert from a general normal variable x to a standard

normal variable z, we set

99.99

Porosity-Thickness

Cumulative Probability

99.9

99.8

99

98

95

90

80

70

60

50

40

30

20

10

5

2

1

0.2

0.1

0.01

0.

4.0

8.0

phiH

12.

ft

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

We have already seen how we can represent any dataset as a cumulative distribution

function or cdf. The cumulative scale is essentially a linear probability scale between

0 and 1 in other words it is a uniform distribution. Clearly if the data cdf followed a

uniform distribution then they would plot as a straight line.

The probability points for a standard normal distribution N(0,1) are well known (they

can be easily obtained from statistical tables or a spreadsheet). It is therefore a

straightforward matter to take the original distribution, calculate the probability value on

the cdf for each sample and use this probability to compute the value with the same

probability in a standard normal distribution. We then replace each data value with its

standard normal value and have transformed the data from its original distribution to

a normal distribution. Clearly the process is reversible, although there is potential for

some slight loss of accuracy when back transforming the data.

Note that the normal score transform (sometimes called anamorphosis) does not

alter any spatial properties of the data, such as trends. Also, whilst the process is

reversible, if we work with our data and generate values which have a larger range than

the input data (as happens during a Monte Carlo process) then we have the problem of

how to back-transform values for which we do not know the correct cdf relation. This

problem is usually solved by making an assumption about the behaviour of the tails of

the distribution, such as by linear or power law extrapolation.

Although geostatistics does not rely on an assumption that the input variable distribution

be Gaussian, some procedures only work properly if the data are Gaussian. Similarly,

the term best estimate is only guaranteed to be best if the data are Gaussian. One

geostatistical procedure for which a Gaussian distribution is very important is sequential

Gaussian simulation (SGS).

Porosity-Thickness - NScore

Number of Data55

mean 0.00

std. dev. 0.99

coef. of varundefined

maximum2.36

upper quartile0.67

median 0.00

lower quartile-0.67

minimum-2.36

0.160

Frequency

0.120

0.080

0.040

0.000

-3.00

-2.00

-1.00

0.00

1.00

2.00

3.00

4.00

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

Cumulative Frequency

1.00

Porosity-Thickness - NScore

0.80

Number of Data55

mean 0.00

std. dev. 0.99

coef. of varundefined

maximum2.36

upper quartile0.67

median 0.00

lower quartile-0.67

minimum-2.36

0.60

0.40

0.20

0.00

-3.00

-2.00

-1.00

0.00

1.00

2.00

3.00

4.00

99.99

Porosity-Thickness - NScore

Cumulative Probability

99.9

99.8

99

98

95

90

80

70

60

50

40

30

20

10

5

2

1

0.2

0.1

0.01

-3.0

-2.0

-1.0

0.

1.0

2.0

3.0

Bivariate statistics

The scatterplot or crossplot is a graphical tool familiar to geoscientists, especially those

working with well and log data. The scatterplot is the most common way of visualising

bivariate data. Bivariate refers to data for which there are two measured attributes for

each sample. Multivariate data would have many attributes measured for each sample.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

variables can be positively or negatively correlated or uncorrelated. Positive correlation

is where larger values of one variable are associated with the larger values of the

second variable. A negative correlation is where the larger values of one variable

are associated with the smaller values of the second variable. If the variables are

uncorrelated, then an increase or decrease of one variable will appear to have no effect

on the other variable.

The most common statistic to summarise the relationship between two variables is the

correlation coefficient:

The number of data is n; x1,...,xn are the data values for the first variable, mx is their

mean and x is their standard deviation; similarly for y. The correlation coefficient may

be given the symbol r or R as an alternative to

.

The numerator of the above equation is known as the covariance:

The covariance is often used as a summary statistic of a scatterplot and also forms an

integral component of geostatistical analysis, particularly in the key step of establishing

measures of spatial continuity. Note that the covariance between two variables

depends on the magnitude of the data, whereas the correlation coefficient is rescaled

to lie in the interval -1 to +1 by normalising with the standard deviations of the two

variables.

The correlation coefficient is only a direct measure of dependence or independence

when it takes on either of its two extreme values. If X and Y are independent then

r = 0. If r = 1 then we are saying that y and x are dependent (there exists a linear

functional relationship between y and x). Note that r = 0 does not necessarily imply

independence.

The magnitude of the correlation coefficient (r) between these two extremes has no

special significance - it is a relative measure. The quantity r2 is a useful approximation

for the relative increase in dependence. R2 should be interpreted as the amount

of variability of one attribute that could be predicted from knowledge of the second

attribute. Take note of this that means that for a correlation of 0.71, only half of the

variability can be explained from the other variable and half cannot. Similarly, for a

correlation coefficient of 0.5, only 25% of the variability can be explained, 75 % cannot

be explained from knowledge of the second variable.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

For example, one study showed that the r2 between seismic relative acoustic

impedance (AI) and porosity is 0.3 and increased to 0.6 when seismic AI plus a

low frequency model was compared to porosity. This suggests that, to a first order

approximation, the low frequency model is about of equal importance as the high

frequency information extracted from the seismic in estimating AI.

It cannot be stressed too strongly that a statistical correlation cannot, under any

circumstances, be used to establish or infer a causal relationship between x and

y. At best a correlation can be used to guide the formulation of physical models or

hypotheses that might form a predictive tool whose operational constraints (accuracy/

precision) might be summarised in some statistical way. Note also that the correlation

coefficient is only indicative of a linear dependence between y and x. A non-linear

dependence is not considered.

An alternative and useful test that can check for correlation between two variables

irrespective of the form of any relation (linear or non-linear) is the rank correlation

coefficient. This is based on comparing the rank values instead of the measured

values for each attribute, using the following equation

An algorithm for computing rank correlation can be found in Press et al (1992). If two

variables exhibit a significant rank correlation then it can be taken as strong evidence

that some association exists between them, although not the form of the association.

It is straightforward to create a second variable Y correlated with an initial variable X

as follows

If x1 is a N[0,1] variable, we can create any other Gaussian variable X with mean

= and standard deviation = using:

random variable x2 and combine them as follows, where is the correlation

coefficient

X and Y will then be correlated with each other according to the chosen

correlation coefficient

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

Spurious Correlation

Whenever two variables are plotted against each other in the form of a crossplot, it is

common practice to calculate a line estimating the linear relation of the two variables

and to compute the correlation coefficient between the two variables.

The correlation coefficient measures the degree of dependence between the two

variables. It is used for a number of purposes such as selecting seismic attributes

to assist in predicting reservoir properties and for choosing a functional relationship

for time to depth conversion. The problem is that the correlation coefficient is often

estimated using only a small number of sample points - the data at well locations. As

with any statistical measure, as the sample size decreases so the range of the true

parameter, for a given confidence level, increases.

A test for spurious correlation is made using a two-tailed Students T test with two

degrees of freedom. This is very simple to calculate in Excel, and the formula is

given below. 3 cells are set up for entering the values for number of wells, correlation

coefficient and number of attributes considered. A fourth cell contains the formula to

calculate the probability of spurious correlation (Psc). The calculation cell contains the

following entry :

=1-(1-(TDIST((R * (SQRT(Nwells-2))/(SQRT(1-R^2))), Nwells-2, 2)))^Nattr

where Nwells, R and Nattr are the 3 cell references where the values are entered.

The Psc may be expressed as a percentage. Alternatively, a Java Applet for the same

computation can be found at http://www.sorviodvnvm.co.uk.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

Psc

(1 attribute)

Psc

(3 attributes)

Al vs

R

(7 wells)

Rmin

Rmax

Porosity

0.913

0.511

0.987

0.4%

1.2%

PorosityThickness

0.939

0.634

0.991

0.1%

0.5%

Thickness

0.719

-0.074

0.954

6.8%

19.2%

The number of attributes shows the effect of trying more and more correlations with

different attributes until a correlation is found. The chance of finding a spurious

correlation increases as the number of comparisons increases. The tables overleaf

show the Psc for a range of values of R and Nwells.

Table shows probability of spurious correlation for a given R (left hand column)

and number of samples (column headings).

number of attributes (left hand column) and number of samples (column headings)

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

Note that for spatially correlated variables the results given above will be overly

optimistic because the degrees of freedom assumes there are N independent samples.

If the samples are close enough together then they may be spatially dependent and so

N is effectively reduced, rendering some data in the regression redundant and therefore

the chance of spurious correlation will increase. In other words, real data cannot be

better than given here and will often be worse!

Whenever a correlation is used for estimation, the correlation coefficient and the Psc (or

alternatively, the number of samples) must be reported if the meaning of the correlation

is to be communicated to the reader.

In addition to computing the Psc, we can also calculate, for a given confidence interval,

the limits of the true correlation between the variables. This is an equivalent calculation

the Psc calculation computes the probability of this confidence interval including

zero correlation. The correlation range is computed approximately using Fishers

Z-transform. The already mentioned Java Applet gives the minimum and maximum R

for a 95% confidence interval about the input R.

PREDICTION ERRORS

Actually

uncorrelated

Actually

correlated

Select

correlation

Type l Error

Correct decision

(no error occurs)

Reject

correlation

Correct decision

(no error occurs)

Type ll Error

When choosing to use or ignore a correlation, there are four possible outcomes. If we

elect to use a correlation and the variables are genuinely correlated we have made

a correct choice and reduced uncertainty. If the variables are not correlated and we

reject the correlation we have correctly avoided being misled by an inappropriate

relation.

However, there are two possible prediction errors that we can make, depending on

our choice.

In a Type I error we elect to use a correlation but in fact the variables are not

correlated. This results in a prediction using the erroneous variable which is less

accurate but more precise. In other words, we mistakenly are led to reduce the

uncertainty resulting in an inaccurate prediction made with confidence.

With a Type II error we reject a correlation, even though it would have been

useful in reducing uncertainty. The result is a prediction which is both less

accurate and less precise than if we had used the correlation. We have a

prediction with more uncertainty than necessary because we have rejected

useful information.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

expressing a quality, it should be remembered that this is not the same as measuring it.

For example, if X and Y are independent of each other then the correlation coefficient is

zero. However the converse is not necessarily true: if the correlation coefficient is zero

it cannot be concluded that X and Y are necessarily independent. Finally, correlation

coefficient is a relative indicator of dependence. The most useful description is the

R2 value, which expresses the proportion of variance which is shared between the

two variables.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

as predictors of reservoir properties

&<17+,$ 7.$/.20(<0RELO( 3 7HFKQLFDO&HQWHU'DOODV7H[DV

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-

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.

attributes to predict reservoir properties.

means we have happened to sample at locations that

yield a large sample correlation when in truth they are

MARCH 1997

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

247

STATISTICS

EARTH

WORKS

7DEOH3UREDELOLW\RIVSXULRXVFRUUHODWLRQEHWZHHQD

UHVHUYRLUSURSHUW\DQGDVLQJOHVHLVPLFDWWULEXWH

6DPSOHVL]H

5

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 (l) is zero can be found from

R n<2

p sc Pr r * R Pr t *

1< R2

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

7DEOHV 3UREDELOLW\ RI REVHUYLQJ DW OHDVW RQH VSXULRXV VDPSOH FRUUHODWLRQ IURP D

JLYHQQXPEHURILQGHSHQGHQWVHLVPLFDWWULEXWHV

7DEOH7HQLQGHSHQGHQWVHLVPLFDWWULEXWHV

7DEOH)LYHLQGHSHQGHQWVHLVPLFDWWULEXWHV

6DPSOHVL]H

6DPSOHVL]H

5

7DEOH7ZHQW\LQGHSHQGHQWVHLVPLFDWWULEXWHV

7DEOH)RUW\LQGHSHQGHQWVHLVPLFDWWULEXWHV

6DPSOHVL]H

6DPSOHVL]H

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

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.

attribute as a predictor of a reservoir property.

k

i <1

i 1

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

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

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

Introduction to Geostatistics

MARCH 1997

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

249

20

STATISTICS

EARTH

WORKS

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

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.

D

seismic attribute

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.

Figure 7. (right)

Estimate of porosity generated

using collocated

cokriging with the

seismic attribute

as the covariate.

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

MARCH 1997

more than five porosity units. (b) Areas underestimated by more than five porosity units.

Figure 6. (above)

Scatter plot of

porosity versus

seismic amplitude

at all

locations.

250

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

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

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-

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

predictors. /(

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.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

When using techniques such as multilinear regression or correlation, we should pause

to consider just how many attributes a seismic dataset may have. Chen and Sidney

(1997) published a comprehensive list of possible seismic attributes which could be

determined from seismic data. Clearly, there are an infinite number of attributes which

could be ingeniously defined, but seismic data cannot have an infinite information

content to be given up. This leads us to pose the question How many independent

seismic attributes are there?

In the seminal paper on complex trace attributes, Taner et al (1979) open by clearly

stating the application of complex trace analysis to seismic data as transformations of

data from one form to another to extract significant information. They clearly state the

benefit of these transformations: Interpreting data from different points of view often

results in new insight and the discovery of relationships not otherwise evident.

Applying a transformation to data does not result in generating new variables. For

example, if we have variables X and Y and multiply Y by a constant, we do not have

a new variable, merely a scaled version of the previous version of Y. If X and Y

were linearly related, then the new Y is also linearly related. However, consider that

if ln(Y) was linearly related to X then applying a log transform to Y would enable the

nature of the relationship to be more easily recognised. In other words, transforms

do not result in new variables, they merely assist in highlighting relationships between

different variables.

Given a seismic trace, when we take the Fourier transform we do not obtain additional

independent variables. Through the FFT, amplitude and phase are functionally (but not

linearly) related to the real and imaginary parts of the transformed trace. However, by

computing amplitude and phase spectra we can observe features of the trace data in a

way that is not possible in the time domain. Our data remains the same, only our view

on the data changes.

So for post-stack data, how many seismic attributes are there? Consider the basic

equation for a seismic trace where the real seismic trace f(t) is expressed in terms of

time-dependent amplitude A(t) and a time-dependent phase (t):

The trace clearly has two attributes time and phase. (Of course, pre-stack seismic

data contains additional information in the offset data) . We can derive another

attribute, which is the time difference between events on the trace, providing they are

sufficiently far apart not to be autocorrelated. A wedge model illustrating tuning is a

good example of how the interval time becomes dependent when the two events are

close enough together that the autocorrelation (wavelet) becomes important.

Consider that the most unsymmetrical material has a maximum of 21 elastic constants.

As the degree of symmetry increases, the number of independent elastic stiffnesses

decreases to 9 for an orthotropic material, 5 for a transversely isotropic material and 2

for an isotropic material (Gregory, 1977).

Clearly there is a benefit in transforming seismic through complex trace analysis in

trying to discover useful relationships. However, it should not be assumed that new

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATISTICS

EARTH

WORKS

variables are gained in this way. The variables are all wholly dependent on each other,

although sometimes cryptically because of the nature of the transform itself. Methods

such as multi-linear regression rely on the input variables being independent. There is

no value in using two variables which are themselves correlated, almost certainly one

will do.

There are other dangers in taking multiple derivatives of a seismic trace, through

complicated transforms. One is that the probability of generating a spurious correlation

increases the more comparisons are made. Also, high order derivatives are much

more sensitive to noise. It might be tempting in a geostatistical study to find reasons

to reject outlying samples caused by noise, which could give rise to a very high (but

spurious!) correlation coefficients for the remaining samples.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

STOCHASTIC THEORY

Random function model

There are essentially two models which we can use conceptually to describe data and

samples. We will illustrate the difference between these two models by considering

the quantity porosity.

The notion of porosity is simple to understand but it poses some problems if we want

to define it with precision. There are two accepted ways of defining local properties of

a medium: the notion of the representative elementary volume (REV) and that of the

random function (RF). These two alternative models implicitly influence descriptions

of spatial variation.

REV

Random Function

The problem stems from the fact that porosity, like many quantities, cannot be defined

or measured at single points, since a porous medium is a conglomeration of solid grains

and voids. Below a certain scale of volume, porosity has no physical significance.

The REV model consists of saying that we give to one mathematical point in space the

porosity of a certain volume of material surrounding this point, the REV, which will be

used to define and possibly measure the mean property of the volume in question.

This concept involves an integration in space. It is obviously the first method that

comes to mind and behind it lies the idea of a sample of finite size which is collected

and from which the relevant property is estimated by measurement.

The size of the REV is defined by saying that it is

mean global property, while ensuring that the effect of the fluctuations from one

pore to another are negligible.

Sufficiently small so that the parameter variations from one domain to the next

may be approximated by continuous functions, in order that we may use the

infinitesimal calculus, without introducing any error that might be detected by

measuring instruments at the macroscopic scale.

It should be noted that in a fractured medium the size of the REV may be quite

astonishingly large to satisfy the first condition, but therefore failing to satisfy the

second condition of a continuous function at the scale of the measurement. The

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC THEORY

EARTH

WORKS

size of the REV is often linked to a flattening of a curve connecting the property with

the dimension of measurement. However, nothing allows us to assert that such

a stabilisation must exist. The size of the REV is thus arbitrary and there may no

characteristic scale at which it can be said to exist.

The REV concept has other important limitations. Firstly, it is unsuited to handling

discontinuities in a medium. Second, and most importantly, is that it gives no basis for

studying the structure of the property in space. The most that can be said is that the

spatial variations of the studied property must be smooth.

The RF model is a more powerful concept. It comprises in saying that the studied

porous medium is a realisation of a random process. A property like porosity is

then defined at a given geometrical point in space as the average over all possible

realisations of its point values, defined either as a grain (0) or a pore (1). We then have

an ensemble average of grain or pore at a point and therefore refer to porosity at a point

as having some probabilistic interpretation. Thus the ensemble average is in fact the

expected value of the point porosities. Over the whole volume (e.g. the right hand side

of the picture) the expected value will be the same as the space average defined by the

whole sample - the REV. However, the ensemble average will be the same probability

at any point within the sample.

The significant advantage of the stochastic approach of the random function model is

that it becomes possible to study statistical properties of the ensemble other than just

the expected value.

Note that the real world from which the sample is drawn is just one complete realisation

of some random function and not the random function itself. In order to make this

concept work we require some additional assumptions, the most useful of which are

stationarity and ergodicity. (From de Marsily, 1986)

We shall now demonstrate a simple stochastic model. The word stochastic means that

we are concerned with probabilities. Consider an experiment involving tossing a fair

coin (ie one that can only land heads or tails, not on its edge nor can it be grabbed by

passing birds, roll down cracks in the floor or any other outcome!)

The Expected Value Ev is (P0.5 * 0 + P0.5 * 1) = 0.5

Ev is the probability term for the mean or average result

Our best estimate of the result of tossing the coin is always 0.5 because this

gives the smallest average square error (least squares)

The only possible realisations (outcomes) for tossing the coin are 0 or 1

When we consider a discrete event the distinction between expected value and

realisations is very clear. However, the same principal applies to continuous variables.

The mean is not a real value, it is a mathematical artifice describing the probability

weighted sum of the possible outcomes.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC THEORY

EARTH

WORKS

Regionalised variables

A key difference between statistics and geostatistics is the recording and extracting of

information related to the location of every observation. Attributes that have a location

(coordinate reference) associated with each observation are called regionalised

variables. Geostatistics models regionalised variables as realisations of random

functions.

We start with the observation that two data points which are close together are much

more likely to have similar values than two points which are far apart. This phenomena

is used quite naturally in procedures such as hand contouring of data. In geostatistics

we describe this property as spatial continuity or spatial correlation. Unless a variable is

spatially correlated, it is meaningless to attempt to map it.

Variography is a method for measuring and modelling spatial correlation. There are a

number of related statistical tools to describe spatial correlation

h-scatterplots

Correlation function

Covariance function

Semivariogram

the offset between two points. In the case of a spatial variable, the lag is the distance

between two points. The lag h is a vector, so it has both separation and direction.

To construct an h-scatterplot for a chosen lag value

find all the pairs of values which are separated by h (to within a given tolerance).

Construct a cross-plot of second value in pair against first value in pair

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC THEORY

EARTH

WORKS

h=100m

h=200m

h=300m

h=400m

h=500m

h=600m

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC THEORY

EARTH

WORKS

If we choose h to be very small, we would expect to see very little difference in the

values of the first and second data making up each pair. If h is large then the first and

second value of the pair may be quite different.

We can summarise a linear relation on the cross-plot using any of the common

statistics. For example we could use correlation coefficient, covariance (numerator of

the correlation coefficient) or semivariance (the deviation of the cross-plot values from

the 1:1 line).

If we calculate the correlation coefficient for each h-scatter plot and then plot the

correlation coefficients against h, we obtain the correlation function. Similarily we can

obtain the covariance function and the semi-variogram by plotting the semivariance

against h. In practise we rarely draw the h-scatterplots. Instead we compute the

semivariogram function directly using

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC THEORY

EARTH

WORKS

semi-variogram) is the main spatial

analysis function used in geostatistics.

If we turn this function upside down,

we have the covariance function. If we

normalise the covariance function to a

range of 0 1 we have the correlation

function.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC THEORY

EARTH

WORKS

Classic variogram

The following terms are used to describe the variogram

As the separation distance between pairs increases, the variogram value also

increases (in the ideal case). Beyond a certain separation no further increase

occurs and the variogram reaches a plateau. The separation at which this

plateau is reached is known as the range (symbol a) and it defines the spatial

separation greater than which there is no further spatial correlation apparent

for the variable. Points separated by distances greater than the range only

contribute information about the mean value - this is the realm of classical

statistics and implies the data cannot be mapped.

The magnitude of the plateau which the variogram reaches at the range is

called the sill. Under certain special conditions, the sill is equal to the classical

variance of the data. The sill is given the notation C+C0, strictly the sill is C, but

the nugget (described below) is often added into the definition.

The strict value of semivariance for h=0 is zero. Any point compared to itself

will have no difference. However a number of factors including measurement

error and short scale variability can cause sample values separated by short

distances to be quite dissimilar. This gives rise to a discontinuity at the origin

referred to as the nugget effect. It is usually given the symbol C0.

Y axis (North)

ng

=6

:a

r

cto

lag tolerance

xltol = 2.0

nv

tio

ec

dir

g6

La

ang = 60

g5

La

g4

La

g3

lag distance

xlag = 4.0

La

bandwidth

bandw = 5.0

2

ag

L

g1

La

X axis (East)

An omnidirectional variogram is a variogram where all sample pairs are used in its

construction and the direction element of the vector between a pair of samples is

ignored. The variogram is therefore based on separation (distance between) pairs

only. Since omnidirectional variograms use all the availble pairs they tend to be

robust but clearly if the geological property has an anisotropic behaivour then the

omnidirectional variogram will average it.

A directional variogram is where the binning scheme only uses pairs oriented in a

given direction such as N-S. This may be rather restrictive and no pairs may be found

that exactly meet this criteria and so it is usual to consider pairs for which the angle

between the two points is within somne tolerance of the ideal angle. For example, N-S

+/- 20 degrees. A typical scheme for directional variograms is illustrated in the diagram

(above right).

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC THEORY

EARTH

WORKS

Note that a directional variogram for which the tolerance is set as 90 degrees is in fact

an omnidirectional variogram as all pairs will be included. Using an angular tolerance

of 90 degrees and then changing the direction will always give the same calculated

experimental variogram. To observe directional behaviour the angular tolerance

should be reduced to make the selected pairs specific to the direction being analysed,

but the tolerance should not be too narrow or few pairs will be found and the reliability

of the resulting experimental variogram will be impaired.

When computing directional variograms it is usual to analyse several directions,

typically orthognal to each other or at regularly spaced angles around the compass.

Typically, four directions might be analysed, starting N-S (0 degrees), then NE-SW

(+45 degrees), E-W (+90 degrees) and NW-SE (+135 or -45 degrees). The angular

tolerance should usually be set no larger than 22.5 degrees to avoiding double

counting of pairs. Ideally, if sufficient data are available, the angular tolerance could

be set to 5 - 10 degrees.

The objective of directional variograms is to investigate whether the geological property

exhibits anisotropic behaviour, that is its covariance function is elliptical rather than

circular when observed in plan view. Remember that a directional variagram is a

radial cross-section from the centre out through the covariance function. The vertical

variogram will almost certainly show anisotropy compared to the horizontal variograms,

the vertical range being much shorter (typically 20:1 to 30:1 ratio between horizontal

and vertical range).

The presence of a drift or trend will strongly mask any anisotropic bevaviour and

detecting anisotropy in these circumstances (other than vertical versus horizontal) will

be difficult. In many instances, a drift or trend is mistakenly interpreted as anisotropy.

If anisotropic behaviour is suspected, the directional variograms should be aligned with

it to enable the anisotropy to be fully odentified and subsequently modelled.

A Good Variogram

It will become clear from the above practical examples that a good stable variogram

requires the data both to be stationary and be comprised of a sufficient number of

samples. In fact, these two criteria are to some extent related, in that stationarity is a

function of low frequency aliasing, that is, under sampling.

Assuming a stationary function, a question often asked is how many values are

required to ensure a stable variogram? Certainly it is true that one can never have too

many samples, but it is common to see attempts to compute variograms on far too few

samples. Webster and Oliver (1992) investigated this question for a synthetic variable

of a soil property.

Firstly, they noted that the advice of Journal and Huijbrets (1978, p194) is misleading.

That advice suggests the number of pairs need be a minimum of 30 - 50, and that the

maximum h considered should be less than L/2 where L is the dimension of the field

from which the pairs are computed. If we take 45 pairs to be the total number of pair

combinations, we require only 10 samples to achieve this. In hydrocarbon exploration

when dealing with well data this may be the typical number of samples under

consideration. Unfortunately, variograms derived from this number of samples are

likely to be unreliable, difficult to interpret, or both. A better rule of thumb is to assume

that the average number of pairs contributing to each lag should be in the range 30

- 50.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC THEORY

EARTH

WORKS

Webster and Oliver observe that variograms computed on fewer than 50 data are of

little value, and that at least 100 samples are needed. They suggest that for a normally

distributed isotropic variable a variogram calculated from a sample of 150 points might

often be satisfactory, while one derived from 225 data will usually be reliable. As an

example of this, they constructed variograms from the synthetic dataset shown below,

which has a known variogram (exponential with range 40 units and a sill of 1), which

was generated from a consistent random function model.

Variograms generated by sampling 49 points on this surface are shown in (a) above

Note how disparate the variograms are, and also how difficult it would be to determine

a reliable nugget effect. In (b) above, the variograms have been derived from the

synthetic data of the previous page using 225 points per variogram. The difference

between this result and (a) above should speak clearly of the problems of inadequately

sampling to generate variograms.

a 7 x 7 grid at 15 unit intervals

a 15 x 15 grid at 7 unit intervals

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC THEORY

EARTH

WORKS

an anisotropic variable will have only pair contributions in certain directions. For the

example variogram looked at previously for hand calculation, in the X direction there

are 61 pairs, an average of 10 per lag, but in the Y direction the total number of pairs

possible is 31, and there are no lags greater than 300 m. The total number of pairs

possible (omnidirectional) for this example of 24 samples is 276, with a maximum

range of 671 m (2 pairs).

Another aspect of good variograms not often considered is the behaviour close to the

origin at short lags. Perversely, with a very regularly sampled data set, such as from

a 3-D seismic survey, there are no lags at an offset less than the grid interval, and so

very little is known about the spatial behaviour at short lags. With a scattered data set,

the separations are more evenly distributed across many values of h, and so although

an overall stable variogram may be more difficult to interpret, there may be significantly

more pair information at very short offsets, which often is the most critical part of any

variogram analysis. This problem was considered in an excellent paper by Laslett

and McBratney (1990), and we shall briefly look at their recommendations concerning

sampling strategy in a short section later in this course.

Therefore, should we conclude that it is pointless to attempt to construct variograms

from too few samples? The only answer is to try it and see. A well behaved, simple,

continuous function may require only a few samples to describe it quite well, so it is

always worthwhile constructing a variogram just to see.

Variogram of 55 Wells

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC THEORY

EARTH

WORKS

Variogram of 7 wells

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC THEORY

EARTH

WORKS

We will define a very simple probabilistic spatial model. It is defined by a univariate

and a bivariate (joint) probability law. as follows:

Starting condition:

V(0) =

1 with probability = 0.50

V(x) =

V(x-1) with probability = 0.75

1-V(x-1) with probability = 0.25

The univariate law shows that the possible outcomes are 0 or 1, with equal probability.

This law is independent of the location x. Given the probabilities and magnitudes of the

outcomes, it is trivial to calculate that the expected value is:

The joint probability law shows that the value at the next location of x is the same as

the previous value with probability 0.75, or that it may take on the other value with

probability 0.25. Therefore two points at one unit distance apart have a joint probability

distribution. Note that this law does not depend on whether the previous value is 0 or

1, therefore it is independent of the x location.

Considering an adjacent pair, located at x and x+1, the set of possible outcomes is:

{(0,0), (0,1), (1,0), (1,1)}

and the associated probabilities are:

{3/8, 1/8, 1/8, 3/8}

There are two possible outcomes that we are interested in.. These are when the pair

are the same and when the pair are different values:

(0,0) and (1,1) = 3/8 + 3/8 = 0.75

(0,1) and (1,0) = 1/8 + 1/8 = 0.25

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC THEORY

EARTH

WORKS

which (not surprisingly!) are the joint probability values we defined. The semivariance

for h =1 is

(0.75 * 0 + 0.25 * 1)/2 = 0.125.

If we know the joint distribution of adjacent pairs, we can compute the joint distribution

between pairs separated by any separation. For example, considering the pairs at V(x)

and V(x+2) the following permutations are possible:

{(0,0,0), (0,0,1), (0,1,0), (0,1,1), (1,0,0), (1,0,1), (1,1,0), (1,1,1)}

and the probabilities are:

p=9/32

p=3/32

p=1/32

(0,0,0), (1,1,1)

(0,0,1), (0,1,1), (1,0,0), (1,1,0)

(0,1,0), (1,0,1)

if we group these according to the pairs separated by a lag of 2, we get the outcomes:

{(0,0), (0,1), (1,0), (1,1)}

and the corresponding probabilities are:

{10/32, 6/32, 6/32, 10/32}

considering the two outcomes where the pair x and x+2 are the same number or the

opposite number we get:

(0,0) and (1,1) = 10/32 + 10/32 = 20/32 = 0.625

(0,1) and (1,0) = 6/32 + 6/32 = 12/32 = 0.375

The semivariance for h =2 is

(0.625 * 0 + 0.375 * 1)/2 = 0.1875.

We can carry this process on to calculate the joint probability for any separation of

points. Given that the expected value can be computed from the probability multiplied

by the outcome and the variance is the squared difference of the outcome from the

expected value, we can calculate variance for any separation of points. The variance

plotted as a function of separation h has been calculated and is shown on the next

slide. This graph shows the characteristic semivariogram form. In this case it will

reach the asymptotically at infinite, where the sill will attain a value of 0.25 (the global

variance calculated from the univariate statistics).

To summarise the key points from this example, note that:

The joint probability law results in spatial correlation

There are an infinite number of possible realisations of this random function.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC THEORY

EARTH

WORKS

If we draw realisations of this model, ie actually input the two probability laws to

a spreadsheet and draw random numbers according to the laws, each time we

recalculate we obtain a different sequence. Three example sequences are shown

below for the function V(x). Note that each realisation has the same univariate and

joint probability. The joint probability law results in spatial correlation : locations

close together tend to have the same value. Each of these realisations has the same

variogram.

Consider now a change of parameter in our function V(x). We will change the value of

0.75 to 0.875. We will call this function U(x). This change means that adjacent values

are more likely to be the same. However, the univariate probability law is unchanged

and therefore the expected value and variance remain the same. If we calculate the

variogram, we now find it looks like the graph shown below and if we draw realisations

we find that they have a greater degree of spatial continuity, as we would expect.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC THEORY

EARTH

WORKS

Finally consider the effect of changing the joint probability value to 0.5. This means

that adjacent values are just as likely to change as stay the same. This law results in

no spatial correlation : its as though a 1 or 0 is picked at random at each new location

of x. The variogram for this function is a constant value of 0.25 for all values of x

and the realisations appear random. For an expanded discussion see Isaaks and

Srivastava, (1989) from where this example came.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

STATIONARITY

Stationarity and ergodicity

Stationarity is the assumption that any of the statistical properties of the medium (mean,

variance, covariance or higher order moments) is stationary in space i.e. does not vary

from place to place. Note that this does not mean that the property is uniform, just

that its probability (expected value) at any point is invariant spatially. As we will see,

the assumption of stationarity is rather restricting. It is the hypothesis or decision of

stationarity which allows inference. Stationarity cannot be checked from data and is a

function of the sampling window dimensions.

Ergodicity implies that the unique realisation available behaves in space with the same

pdf as the ensemble of possible realisations. In other words, by observing the variation

in space of the property it is possible to determine the pdf of the random function for all

realisations. This is called the statistical inference of the pdf of the RF Z(x).

In stochastic terminology a phenomenon that is stationary and ergodic is called

homogenous. The term uniform is used to describe a medium in which some property

does not vary in space - which would usually be called homogenous to geologists or

geophysicists.

We use these properties along with the concept of a random function in many

scientific tasks, although we may be unaware of the implicit assumptions. For example,

calculating porosity from thin section by point counting assumes the sample is ergodic

and stationary. Moving the slide around is the equivalent to sampling of multiple slides

at the same point.

Weak stationarity is where only the first two moments are stationary. I.e. the RF

Z(x) has an expected value (mean) and a covariance which are not a function of the

co-ordinate space. The covariance is then assumed to be a function of the distance

between points. This is known as second order stationarity.

The intrinsic hypothesis consists in assuming that even if the variance of Z is not

finite, the variance of the first order increments (lags) of Z is finite and that these

increments are themselves second-order stationary. The variance of the increment

defines a function called the variogram:

which is the function of the mean squared difference in the property for points

separated by a distance h.

Stationarity is a property of the model, not of the data. The decision of stationarity

allows inference. We trade unavailable replication of the random function at a point for

another replicate elsewhere (in time/space) ie our conventional notion of collecting data

samples. This trade is the hypothesis or decision of stationarity. This decision cannot

be checked from data, is a function of the sampling window (scale of observations) and

to some extent contradicts fractal relations.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATIONARITY

EARTH

WORKS

cells and fill them with the formula =rand()*10. This formula instructs the program to

generate uniformly distributed random number on the interval [0,1] and then multiply it

by 10, giving a random number uniformly distributed on the interval [0,10].

However when we look at the spreadsheet we see the result of the computation, a

random number is displayed in each cell. The formula has the same expected value in

each cell (5) but the display shows a realisation of this formula the random number.

To an observer who is not allowed to examine the formula in any cell, they can only

assume the expected value is the same in every cell and approximate it by taking the

average of the results. To generate a new realisation (new universe) we can press the

recalc button. The result is a different array of random numbers.

Most inversion examples demonstrating a particular contractors methodology are based

on stationary models. These are the easy ones to solve. The real problem with

inversion is the reasonableness of the stationarity assumption and subsequently the

expected value predicted away from the well.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATIONARITY

EARTH

WORKS

Global Warming

The stationarity argument is often at the root of scientific and technical disagreement.

For example, when different seismic interpreters put different maps on a wall based

on different geological concepts, they are often disputing the stationarity assumption

with each other. A more famous example is the scientific dispute over the presence

or otherwise of global warming.

The global temperature graph with time shows a steady rise of average temperature

with time over the past 100 years. It should be remembered that it is very difficult to

compute these estimates and they may not be reliable. It is also worth contemplating

that the graph is falling towards the temperature minima reached around 1893, and it

is doing so for data prior to 1880, which is not often shown although it has the same

reliability back to about 1859.

Based on these data, it has been proposed that we are in a period of climate change

(which is indisputable, given our knowledge of the geological record) and that this is a

consequence of global warming caused by a buildup of greenhouse gases. We have

to consider very carefully

Is this is just natural fluctuation around a steady long term value (stationary

process)

Even if the first model were correct, it is still a further hypothesis that the cause is

release of hydrocarbons from burning fossil fuels.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATIONARITY

EARTH

WORKS

Those of us with longer memories will recall that in the mid 1970s a number of

prominent scientists became very concerned about climate change, based on the

average temperature data chart up to that point. In Britain, they included Sir Fred

Hoyle, then Astronomer Royal. They all pointed to the gentle downward trend in

temporature since 1940 as potentially indicative of the onset of a new ice age.

In 1976 the Australian Academy of Science prepared a report on climate change for the

Australian government. The introduction stated:

This report was prepared at the request of the Minister for Science

for advice on statements by some climatologists that the earth was

undergoing a continuing cooling trend. Examination of recorded

temperatures in the northern hemisphere suggest that, after nearly

100 years of slowly rising temperatures, there has been fall since

1940, most pronounced at high latitudes and averaging -0.3 degrees

Celsius. Combined with extreme climatic events elsewhere during

the early 1970s, notably droughts in the Sahel and the Ukraine and

failures of the monsoon in India, this fall in temperature has led

some climatologists to suggest that the worlds climate is progressing

rather rapidly towards another glacial phase, or at least another Little

Ice Age

Prominent British Scientist Sir Fred Hoyle, Astronomer Royal, wrote articles for serious

newspapers describing that we had to act quickly to prevent the ice age. He proposed

huge pumping systems to bring deep ocean water to the surface and stabilize the

temperature.

In September 1989 Scientific American devoted an entire issue to the theme Managing

Planet Earth. In introducing an article entitled The Changing Climate, it was stated

that:

The earth owes its hospitable climate to the greenhouse effect but

now the effect threatens to intensify, rapidly warming the planet. Rising

concentrations of carbon dioxide and other gases are the cause. The

danger of warming is serious enough to warrant prompt action.

(The above are taken from McTainsh and Boughton, 1993, p142).

In just 15 years we had swung from ice age to runaway heating. Even the language is

the same: extreme weather events are to be expected.

If we remain objective, we can see that this is an argument about stationarity

assumptions. If we look through different length time windows, we may draw different

conclusions. For example, the Intergovernmental Panel on Climate Change report

in 1995 which stated there is a discernible human influence on global climate and

formed the basis of the Rio Summit) showed the temperature graph deduced for the

previous 1000 years. The Little Ice Age during the 17th century is clear, as is the

Medieval warm period, when temperatures were several degrees higher than they are

today. Based on this, some scientists might consider a stationary temperature model,

with fluctuations around a mean temperature similar to today.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STATIONARITY

EARTH

WORKS

If we take an even longer view, we might consider the gradual temperature decline of

the past 3000 years shown by the data of Keigwin (1996). A stationary temperature

assumption looks improbable, and a warming trend highly implausible.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

ESTIMATION - KRIGING

Mapping and Prediction

Mapping is all about estimating the value of an attribute at an unmeasured location.

It is not the same procedure as undertaken by cartographers, whose objective is to

represent on a sheet of paper as accurately as the scale allows that which they already

know.

In order to make an estimate an assumption must be made concerning a suitable

predictive model. All estimation or prediction requires the use of a model. Many

commercial gridding packages use arbitrary models which are not visible to the user.

In selecting a model it is important to consider the processes which have given rise to

the phenomenon under consideration. For some processes the underlying model may

be so well understood that only a few sample values are required in order to estimate

the values at unmeasured locations. In this circumstance a deterministic model is the

most appropriate.

Very few earth science processes are sufficiently well understood to allow the use of

deterministic models in estimation. Instead we must accept that there is uncertainty

about the behaviour of a phenomenon between sample locations. A random function

model as used in geostatistics recognises this uncertainty and provides the tools to

estimate values at unmeasured locations based on assumptions about the statistical

characteristics of the phenomenon.

As well as estimating a value it is also desirable to know the quality of the estimate.

Without an exhaustive reference set (which would obviate the need for prediction!) the

goodness estimate is rather qualitative and depends greatly on the appropriateness

of the underlying model. Models are neither right or wrong : they are appropriate or

inappropriate.

Methods of point estimation, such as polygons, triangulation or inverse distance

methods, may be best, that is predict the true value, depending on the estimation

criteria. All of these estimation methods are linear and theoretically unbiased.

However, of all the linear estimators, kriging is known by the acronym B.L.U.E., which

stands for best linear unbiased estimator. Ordinary kriging is linear because its

estimates are weighted linear combinations of the available data; it is unbiased since it

attempts to have the mean residual, or error, equal to zero. So why is kriging best? It

is best because it also aims to minimise the variance of errors.

Actually, these aims (mean residual equal to zero and minimisation of error variance)

are not attainable for a sample dataset for which there is no final reference. Because

we never know the mean residual, nor the error variance, we cannot guarantee that we

have made them exactly zero, or minimised respectively. The best that can be done

is to build a model of the data and work with the average error and error variance of

the model. This is done in ordinary kriging by using a probability model in which the

bias and error variance can both be calculated, and then we choose weights for nearby

samples used for estimation to ensure that the average error for our model is exactly

zero and that the modelled error variance is minimised.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

ESTIMATION - KRIGING

EARTH

WORKS

At every location we wish to predict a value, we will estimate the unmeasured value

using a weighted linear combination of the available samples. This set of weights will

change as we estimate new unknown values at different locations.

First we assume that for every point at which we want to estimate a value our model

is a stationary random function that consists of several random variables, one for each

surrounding sample used for estimation, plus the sample location to be estimated.

Each of these random variables has the same probability law and any pair of random

variables has a joint distribution that depends only on the separation between the two

points, and not on their locations (i.e. stationarity). The covariance between pairs of

random variables separated a distance h is Cv(h).

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

ESTIMATION - KRIGING

EARTH

WORKS

Without going into the derivation (which we are trying to avoid in this course), based on

the assumption of a stationary random function, it can be shown firstly that in order to

arrive at the unbiasedness condition, the expected value is set to zero and the result of

this is that the sum of weights to ensure unbiasedness equals 1. Next, it can be shown

that to minimise the error variance, we can express the variance of error as a function

of the covariances of the pairs of points for estimation. This gives an expression

for the error variance based on the random function model parameters variance

and covariance as a function of

n variables - the weights for each

sample to estimate the unknown

location value.

The minimisation of a function

of n variables is conventionally

accomplished by setting the n partial

first derivatives to 0. This produces

a system of n equations that can be

solved by any standard procedure

for solving simultaneous linear

equations. However, for ordinary

kriging, we imposed the additional

restriction that the set of weights

must sum to 1 and so we need to

convert our constrained minimisation

problem to an unconstrained one.

Without the constraint we have n

equations and n unknowns, but

the constraint means we have n+1

equations, but still only n unknowns. Solution of this problem is achieved using the

Lagrange Parameter, a conventional mathematical solution. The result is that we

have to solve n+1 simultaneous linear equations that will produce the set of weights

that minimises the error variance under the constraint that the weights sum to 1. In

addition, we obtain the Lagrange Parameter value, and this is used in calculating the

resulting minimised error variance.

The final set of equations to be solved are known as the ordinary kriging system, and

they describe the set of weights that minimise the error variance under the constraint

that they sum to 1 using the following n+1 equations stylised on the slide above

(actually, simple kriging is shown), where w are the weights, ij are the pair combinations

of known samples, m is the Lagrange Parameter, C is the covariance and 0 subscript

denotes the sample to be estimated.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

ESTIMATION - KRIGING

EARTH

WORKS

pair of samples ij, and a vector of covariances i0.

Function Z at any point.

combination of available data.

The weights are determined by the requirements that the results are not biased

and that the error variance is minimised.

Kriging is an exact interpolator : at the data points the kriging results coincide

precisely with the measured values.

The weighting factors and the kriging estimate do not depend on the sill of

the variogram.

We do not need to know the exact values of the data points themselves to

determine weighting factors. To one sample pattern there is one kriging matrix.

Finally consider the estimation of the value at the centre of a circle from four samples

on its perimeter. If the samples are equally distributed around the perimeter then we

will logically assign equal weight to each sample. Both inverse distance and Kriging

would assign the total weight to each samples.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

ESTIMATION - KRIGING

EARTH

WORKS

However if the samples were arranged such three are grouped very closely, then it

would be illogical to assign the weights in the same manner. The tightly clustered

samples carry very little independent information. However inverse distance weighting

pays no heed to the inter-sample distances, only to the distances to the unknown

point and will still assign of the weight to each sample. Kriging accounts for the

dependency and will ensure approximately the weight is assigned to the lone sample

and the remaining half shared amongst the clustered samples, each receiving 1/6

each.

Simple kriging

So far we have really been describing ordinary Kriging, which is the general form of

Kriging. In addition, we should mention simple kriging. Simple Kriging differs in that it

does not require that the sum of weights sum to 1 but does require strict stationarity

of the mean. The set of equations solved involve a matrix of (n)x(n) covariances and

a vector without the Lagrange Parameter. The weights are still determined to minimise

the error variance. In practical terms, simple kriging is the same as ordinary kriging,

but the extra row and column of 1s is omitted from the covariance matrix before solving

for the weights.

Block kriging

Most of our technical consideration so far has been concerned with the estimation of

point values. Sometimes, though, we may prefer to estimate the average value of a

variable within a block, or local area.

One way of estimating the average value of a block is to estimate many points within the

block and then average them. This is effective, but is not very efficient. Block kriging

achieves the same result by an elegant adaptation of kriging.

Basically, the kriging system is modified to compute average covariances between the

sample points and the block. This only requires changing of the covariance vector D,

not the C matrix. The covariance required is the average of the covariances between

the samples and every point within the block to be estimated. In order to establish the

average point-to-block covariances, we require just a few points to be estimated within

the block, perhaps 4. This means that block kriging is very efficient and convenient

direct estimation, a feature not shared by other estimation methods. Block kriging

should be considered for use more frequently than it is.

Neighbourhood search

A critical choice for any estimation process is to determine how many nearby samples

or neighbours should be used for predicting the value at an unsampled location. For

kriging, this will determine the number of linear equations that need to be solved for

though matrix inversion.

A neighbourhood choice containing a large number of samples will be slow for kriging,

although this is unlikely to be a prohibitive factor with modern computing power.

However, selection of a large neighbourhood with a sparsely sampled data set will also

have a significant smoothing effect on the result. Similarly, small neighbourhoods will

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

ESTIMATION - KRIGING

EARTH

WORKS

have less of a smoothing effect, but if local values are erratic, may not result in reliable

estimates. Therefore, the choice of neighbourhood size will depend on the data type,

the smoothness of the underlying spatial continuity, the data sampling interval relative

to the gridding interval, the clustering of the data and many other factors.

The choice of neighbourhood search parameters is a very important decision in

estimation.

Cross Validation

Cross-validation is a technique which is often used both as a means of comparing the

results of different neighbourhood choices, before selecting final parameters for kriging.

It is a useful tool, but its importance is often over-emphasised, and its results can

sometimes be extremely misleading.

Cross-validation involves removing a single point from the data set at a time and then

re-calculating it using only the remaining points and the selected kriging parameters

and covariance model. The kriging estimate is then compared to the original value

and the difference calculated. By repeating the procedure for every point, successfully

removing and replacing them in the kriging, a statistical summary of the differences

between all the points and their kriging estimates can be generated. Based on a criteria

such as minimising the error difference between the samples and the kriging estimates,

a neighbourhood choice is then determined.

Unfortunately, cross-validation can be adversely affected by the data arrangement, for

example clustering, and so the results of a cross-validation exercise are subjective and

require careful interpretation. Notwithstanding these reservations, a properly applied

cross-validation study is an important component of any geostatistical analysis.

Variogram modelling

We can see that a prerequisite of kriging is to obtain a variogram which is

representative of the spatial continuity of the phenomenon being studied. In order

to obtain a covariance value for any separation h we need a continuous covariance

function. We can obtain this by fitting a model to our variogram and then using the

model to define the covariance function at any separation by calculation from the model

parameters.

In fitting a model, certain constraints are applicable. If we simply interpolated between

the points on the experimental variogram, the solution of ordinary kriging equations

derived using these numbers may not exist, or if it does exist, may not be unique.

In order to ensure a stable matrix solution in the kriging, all the models used to fit

variograms must have the property of positive definiteness.

Many positive definite models are available with which to describe the variogram shape

for use in kriging. For most purposes, the set from which to select either individual

or combinations of models is quite small, and these standard models are described

below. In addition, a wide range of more exotic models are also available, but their use

is in special circumstances.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

ESTIMATION - KRIGING

EARTH

WORKS

asymptotically and so the practical range is defined as the distance at which the

variogram value is 95% of the sill. This model is linear at short separations, but

rises more steeply and flattens gradually. This model is appropriate for quite

rough phenomena, such as well logs.

The Spherical model is probably the most commonly used model of all. It

has a linear behaviour close to the origin, but flattens off at larger separations,

reaching a sill at the range a. This is an intermediate roughness model. If in

doubt, use this model.

The Gaussian model is also a transition model and is appropriate for the

modelling of extremely continuous (smooth) phenomena. Like the exponential

model, this also reaches its sill asymptotically and the practical range a is

defined as for the exponential model. The Gaussian model is parabolic close

to the origin and has a flexure. Gaussian models should always be used with a

small nugget effect to guarantee stability of the kriging solution. The Gaussian

model is generally not recommended. The presence of a Gaussian model form

on an experimental variogram often indicates some form of smoothing process

has already been applied to the data.

The Linear model is not a transition model since it does not reach a sill, but

increases linearly with h. Sometimes useful as a nested component or for quick

and dirty results.

nugget effect model can be thought of as a moving average type of filter, based

on the statistics of the local samples used in estimation. A very common nested

component.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

ESTIMATION - KRIGING

EARTH

WORKS

* Nugget model +

-sill = 0.4

* Gaussian model

-sill = 1.0

-range = 10

* Nugget model +

-sill = 0.4

*Spherical model +

-sill = 1

-range = 5

*Exponential model

-sill = 1

-range = 1-0

Anisotropy

Anisotropy can be identified in variograms. Anisotropy can take one of two forms

geometric

zonal

Geometric anisotropy is where the variogram computed for different directions has the

same sill but the range is interpreted as different. This can be viewed as an elliptical

spatial correlation. The longer range direction will be in the strike direction of any

spatial features in the dataset.

* Direction 1

-sill = 2200

-range = 12

*Direction 2

-sill = 2200

-range = 5

* Nugget

-sill = 300

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

ESTIMATION - KRIGING

EARTH

WORKS

* Direction 1

-sill = 1600

-range = 7

* Direction 2

-sill = 2200

-range = 7

* Nugget

-sill

Zonal anisotropy is often described in mining examples and corresponds to the situation

where the variogram in one direction has a higher sill than that in another direction.

This may occur because one direction of the variogram is across layers or zones,

whereas the other direction(s) is/are within a layer or zone. Zonal anisotropy is really

an artefact of the way zones are mixed and it is difficult to conceive of a property which

is implied to have some form of independence between different azimuths. A trend is

often mistakenly interpreted as zonal anisotropy.

Trend

The expression of a trend (or drift) in variograms is a source of much misunderstanding

amongst lay geostatisticians. A trend is frequently confused with anisotropy. The two

concepts are completely unrelated. A trend is where the values of the attribute show

a correlation with the coordinates of the samples. The correct identification of a trend

is important when considering the stationarity assumption to be used for the spatial

correlation model.

The presence of a trend causes the variogram to rise very rapidly. In the presence

of a trend, variograms are not a suitable choice of spatial analysis tool. We are no

longer in a situation where the intrinsic hypothesis applies and our simple stationarity

assumptions are clearly violated. Either the trend should be removed during the

analysis or higher order fitting methods specifically designed to deal with the trend in

the spatial correlation model should be used.

Semivariogram

tail:TWT

ms head:TWT

2000.

1500.

1000.

500.

0.

0.

4000.

8000.

Distance

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

12000.

ms direction 1

ESTIMATION - KRIGING

EARTH

WORKS

By the far the most practical solution is to remove the trend, undertake the

geostatistical analysis and prediction on the residuals and then, after prediction,

add the trend back again. A number of geostatistical algorithms actually work in

this way. To estimate a linear trend, apply multi-linear regression of the variable

against the coordinates to obtain the regression coefficients. Compute the trend

from the coordinates using the coefficients and subtract this from the data. Work with

the resulting residual and then add the trend back after final prediction (kriging or

simulation).

The alternative is to use higher order spatial model fitting techniques, so called irf-k

or intrinsic random functions of order-k. For variograms, k=0. With a linear trend,

k=1. Some programs can also work with k=2. The only commercial software package

capable of explicitly working with these non-stationary problems is Isatis, produced

by Geovariances. The fitting is purely statistical (no graphics) and the benefits over

simply subtracting the trend and working with the residual are marginal. For a further

discussion of irf-k, see De Marsily (1986).

The following components of variogram models must be considered when fitting to

experimental variograms:

Nugget effect

Range

sill

Anisotropy

A discussion of the effects of the model type (primarily determined by the slope at the

origin and the nugget effect) will be found later. However the following diagram shows

the consequence of kriging a sparse data set using different range spherical models.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

ESTIMATION - KRIGING

EARTH

WORKS

DEM example

The data shown here are taken from a digital elevation map (DEM) of the south east

of England. The original data was kindly supplied by Nigel Press Associates Ltd from

their EuroDEM datset. The DEM is supplied sampled on a 100 x 100 m cell size, this

has been re-scaled to give a 20 x 20 m cell size by simply dividing the coordinates in

X & Y by 5. An area of interest (AOI) has been chosen to simulate a tilted fault block

trap and is approximately 10.8 x 6.6 km in extent. This is shown as the white rectangle

on the accompanying figure. The elevation data has been shifted to a depth of about

2000 m. The grid definition for the AOI is

x origin 103400

y origin 28200

x mesh 100

y mesh 100

x nodes 109

y nodes 67

A perfect 3D migration with a 100 m Fresnel zone (50 m radius) has been simulated

by applying a 5 x 5 (25:1) moving average filter to the full dataset. The parameters

are equivalent to a velocity of 3,000 m/s and a dominant frequency of 60 Hz (Yilmaz,

1987). An enlargement of the AOI showing the raw and 25:1 smoothed data is shown

for comparison.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

ESTIMATION - KRIGING

EARTH

WORKS

2D seismic acquisition has been simulated with seismic lines running in a grid pattern

approximately N-S and E-W. The nominal line spacing is 2 km with an along line

sample increment of 250 m. An exact fault line defining the updip closure for the raw

data has been added. This is simply the 2,000 m depth contour along the southern

edge of the escarpment.

Exhaustive variograms for the raw and 25:1 smoothed data have been computed and

displayed. The exhaustive variograms appear very similar at first glance, apart from

a slight reduction in maximum variance. However, closer inspection of the short lags

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

ESTIMATION - KRIGING

EARTH

WORKS

shows the effect of smoothing is significant in changing the short offset behaviour

of the variogram very significantly. The inflection near the origin (like the Gaussian

variogram model) is characteristic of a smooth phenomenon. Variograms of seismic

time horizon data clearly exhibit a similar behaviour, even when snapped picks are

used. It is suggested that this is a Fresnel effect and it would be expected to be more

significant for 2D than for 3D seismic, due to migration of the former being unable to

collapse the out of plane Fresnel zone

Support is a sample volume concept. Two measurements are at the same support

if they sample at the same volume size and dimensions in X, Y and Z. Sampling at

different support sizes results in some well known effects (which were noted by Krige,

back in 1951). Specifically, as the support increases so the variance decreases.

Using the variogram and the pseudo-2D seismic lines, the depths have been estimated

on a regular grid of points. The kriged map is the expected value at a given point, and

with it we obtain the standard deivation map which indicates the confidence we have in

the estimate. The standard deviation map follows the data sampling arrangement. Our

confidence in the prediction is high close to a seismic line and reduces away from the

data control.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

ESTIMATION - KRIGING

EARTH

WORKS

Kriging

Standard Deviation

Based on these results, we can test the accuracy of our depth predictions against the

original satellite DEM data. The results are as follows

5x5 vs Raw

Kriged vs 5x5

Least squares

Inverse distance

MAE (m)

5

11

14

13

MSE (m)

8

16

19

19

Variance (m2)

61

248

342

351

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

ESTIMATION - KRIGING

EARTH

WORKS

Clearly all methods give reasonable prediction accuracy but Kriging is best in that it has

minimised the variance of error.

Next we will look at the gross rock volume prediction accuracy based on a cutoff of

2000 m.

The volumes are:

Raw DEM

5x5

9x9

Kriged

Least squares

Inverse distance

265

55

30

55

55

55

296

261

227

201

186

121

0

-12

-23

-23

-29

-54

Clearly we have serious bias in the volume estimation. The next section will seek to

explain what has gone wrong and how we can correct it.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

STOCHASTIC SIMULATION

Uncertainty is subjective. It is not possible to quantify uncertainty, but we can use

geostatistical tools to explore uncertainty. Although we obtain a confidence interval

from kriging (the standard deviation map) there are aspects of its interpretation which

should be considerd carefully. Firstly its magnitude is solely a function of our model

the sill of the variogram to be specific. This is often a subjective decision.

If we accept our kriging error variance maps as defining meaningful values, then we

could obtain an upper and lower bound for, say, the depth to a horizon by simply adding

or subtracting the confidence limit from the kriging estimate (our best estimate) to

obtain a minimum and maximum depth value for the horizon. Having done this, it may

be tempting to subsequently use the resulting best estimate, minimum and maximum

95% depth confidence maps to derive three volumetric scenarios - best estimate, lower

case and upper case.

An example of this approach is given in the DEM example. The kriging result has a

volume of about 190 Mm3 and adding and subtracting 1.28 standard deviations of depth

(giving P90 and P10 depth maps) results in GRV estimates of about 80 - 400 Mm3 ,

depending on the model chosen. The true volume is 261 Mm3, which makes the result

rather subjective and prone to interpretation as to whether the kriged map is a mean,

most likely or P50 map.

So what is the problem? Nothing really, except that we are making an assumption

which is not true. The error variance is related to the confidence of obtaining a depth

value. In other words, it is a confidence interval around an estimated (predicted)

depth. The estimation of a volume or area is a non-linear function of the error of depth

estimation.

Kriging (and other methods of mapping found in commercial software packages) is

a linear, unbiased estimator. The estimate is linear because it is a weightd linear

combination of sample values. It is just like the mean or average calculation but with

the weights adapted to account for the spatial positions of the samples. The resulting

prediction (expected value) is smoother than the actual values and represents our best

estimate of the true surface at that point. It is not the surface itself, which is unknown.

Kriging, like the mean, is unbiased. This means that the average error is about zero,

with positive and negative errors cancelling out.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC SIMULATION

EARTH

WORKS

If we plot the CDF of the data and of the kriged surface we see that the effect of

smoothing is a CDF which is narrower and steeper. The calculation of a volume does

not use all the data: it is an integration between limits of the height of each grid node

above the OWC. Depths below the OWC are ignored. Inspection of the detail of the

CDF shows that the kriged result has fewer samples shallower than the cutoff and that

their average value (column height) is also smaller. Multiplying the two gives the vlume

(each grid node is 1x 104 m2)

%<2000m

Average

Column (m)

GRV (Mm3)

Raw Depth

13.3

30.5

296

Depth 25:1

13.2

27.1

261

Kriged

12.1

22.8

201

The problem lies in the non-linearity of estimating the value above a cut-off from

a surface estimate. Consider the problem in the diagram above. On the left, the

uncertainty never impinges on the cut-off and so the volume is correctly estimated,

but in the example on the right the volume will be zero for the deterministic case and

+0.25 for a Monte Carlo simulation. The deterministic (smooth) case will always

underestimate the volume (bias).

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC SIMULATION

EARTH

WORKS

.

In the cartoon above, a series of measurements have been made of the sea depth

(yellow dots) between London and New York. A bathymetric cross section has been

constructed by interpolating between these points (kriging, shown in green). The error

envelope around the kriging is shown as the string of sausages (red dashed line) and

is zero at the data points. In order to navigate a submarine between London and New

York as close as possible to the bottom, the kriged result is the correct line to follow,

although a risk averse captain will want to take account of the error and pilot a course

closer to the shallower error line.

However, consider measuring the length of the kriged line, or even the error surfaces,

and using the longest of these lengths to establish the length of cable required to lay

a telephone line between New York and London. Clearly the estimate will always be

too short and the ship run out of cable about where its position is shown in the cartoon.

Different estimations require different methods of prediction.

Another analogy of the same type of problem is the classic measurement of the length

of coastline of the mainland UK. The length of the coastline depends on the length of

unit measurement used to determine it. As the unit of measurement decreases in size,

so the length increases. This is now a well known example of a fractal relationship,

and the idea is closely related to the subject of simulation. This is known as Steinhaus

paradox

Using a high side and low side map based on smooth estimation can never give

correct volumes. The variogram of the kriged surface is not the same as the modelled

variogram estimated using the observed data. The basic RF assumption of ergodicity

implies that a smooth mapped surface cannot be a possible realisation.

Secondly, to assume that all depth values are positive or negative at the same time is

implausible - its extremely improbable that all the depths are simultaneously high or

low. Therefore the high/low volume estimate approach must exaggerate the range of

GRV

The common method of determining the spill point and then using higher and lower

contours on the same deterministic map is tantamount to assigning a degree of fill

range as being the volumetric uncertainty range. This is erroneous.

In order to correctly estimate volume we need to use the geostatistical technique of

conditional simulation. This requires that we generate possible surfaces which honour

the observed data points and the variogram model. This cannot be achieved uniquely,

and so we are required to generate many equi-probable realisations of the depth map,

estimate the volume for each and construct a GRV expectation curve. Additionally, by

converting each grid node to a Boolean result of above the OWC, and adding many

realisations together, we can also compute the isoprobability closure map, expressing

the likelihood at each node of being above OWC.

As in the simple case of tossing a coin or in the random function examples shown

earlier, we may not be interested in the expected value but in an outcome or realisation

of the variable. In the case of depth, this is the case if we wish to estimate volume

above a cutoff. The reason is that the estimation process obtains accuracy at the

expense of the higher order statistics. This means that the CDF is not reproduced.

Instead, the CDF of the estimate is narrower and steeper than the that of the raw data

(which is a realisation). However, the CDFof the estimate is centred such that it is

unbiased.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

Introduction to Geostatistics

Stochastic Simulation

WORKS

EARTH

STOCHASTIC SIMULATION

EARTH

WORKS

at a threshold and only consider part of the curve. The predictions will then be biased.

The solution is to work only with realisations for any procedure that requires the CDF

to be truncated. Because each realisation is not unique, we must generate many

realisations, analyse each one and then look at the average of the computations.

In the DEM example, we can generate realisations of the depth map. 100 maps were

computed, and six are illustrated. We compute the gross rock volume for all 100 maps,

sort the results into magnitude order and plot the CDF of GRV. The average of these

volumes is then our best estimate of the GRV.

In addition we can estimate the probability of a grid node being above closure. For

each map we test each node. If the node is above contact we set it to a 1, otherwise

to 0. We then sum the 1/0 maps and obtain the probability of the nodes being above

closure. This is known as the isoprobability closure map.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC SIMULATION

EARTH

WORKS

Simulation Algorithms

There are three important criteria to be met in generating acceptable realisations,

which are:

Reproduce the CDF (which will ensure the mean, variance, histogram and other

statistics are honoured)

Reproduce the spatial correlation behaviour represented by the variogram

model.

For the continuous variable types considered so far in this course, there are three

basic algorithms by which a realisation can be generated. All of these methods are

maximum entropy solutions. In other words they generate a realisation that is the

most disorganised result whilst still reproducing the required spatial correlation function

defined by the variogram model.

Turning Bands

Turning bands is the oldest algorithm implemented for the generation of conditional

realisations. It is based on exact 1D simulations along lines which constitute a

regular partitioning of a 2D or 3D space. The method is very fast, but generates nonconditional realisations. Each realisation is then conditioned to the actual data points

by extracting the simulated values at the positions of the original data. These values

are kriged and the resulting surface subtracted from the non-conditional realisation

giving zero values at the data points and a simulated residual everywhere else. The

simulated residual is then added to the surface obtained by kriging the real data points

to give the conditional realisation.

The additional kriging step makes turning bands comparable in speed to other methods

when used for conditional simulation. Turning bands suffers from a tendency to

generate linear artefacts (related to the lines constructed to partition the space) and this

can be problem. The solution is to increase the number of lines used for partitioning

but this can slow down the method sufficiently as to be impractical.

Spectral Methods

A modern and very fast alternative to the turning bands method is to use a fast Fourier

transform (FFT). The covariance function, which is the spectral density in xyz space

equivalent to the autocorrelation function in geophysics, is transformed using the FFT.

In the frequency domain the covariance is than the amplitude spectrum with a zero

phase spectrum. A non-conditional realisation is obtained by adding random phase

and performing an inverse FFT. Conditioning is by klriging, as in Turning Bands.

By saving the weights for the kriging, the method can be made very fast. It also avoids

the artefacts of turning bands.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC SIMULATION

EARTH

WORKS

SGS is the most straightforward algorithm for generating conditional realisations. A

simple sequential method was encountered when we constructed the random function

models V(x) etc. SGS is now regarded by most vendors and geostatisticians as the

standard method of generating conditional realisations.

The method is very simple. Randomly select a grid node and krige using the input

data. This will give an expected value Z* (the kriged result) and a variance 2 (or

usually the standard deviation). Assuming a Gaussian distribution, draw a random

number from the distribution N[Z*, 2] and place the random number on the grid.

Repeat the above procedure at a new randomly selected grid node. However, when

kriging the next position, the previously simulated node is included as though it were a

real data point. Thus the realisation is sequentially generated and converges to have

the appropriate CDF and spatial correlation function.

Clearly, being based on kriging, SGS is conditional by construction. There are detail

considerations in the algorithm. For example, the speed is usually improved by moving

the real data onto the nearest grid as the first step. Including too many simulated

nodes will swamp the original data, resulting in a simulation with uncorrelated noise,

so the number of previously simulated is commonly a user-defined control parameter.

The method is strictly only valid for simple kriging and so some initial normalisation

steps are usually required to satisfy this condition. Finally, the long range behaviour

of a variogram model is best reproduced if the early grid nodes randomly selected are

forced to be far apart.

Simulated Annealing

Although rarely used for something as straightforward as a conditional realisation

based on a variogram model, simulated annealing is a very adaptable (but slow)

method which deserves a mention.

To generate a conditional realisation on a regular, usually the original data are moved

to the nearest grid nodes and then the remaining grid nodes filled in with random

numbers so as to recreate the desired CDF. However, this then results in a realisation

comprised entirely of uncorrelated random noise.

The conditioning to a spatial function defined by the variogram model is generated

by a process rather similar to how materials cool. Firstly, the variogram of the grid is

computed and the difference between this and the required variogram is calculated.

The required variogram is known as the objective function. Next, two grid nodes are

selected at random and their values swapped over. (Grid nodes where real data

values are were placed at setup are disallowed in this selction.) Afte the swap has

been made, the variogram is recalculated and compared to the objective function. If

the result of the swap is to reduce the difference between the calculated variogram and

the objective function then the swap is retained. Otherwise the swap is rejected and

the values returned to their previous positions.

The procedure of swapping and comparison with the objective function is repeated until

an acceptable match between the calculated and desired variogram is obtained, at

which point an acceptable realisation is generated.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC SIMULATION

EARTH

WORKS

The method is very slow and can suffer from periodic artefacts, depending on how the

calculations are made. For straightforward generation of conditional realisations, any

of the alternative methods are preferable. The specific advantage of the simulated

annealing technique lies in its flexibility. Providing an objective function can be defined,

any information can be used for conditioning.

If we map a small number of data points using different variogram model types, we

obtain maps of the expected value. As shown in the accompanying diagram, these

maps are quite simple in form and fairly similar, all comprising of bullseyes around the

data points, dropping away to the mean value at greater distances.

However, if we generate a realisation from each variogram type the results are very

different. The realisation generated from a nugget variogram model comprises

uncorrelated noise which is standard Normal in its distribution. Then, in varying

degrees of smoothness, we progress through the exponential, spherical and Gaussian

variogram models. The PDF in each case is standard Normal, what differs is the

organisation of the points. If we cut the whole map at a given threshold, every map

would give the same volume (approximately), but the connectivity between the islands

where the samples are above the contact would be very different. The exponentional

would give rise to a large number of small, disconnected islands whereas the Gaussian

would result in a few large islands. The spherical model result would lie between these

two cases. This can be also looked at in the cross-sections through the realisations.

Kriging

Simulation

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC SIMULATION

EARTH

WORKS

Cross Sections

Connectivity

We can measure connectivity through the use of multipoint statistics. The variogram

is a two-point statistical measure. Journel et al (2000) discuss the difference between

Kriging and simulation in terms of connectivity.

A semivariogram or covariance model represents a measure of variability or

connectivity between any two points u and u+h separated by a given vector h (twopoint statistics). We can define connectivity by looking at indicator values, where a

threshold Zc is exceeded or not. We can compute the probability of n locations along a

given direction to be all valued above the threshold using:

The greater the spatial entropy of the field, the faster the decrease to zero of the

probability P(n;Zc) as n increases. Because of smoothing, kriging will generate

surfaces with too large P-connectivity functions for certain threshold values Zc. In

other words, calculating connectivity through any result which is Kriging-like in its

computation (ie a minimisation) will result in connectivity being significantly overestimated.

If we consider these statements along with the underestimation of volume previously

described them we have the following, apparently contradictory consequences of

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

STOCHASTIC SIMULATION

EARTH

WORKS

(for a depth /OWC case)

than it really is

The last point can be illustratd with 2D and 3D seismic mapping. Structures or fields

mapped with 2D seismic generally appear more compartmentalised when subsequently

mapped with 3D seismic.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

With the adoption of geostatistics and stochastic modeling it has become very

popular to refer to conventional maps as deterministic and to describe the results

of geostatistics as stochastic. Deterministic maps which have been calculated with

different parameters (eg different depth conversions) are sometimes referred to as

realizations. This is incorrect.

via a functional relationship. Examples are Darcys Law, chemical reaction rate

and Newtons laws of motion

Best Estimate is a model which minimises the prediction error. This includes

kriging, regression models and Wyllies Equation.

Stochastic is a model which generates non-unique solutions and honours

higher order statistics such as spatial correlation and the histogram.

All estimation or prediction requires the use of a model. Very few earth science

processes are sufficiently well understood to allow the use of deterministic models in

estimation.

Deterministic model

input parameters via a process model.

Error on output is proportional to error on input

Few (no?) earth science processes can be modelled deterministically. Yet.

Deterministic reservoir modelling schemes are the subject of research and

development.

Prediction accuracy is obtained by estimator shrinkage - output range is less

than input.

Estimator shrinkage means prediction accuracy is at the expense of higher

order statistics - reduction of variance

Best estimate models are for prognosis, not higher order statistics.

Origin of uncertainty

Uncertainty arises because of

specified in variograms via the nugget component

Non-exhaustive sampling, which is the uncertainty we associate with spatial

interpolation. This is the principal objective of kriging.

Uncertainty = (Z* - Z) and is therefore not quantifiable. In other words, it is

the difference between a known and unknown value, which by definition is not

known.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

at an unmeasured location. The method of analysis should capture the essential

features of the data for prediction. It should not necessarily treat the data as

exact, nor should unnecessary cosmetics that may have no physical basis, such

as smoothness, be imposed on the prediction method. (Laslett and McBratney,

1990)

Uncertainty arises from the decision of a model.

Cui and Blockley (1990) suggest that uncertainty contains three distinct components :

fuzziness, incompleteness and randomness (FIR).

Fuzziness is the imprecision of definition. It occurs because of the way we choose

to express the parameter with which we are dealing. For example, we could describe

top reservoir depth in a well as :

(a)

Very deep

(b)

3,200 m

(c)

3,226.59 m

These three descriptions have varying degrees of precision associated with them. The

level of description (or fuzziness) must be chosen appropriately for the problem so that

extremes of spurious precision or unhelpful vagueness are avoided.

The traditional way of expressing a crossplot is by fitting a function (often a straight

line) through the points by a regression analysis. The function then expresses the

relation but the obvious scatter has to be accommodated. This is done by modelling

the difference between the function and the actual measurements as a random

quantity.

Incompleteness exists because there is that which we do not know. Since a model

is by definition not the reality, it is only partial and is incomplete. Incompleteness falls

into two groups : that which cannot be foreseen, which would include the existence of

phenomena which were previously unknown, and causes which are foreseeable but

are:

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

(a)

(b)

(c)

ignored in error

ignored by choice

ignored because of lack of resource

All of the above are sometimes referred to as bias. Incompleteness resulting from lack

of resource is dramatically illustrated when a comparison of interpretations on based

2D and 3D seismic is made.

Randomness is the lack of a specific pattern in the observation or series of

measurements. If the variation in some measurement cannot be explained in terms of

a deterministic cause and effect it is usually referred to as being random. Statistics and

probability theory is about deriving patterns in populations of data where there is no

discernible systematic relationship.

Consider the depth to top reservoir in a well. If we make many measurements of the

depth under standard conditions they will vary because it is not possible to repeat

the measurement experiment exactly. If we plot a histogram of the results and fit a

curve such that the area under the curve is one then we have a probability density

function (pdf). The depth can now be thought of as a random variable. The pdf may

be represented by characteristics of the distribution such as the moments. The first

moment is the mean (expected value) and the second moment is the variance (square

of standard deviation).

Stochastic Model

order statistics such as the variance, CDF and spatial correlation properties.

Non-linear properties of input can be investigated without risk of bias. These

operations include truncating the CDF with a cutoff or fluid flow calculations.

Uncertainty can be parameterised and explored (but not quantified since this

requires knowing Z - Z*)

The realisations are equi-probable.

Best estimate and stochastic models are complementary. We choose one or other

depending on our estimation objectives: best estimate for prediction/prognosis

(linear problems) or stochastic for volumes/connectivity/fluid flow behaviour (nonlinear problems). The best estimate is the average of the (infinite) set of realisations.

Different best estimate cases produced with different parameters (for example,

different depth conversion layers) are not realisations. In geostatistics, we refer to the

best estimate as kriging - a minimum error variance estimate. In statistics, the best

estimator is the mean.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

It was the observation that cutoff predictions based on best estimators result in bias that

led Danie Krige to his thesis and gave rise to a new science of geostatistics.

If we consider a simple cross-plot relation, such as the permeability-porosity example

shown below, we can predict permeability at unmeasured locations (such as on a

porosity log) by obtaining the linear regression line.

The linear regression equation is the best predictor of permeability from porosity

under certain assumptions, which include Gaussian distribution of error.

It is common to associate a cutoff on permeability with a porosity cutoff. In this

instance 1mD = 9.5% porosity

When we apply the porosity cutoff to the cross-plot, we are counting the

samples in quadrants B+D. This defines 92.5% of the samples as belonging

to net reservoir.

The permeability cutoff really intends to classify samples in quadrants A+B as

net reservoir, which would be 82.5%.

It is only correct to apply the cutoff procedure by transform if r 1!

Cutoff calculations require higher order statistics to be honoured in order to

avoid the risk of bias.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

SUPPORT

SUPPORT

The support of a measurement refers to the volume over which the

measurement has been made. For example, a core plug porosity is made

over a volume of about 20 cm3 of rock. A log measurement, such as bulk

density, is made over a larger volume of rock, dependent on the tool

resolution and depth of investigation. Seismic data measures the earth

response as an average over a large volume determined by the frequency

content and bandwidth.

The support of a measurement is a key issue in geostatistics and accounting

for changes in support is considered a key step in mining applications of

geostatistics. Unfortunately the support question is largely ignored in the oil

industry and very few software packages provide tools for correctly handling

prediction with change of support.

If we krige a grid of values using average porosity data then strictly we are

making our estimates at the same scale of support as our measurements.

During scale-up in reservoir models, the regular support obtained from well

log measurements is often changed to an irregular support within the cells of

the model, due to varying cell sizes, well geometries and the algorithms used

for simulation and estimation.

As the support of a measurement

increases so the variance

decreases. For a linear

averaging property the mean is

generally unaffected by a change

of support.

Changes in support can change

form of the functional relation

between two properties and this

is particularly the case when the

two properties average in

different ways such as porosity

and permeability. Change in

support of seismic measurements will modify or even destroy a relation

between a seismic attribute and property such as porosity.

The figure below is from Partyka et al (2000) and shows how the relation

between acoustic impedance and effective porosity changes with effective

scale of measurement, both in seismic bandwidth and reservoir thickness.

Seismic properties scale according to a Backus average when the effective

layers are below the resolution.

The key statements we can make concerning an increase in the support of a

measurement are:

All rights reserved

SUPPORT

Modelled effects of seismic backus averaging and reservoir thickness from Partyka et al 2000

The histogram (PDF) becomes narrower about the mean

The CDF becomes steeper and narrower

The sill of the variogram decreases

The range of the variogram increases

The effect of the support of the volume over which we averaging will be

greatest for data which have no spatial correlation. Classical statistics then

tells us that the standard deviation of block averages is inversely proportional

to their area or volume. Data sets where extreme values are poorly

connected have high entropy, conversely where extreme values are well

connected we have low entropy. Changing support therefore changes the

apparent connectivity (entropy). Note that most geostatistical methods such

as simulation tend to maximise entropy.

The volume scale change from a well log sample to a reservoir model cell is in

the region of 500,000 : 1. Typically between 10 and 300 log samples are

averaged into a cell. This depends on the cell orientation or, more usually, the

deviated trajectory of the well passing through the cell. Scale-up in reservoir

modelling is largely arbitrary and the cells are unlikely to contain the

appropriate porosity and variability for the support defined by the cells using

currently available methods.

All rights reserved

SUPPORT

Although we will discuss scale-up in a later section, note that resampling into

cells (nearest point) leaves the support of the measurement unchanged.

Averaging is a smoothing operation and changes the effective scale of

measurement.

The scale change between core data and the reservoir model cell is very

large, perhaps as much as 100 million : 1. With properties such as

permeability it is important to remember that they do not average

arithmetically and so the relation between porosity and permeability defined at

core scale does not necessarily apply at log or cell scale.

For seismic data, the seismic attributes or impedance estimates generally

have a lower resolution than the cells of the model. For this reason seismic

generally provides us with estimates of a zone average property. Currently,

the standard algorithms in Petrel used to constrain estimation of properties

such as porosity using seismic data are incorrect.

Support Effect on Variograms

The change of support of a measurement can be observed on variograms and

is widely studied in the mining industry where ore grades are estimated at one

scale (core) and used to predict average grades over large blocks which are

actually mined out. Smoothing of data gives rise to two effects in variograms:

The range increase slightly as the support increases

estimate the missing variance

and true point sample

variogram. To the right are the

DEM variograms we met

earlier, with and without

smoothing. The smoothed

data variogram (blue) has a

lower sill and inflexion at the

origin.

Zooming into the origin, we can

clearly see the change of

support. The points marked in

yellow are the points which

when a straight line is fitted to

them gives the steepest slope

and most negative intercept

with the Y-axis. The fitted line

has an intercept of -107, the

missing variance caused by

the smoothing.

3000

2500

2000

1500

1000

500

-500

0

1000

2000

3000

4000

5000

Raw

25to1

800

1000

6000

7000

8000

1400

1600

9000

10000

Origin

1500

1000

500

-500

0

200

400

600

Raw

1200

25to1

All rights reserved

Origin

1800

2000

SUPPORT

Our true variance is estimated

by adding this to the observed

variance. It turns out to be

slight underestimate the true

variance difference is about

140.

1500

1000

500

y = 1.3584x - 107.58

0

in which observations can be

corrected for the support

effect. The basis of most

methods is to leave the mean

unchanged and correct the data about the mean using a variance adjustment

factor f, defined as the block to point variance ratio.

-500

200

400

600

800

Raw

25to1

1000

1200

Origin

1400

1600

1800

2000

Linear (Origin)

Affine correction

Indirect lognormal correction

Stochastic simulation based estimates

The affine correction is based on the data quantiles. The quantiles are

transformed (rather like a normal score transform) about the mean using the

following formula:

q =

f (q m ) + m

to apply, is not available in current reservoir modelling software.

The kriging implemented in Petrel (and most reservoir packages) is point

kriging. Geostatistics allows for block kriging, where the estimates is a linear

average over the cell volume. Changes of support are important in estimation

and ignoring them is a major limitation of current reservoir modelling software

Porosity and permeability scale changes could be handled much better if the

necessary tools were available. A particular problem is constraining

properties such as porosity to seismic data. Petrel appears to support this

through the collocated co-simulation and locally varying mean algorithms but

in fact their use is incorrect because they fail to take into account the severe

support differences between the upscaled cells used for modelling and the

seismic observations which are closer to zone averages. This will be looked

further in the petrophysical modelling section of the course.

All rights reserved

EARTH

WORKS

Entropy

Although geostatistics relies heavily on the variogram we should consider that variance

does not necessarily act as an adequate descriptor of uncertainty. An alternative is the

quantity entropy.

Entropy H(X) is a statistic that quantifies the intrinsic variability of some variable X and

can be computed from the pdf p(X):

Note that entropy is computed from a histogram, not from individual values.

Consider a categorical value X = {shale,sand}

If P(X) = {0.9,0.1} then H = 0.325

If entropy is reduced, then there is now less disorder, less uncertainty and therefore

more predictability.

If we consider a univariate variable X = {-2, -1, 0, 1, 2}

o Variance = 1.12

o Entropy = 1.10

If P2(X) = {0.09, 0.20, 0.42, 0.20, 0.09}

o Variance = 1.12

o Entropy = 1.44

Variance is a measure of deviation from central tendency and is not always sensitive

to uncertainty. P1 and P2 have the same variance but P2 has larger entropy and is

therefore more uncertain (Mukerji et al, 2001).

The bounded pdf with maximum entropy is the uniform distribution. If the minimum and

maximum bounds of a uniform distribution are a and b then the entropy H is ln(b - a).

This decreases to - as the interval [a, b] becomes narrower, corresponding to greater

certainty. For a given fixed variance, the unbounded pdf that maximises entropy is the

normal (Gaussian) distribution (Journel and Deutsch, 1993).

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

Bayes Theorem

Consider the following problem:

A man is knocked down and injured by a taxi cab in a hit & run incident (the

driver doesnt stop).

In the city where it happens, taxis are either green or blue. 85% are green and

15% are blue.

A witness reports seeing the man hit by a blue cab, but when tested later during

the court case the witness was only able to distinguish correctly the colour of a

cab 80% of the time.

What is the chance that the cab was blue?

The solution to this type of conditional probability can be found by applying Bayes

Theorem, named for Thomas Bayes, a 19th century English Clergyman. This can be

stated as:

Which means that the joint probability P(A|B) of both A and B occurring is equal to the

conditional probability P(B|A) that B will occur given that A has occurred multiplied by

the probability P(A) that A will occur. By a series of further steps (see Davis, 1986) we

can deduce the following general equation

The notation and terminology used with conditional probability is rather confusing, so a

worked example may help. We will illustrate the application of Bayes Theorem using

the taxi problem posed above.

The confusion with probabilities is usually people do not work out the outcomes

properly in the form of a tree and then work the probabilities backwards. In the

taxi example, there is a temptation to think of four outcomes, trying to decide if the

witness was right or wrong and which colour cab hit the man. In fact there are just

two outcomes. The man was hit by a cab and it can only be blue or green. The

probabilities from the proportion of cabs of a particular colour in the total population of

cabs are

Hit by green cab (B2 = 0.85)

It really was green, P(A|B2)P(B2) = 0.20*0.85 = 0.17

Therefore the chance it was blue = 0.12 / (0.12 + 0.17) = 41.4%, which amounts to

normalising these two outcomes to sum to 1.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

Full Co-Kriging

The external drift approach described in the previous is largely based on linear

regression and indeed follows many of the assumptions used in linear regression. Cokriging is the simultaneous estimation of two or more variables under an assumption of

a joint spatial model of covariance.

Recall from our basic statistical theory that the correlation coefficient is simply a

normalised expression of the covariance between two variables. If we consider the

covariance between two variables, we perhaps think of a crossplot in order to visualise

the relationship. A conventional crossplot plots the value of one variable against

the value of a second variable measured on the same sample. In other words, the

conventional correlation coefficient (or covariance) is measured on a zero lag (h=0)

between the two variables. It is therefore the first point on a cross-variogram between

the two (or more) variables. We can compute the covariance between variables for

any other lag by taking pairs a vector h apart where one value of a pair is from one

variable and the second value of the pair the other variable. In doing so we obtain

the cross-covariance (or cross-semivariogram if we choose). This describes the joint

spatial behaviour of the two variables as a function of lag. Note that cross-variogram

can be negative as well as positive, because correlation between variables can also be

negative or positive.

In addition, we also need to consider the variogram or covariance function for the

variables themselves, in the usual manner which we have seen so far in the univariate

case. Thus to map a pair of variables simultaneously we require three variograms

that of the primary and secondary variables and the cross-variogram. For three

variables we require six variograms primary, secondary and tertiary plus primarysecondary, primary-tertiary and secondary-tertiary cross-variograms.

When mapping two variables simultaneously, we use the cross-variogram to reduce

the uncertainty in mapping each variable using the other variable and use the individual

variograms in the usual way. The matrix inversion includes all these covariances

and results in weights for mapping primary and secondary variables being derived

simultaneously.

Full co-kriging is most appropriate for mapping variables such as porosity and

permeability. The procedure is as follows:

between primary and secondary data.

Fit authorised joint model to the variograms.

Co-krige.

No approximation is made about the correlation coefficient or spatial

dependence of the cross relation between the variables. Rolls-Royce

integration

Uncertainty of primary and secondary data is accounted for

Uncertainty of cross-relationships are accounted for.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

to estimate unless there is a lot of primary data.

Modelling can be difficult, especially if more than 2 variables. Apparently

acceptable models may not have kriging solutions.

If there is plenty of primary data, secondary data contributes very little and will

receive very little weight or no weight in the Kriging system.

The significance of the last point has led to a simplified co-kriging approach which will

be described next.

Collocated cokriging

It was observed in cokriging that where the secondary variable was available at the

target estimation position, then the only secondary data point which receives any

weight in the cokriging system is the value at that location, all other secondary points

have a weight of zero (which is generally logical, provided the secondary variable does

not contain a noise term). Therefore including secondary data in each estimation,

other than the secondary point coincident with the target node, is superfluous. This is

strictly applicable to problems involving surfaces or attributes obtained from 3D seismic

data.

variable at the target node only is retained

The method is very fast

Similar to, but more rigorous than using Kriging with external drift because a

joint model including the primary and secondary cross terms is used

The variogram of the secondary variable is not required in this approach (because

there is no requirement to estimate the secondary variable since it is required to be

known at all target locations)

Markov-Bayes hypothesis.

There are a number of practical considerations which have been noted in the previous

methods.

appropriate as we can only estimate locations where the secondary variable is

known

Interpolating the secondary variable to all target locations and then applying

collocated or external drift methods ignores the uncertainty in predicting the

secondary variable

Variography for co-kriging can be difficult because of too few data and the fitted

models may then be somewhat arbitrary.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

Markov-Bayes model we can achieve some of this.

problems

The secondary variogram is usually straightforward to compute (lots of

samples)

The primary variogram is a scaled version of the secondary variogram - the

scalar is the variance ratio

The cross-variogram is computed from the variance ratio and the linear

correlation coefficient, r.

When r = 0.999, we have kriging with external drift

When r = 0.001, we have primary variable only

Correlation coefficient, r

The variograms for the primary variable and the primary to secondary cross-variogram

are then calculated as follows:

Cross-covariance model

The effect of the weighting function on the primary and secondary variables under the

Markov-Bayes hypothesis is shown for various correlation coeffecients. When the

correlation is effectively zero then the primary variable only is kriged, the secondary

receiving no weight. When the correlation is 1 the primary only is used at a primary

variable location, but the weight applied to the secondary builds up very rapidly when

moving only a short distance from the primary data point.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

by Doyen et al, 1996. In this method, which requires simple kriging and therefore strict

stationarity, the results are obtained by combining the primary and secondary kriging

and standard deviation maps post calculation, blending them using the correlation

coefficient and the following equations

Kriging estimate

Standard Deviation

The advantage of this approach is the computational speed is very high and the

blending, controlled by the correlation coefficient, can be done on the fly.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

This case study shows the application of geostatistics to integrate a relative acoustic

impedance slice from a 3D seismic survey with porosity data at the wells. The example

is used by Geovariances as a demonstration data set. Originally the case study from

Amoco was presented in Yarus and Chambers (1994) using only the external drift

method. I am grateful to Richard Chambers for providing me with an original copy of

the full data set to work up the examples shown here.

The full data set comprises 55 wells at which porosity, thickness and porosity-thickness

are measured. A 3D seismic horizon slice through a relative acoustic impedance

volume is available.

Initially we are going to demonstrate some procedures with a subset of the data

comprising just 7 wells.

10000

8000

6000

4000

2000.

0

0

2000

4000

6000

8000

10000

Variogram analysis

Variogram modeling

The variogram is difficult to estimate and so we have cheated by stealing the variogram

from the 55 wells later on in the analysis.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

The result of mapping just the seven wells is shown first, along with the confidence we

have in our estimate, shown by the standard deviation map.

Next we will integrate the seismic information. We start by finding the nearest

impedance values to each well and then cross-plotting the porosity thickness against

the seismic impedance, as we did for the external drift exercise.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

correlation. This can done with a calculation of spurious correlation using a Student

T test, a standard statistical test (see Kalkomey, 1997). A Java applet to make this

calculation can be found at www.sorviodvnvm.co.uk. This is very simple to calculate

in a spreadsheet as well, and the formula is given below. 3 cells are set up for

entering the values for number of wells, correlation coefficient and number of attributes

considered. A fourth cell contains the formula to calculate the probability of spurious

correlation (Psc). The calculation cell contains the following entry :

=1-(1-(TDIST((R * (SQRT(Nwells-2))/(SQRT(1-R^2))), Nwells-2, 2)))^Nattr

where Nwells, R and Nattr are the 3 cell references where the values are entered. The

Psc may be expressed as a percentage.

Next we use collocated Kriging under the Markov-Bayes hypothesis to integrate the

seismic impedance into the estimation. We can easily define the variograms from the

seismic impedance data and then use the Markov-Bayes relation to define the primary

and cross variograms.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

Based on the spurious correlation calculation, we can also define the confidence limits

and thus the level of weighting we should apply to the seismic. The 95% confidence

interval on the observed correlation suggests the actual correlation could be as low

as 0.635 or as high as 0.991. Adjusting the weighting gives two alternative results,

representing a low and a high case confidence in the seismic impedance to help us

map porosity-thickness.

We can illustrate the effect of changing the weight on the mapped result by

systematically changing the correlation coefficient used in the Markov-Bayes and this is

shown next.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

Introduction to Geostatistics

WORKS

EARTH

EARTH

WORKS

Finally we repeat the procedure, but this time using all 55 wells. The first observation

is that Kriging with all the wells results in a reduction in uncertainty, as shown by the

standard deviation map.

Regenerating the cross-plots but with all 55 wells shows we were misled by the 7

wells. There is no correlation of thickness with impedance at all and the best relation is

between impedance and porosity.

show that it is a quite reasonable choice of model, and also to show the primary,

secondary and cross-variograms.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

EARTH

WORKS

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

INTRODUCTION

Geostatistics has found a range of applications within the oil industry in the

last 20 years. Initially it found favour in depth conversion and volumetrics,

then in combining seismic attributes with reservoir properties at wells. But by

the far the most significant application has been the development of the 3D

stochastic reservoir model.

In the simplest sense a reservoir model is simply a set of grids, one for each

reservoir layer, containing estimates of reservoir properties. The grids may be

irregular and their geometry is often complicated by faults and structure. But

a 3D stochastic reservoir model is far more than that, providing as it does a

platform for the integration of geophysical, geological and engineering

knowledge and the ability to view the results in 3D. These are geologic

models which include geostatistical tools in their construction.

From a geostatistical perspective, a good reservoir model has:

A spatial correlation model

Honours the CDF and spatial geometries through realisations

Can generate best estimates and realisations of poro-perm data

Many reservoir models are too ambitious in their scope. The principal

problems are defining a grid and layer definition at too high a resolution. This

leads to large models, well beyond what can be input to fluid flow simulation,

which are also slow to simulate. The second problem is in specifying to many

facies. It is important to try and cut down the number of facies or flow units to

the fewest which will capture the essence of the reservoir. This does not

necessarily result in large, average blocks in the reservoir. A thin interval may

be important if it has extreme permeability characteristics a low permeability

interval may act as a vertical barrier, a high permeability streak may give rise

to early breakthrough. Remember also that two different facies types may

have very similar flow properties, so it may be possible to combine them.

Caution should be exercised here as facies with similar flow properties may

have different geometries which may need capturing in the model.

There are five basic steps in constructing a geological model:

Grid and layer definitions

Scale-up of well logs into grid cells

Categorical prediction of facies or flow units

Continuous prediction of petrophysical properties within facies

All rights reserved

FRAMEWORK

For most 3D reservoir models, the first step is constructed directly from the

depth converted seismic interpretation. Few reservoir models include

uncertainty in the structural framework and this is a major limitation of the

approach used currently in many companies.

The zones are the main intervals representing the geological intervals. They

will usually be mapped as isochores between the depth converted seismic

horizons.

The layers are the fine detail sub-dividing the zones. The layers can be

constructed within each zone as:

Proportional

Base or top conformal

Fractions eg 1,2,1 proportions

The grid and layer definitions are generally selected in relation to layer

characteristics of the geological setting and are a compromise between

capturing significant detail and creating an over large model with too many

cells.

Segments are areas of the model defined by faults. The main use of

segments is to allow different fluid contacts in different parts of the field

The simbox is the name given to the layering after flattening the model to

remove faults and structural shape. Simbox geometries are essentially

continuous within the layers and used during the mapping and simulation of

the geology. They can also be selected as a visualisation option.

All rights reserved

A question which arises is how thin to construct the layers. One approach is

to build coarse layers. This gives simple models which are quick to compute.

However it should be noted that this involves cell averaging which may be

quite severe and is not the same as upscaling for permeability.

The alternative is to build fine-scale layered models and then upscale after

estimating or simulating the geological facies and petrophysical properties.

This has the advantage of offering more rigour for the treatment of scale

changes of permeability and allows sensitivity to upscaling tobe tested but it

may result in large models which are slow to load, save and compute.

All rights reserved

STRATTON FIELD

STRATTON FIELD DATA SET

The data-set used here as the training set for stochastic modelling is the

Stratton Field 3D seismic and well log data package, prepared by the Bureau

of Economic Geology, Austin, Texas, USA (Levey et al, 1994).

Stratton Field is an on-shore

gas field producing from the

Oligocene Frio Formation in

the NW Gulf Coast Basin,

Gulf of Mexico. The Frio

Formation is a sediment

supply dominated

depositional sequence

characterised by rapid

deposition and high

subsidence rates and forms

one of the major

progradational off-lapping

stratigraphic units in the basin. Commercially, the Frio Formation of Texas is

volumetrically the largest gas producing interval of ten major depositional

stratigraphic packages in the Cenozoic of the Gulf Coast

Basin.

Gulf Coast Basin (from Levey et al, 1994)

The top of the middle Frio formation is about 1200 ms, corresponding to a

sub-sea elevation of around -4,500 ft. There is little faulting in this interval and

the formations relatively undeformed and flat lying. Reservoir facies of the

middle Frio are interpreted as multiple amalgamated fluvial channel-fill and

splay sandstones. The composite channel fill deposits range from 10 to 30 ft

All rights reserved

STRATTON FIELD

thickness and show either an upward fining or a blocky log profile. The

composite channel deposits can be up to 2,500 ft in width. Splay deposits

show typical thicknesses of 5 to 20 ft and are proximal to the channel

systems. Porosities in these fluvial reservoirs range from 15 25 % with air

permeabilities of less than 1 to greater than 4,000 milliDarcies.

There are a total of 21 wells in the field. A well display for each of the wells is

included at the end of this section. Wells 01, 02, 03, 04, 05, 06 and 21 have

limited log suites and in particular do not have acoustic impedance data

available. Well 16 is only logged over a deeper section and does not have

any data over the modelled interval.

Frequency

Sands are generally indicated by high impedances and have typical velocities

of 12,000 ft s-1, a 30 ft sand thus being around 5 ms thick in time. In the

example model a simple two

16%

facies definition has been

14%

adopted, comprising

12%

sand and shale. Sand facies

10%

have been identified with a

Shale

combination of ILM

8%

Sand

(resistivity), Vclay and

6%

acoustic impedance cutoffs.

4%

An impedance cutoff of 8,000

2%

ft s-1 * g cm-3 gives a good

0%

seismic discrimination

5500

6500

7500

8500

9500

10500

11500

between sands and shales

Impedance (m.s * g/cc)

(see histogram).

Histogram of acoustic Impedance by facies type

All rights reserved

STRATTON FIELD

StrattonCleanModel.pet

All of the data has been loaded to Petrel and the model framework

constructed. Two seismic horizons have been picked form coloured

impedance data and used to provide the structural framework. These are a

middle Frio pick (MFRIO) and a pick which corresponds to the top of the

Upper C38 formation.

A simple velocity model has been used for

depth conversion. Seismic volumes

including coloured impedance,

deterministic inversion and stochastic

inversion data, including probability cubes

have been loaded, depth converted, and

re-sampled in to the model.

Two zones have been defined for the

model and these have been sub-divided

into five layers.

Zone 1

o MFrio

o B46

Zone 2

o C38 Upper

o C38 Lower

o D11

All rights reserved

STRATTON FIELD

East-west section showing Coloured Impedance seismic data and Vclay/ILM logs

All rights reserved

STRATTON FIELD

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

All rights reserved

SCALE-UP LOGS

SCALE UP OF WELL LOGS

Prior to use the loaded well data must be scaled-up to the vertical resolution in

the 3D grid. Scale-up is usually based on some form of averaging. Scale-up

will be applied to continuous measurements such as porosity but also to

discontinuous or discrete logs. Scale-up of discrete measurements such as

facies is usually different to the averaging applied to continuous

measurements and is often based on selecting the most frequently occurring

value. After this type of scale-up it is particularly important to check that the

results are still consistent with the original well logs, particularly with regard to

the scaled-up logs correctly reproducing the original facies proportions and

thickness distribution.

There are many options in reservoir modelling packages allowing different

approaches to scale-up. This includes biasing the method of property

averaging to a discrete log, usually the facies log. Petrel offers the following

algorithms:

Scale-up methods for discontinuous logs:

Median

Minimum

Maximum

Arithmetic

Weighted

Arithmetic mean

Harmonic mean

Geometric mean

RMS

Minimum

Maximum

Scale-up in Petrel can be found under Property Modelling. For facies logs,

the scale-up is usually made using the Most Of algorithm. For porosity log

scale-up the arithmetic mean is usually used, with the additional criteria of

biasing the scale-up using the facies log.

In addition to selecting the averaging method, a decision must also be made

on the method by which to populate the model cells penetrated by the well

path. In common with other reservoir modelling packages Petrel offers a

several methods. Although it depends on the method used, scale-up

generally causes a change in the PDF of the data. Thought must also be

given to the problem of well geometry. The averaging will not be regular

All rights reserved

SCALE-UP LOGS

within the cells if there is significant well deviation. Typically, between 10

and 300 log samples are averaged into a cells depending on the deviated

trajectory of the well passing through the cell.

volume between a well log sample and the modelled cell is of the order of

500,000 : 1. No matter what method of scale-up is chosen, the cell will almost

certainly not contain the appropriate average for the scale of the cell. This

has implications both for the estimated PDF (histogram) and the predicted

range of uncertainty and is a limitation of current reservoir modelling software.

Resampling by selecting a subset of data does not change the scale of

measurement and so the CDF/PDF remain the same as the original measure.

Populating model cells by this approach will result in values measured at log

scale being used to populate cells defined at a much larger scale.

All rights reserved

SCALE-UP LOGS

Averaging is a smoothing operation and changes the effective scale of

measurement. The PDF becomes narrower and the CDF narrower and

steeper. For a reservoir model, simply averaging the logs into the cells does

not result in the correct average for the cell size. Depending on cell size

variation and deviated well path geometries, typically between 10 and 300 log

samples will be sampled into a cell. This is not the correct scale-up and the

result may also give rise to a variation of the PDF from cell to cell. For a data

set with cells oriented perpendicular to the well path (eg low relief and vertical

wells) the arithmetic scale-up is the best choice with current software.

The scale-up options available in Petrel to control which cells are populated in

the model include:

Simple all cells that the well

trajectory goes through will be

populated with values according

to the scale-up method chosen

populated if the well trajectory

passes through two opposite

walls of the cell e.g. top and base.

based on all cells immediately

adjacent to the traversed cell

which are in the same layer

All rights reserved

SCALE-UP LOGS

After scale-up it is important to undertake quality control to ensure that the

scale-up logs are still representative of the original well logs. Depending on

the log type being scaeld-up, these checks will include looking at:

Facies Proportions

Property Histograms

Thickness Histograms

As a minimum the original and scaled-up logs should be overlaid and their

histograms compared. In addition basic statistics should be checked, looking

to ensure reproduction of the mean and to verify that facies proportions are

preserved.

vertical proportion curves (vertical histogram) within each layer for the original

and upscaled logs. This will be illustrated in the Data analysis section.

All rights reserved

SCALE-UP LOGS

STRATTON FIELD SCALE-UP

Facies Logs

Scale up of well logs is found on the

Processes tab under Property

Modelling.

At the top of the menu choose create

new property. In the centre select

input from well logs and select the

logs as Facies.

The averaging method to use is

Most Of, treat logs as lines. As all

the wells are vertical, the Simple

method for populating all cells

penetrated by a well will be a good

choice.

After scale-up, it is important to

check the model facies log against

the original facies log (see next

page). In particular, it is important to

check that the scale-up manages to

capture sufficient detail in the

layering and that there are no major busts between the original and upscaled

logs. The most likely problems will be encountered if the cell size of the

model is too coarse or if the facies changes are strongly cyclical.

Next a check should be made of the facies proportions. This can be done on

the Models tab. Double click the Facies (U) object, select the Disc. Stat. tab

on the pop-up dialogue and check that the upscaled cells for the zone contain

proportions close to those in the original logs. A visual comparison can be

made using the histogram tab.

All rights reserved

SCALE-UP LOGS

All rights reserved

SCALE-UP LOGS

A final quality check is to inspect the thickness histograms for each reservoir

layer. This can be done using the Data analysis option in Property Modelling.

Compare the thickness histograms between the original and upscaled logs.

Note that the after upscaling, the minimum thickness is determined by the

average cell thickness. Typically the thickness distribution for sands will show

an exponential decline such as for the B46 interval for Stratton Field.

MFRIO Zone

B46 Zone

All rights reserved

SCALE-UP LOGS

D11 Zone

Porosity Logs

Scale up of the porosity logs is

relatively straightforward. Again,

create a new property and select

input from the original well logs.

The important element is to use the

Use Bias option. This ensures that

samples averaged only come from

cells with same facies type as

upscaled cell.

An arithmetic average is an

appropriate choice and, as for facies

log upscaling, the Simple method is

fine for vertical wells.

All rights reserved

SCALE-UP LOGS

Recall that averaging changes the shape of the distribution (the PDF and the

CDF). By comparing the original logs with the upscaled cells we can see how

averaging moves the samples towards the mean, causing the distribution to

become more peaked with fewer extreme values.

All rights reserved

DATA ANALYSIS

DATA ANALYSIS

After creating the structural, zone and layer framework and preparing the well

logs by the application of scale-up the next stage in a reservoir modelling

study is to undertake exploratory data analysis.

Petrel distinguishes between data analysis for continuous properties (such as

porosity) and discontinuous properties (such as facies logs). For continuous

properties the tools take the form of Transformations. The data analysis

transformations available for continuous properties are:

Input Truncation

Output Truncation

Logarithmic Transformation

Cox-Box Transform

1D Trend

2D Trend

3D Trend

Scale Shift

Normal Score Transformation

Thickness

Probability

are the variogram analysis. Variogram analysis can be tackled in various

ways in Petrel, not just in the Data Analysis module. Variogram analysis will

be considered in a separate section.

Discrete Properties

Proportion tab allows analysis of the facies

proportions vertically through the model, or within

each zone or layer of the model. The proportion

curve is a vertical histogram of facies type and is

characteristic of the depositional system within a

sequence.

The usual approach would to be to fit a function to

the vertical proportion curve in order to ensure

that the facies have the correct vertical

distribution within each layer. In this example

(right) example, most of the sand is concentrated

in the base of an interval and we would expect

our model to replicate this.

All rights reserved

DATA ANALYSIS

The vertical proportion curve can be analysed for the original logs or upscaled

logs and is a further useful quality control on the results of scale-up of the

facies logs.

and upscaled logs (right)

distribution throughout interval) or more likely either a vertical probability or a

depth trend will be defined. The vertical probability modelling would usually

be applied separately for each layer. In this example (below) the trend has

been modelled as a distribution from the original logs (it could also be

modelled from the upscaled logs) and then smoothed.

Modelling vertical proportion curves: original (left); fitted (centre); smoothed (right)

All rights reserved

DATA ANALYSIS

B46 Zone

All rights reserved

DATA ANALYSIS

D11 Zone

Thickness tab is used to visualize the thickness distribution of the facies. Its

use has been described in the scale-up section as a quality control tool. An

example is given below comparing the original and upscaled facies thickness.

As noted above, remember that the minimum thickness in the upscaled log is

determined by average cell thickness. Also note the characteristic

exponential shape to the sand thickness distribution in the original logs below

a strong argument for not using symmetric distributions for sand thicknesses

in object modelling.

B46 Zone

Probability tab can be used to

constrain a facies model to a

continuous property such as a

seismic attribute. It can be used to

convert the seismic attribute cube a

facies probability cube ready for use

in constraining Facies Modelling using

Sequential Indicator Simulation (SIS).

The usual functions for fitting either

trends or probability distributions are

available in this panel.

All rights reserved

DATA ANALYSIS

Continuous Properties

The most commonly modelled continuous properties would be porosity or

permeability.

Input Truncation limits the minimum and maximum range of values used in

the analysis and modelling. Any values outside the range set will be set as

null.

Output Truncation is usually used as the final step after a Normal Score or

similar back transformation to ensure that the final modelled property values

stay within required (or physical) limits.

Logarithmic Transform is a straightforward logarithm of the data values,

particularly useful for permeability or other data with lognormal distributions.

Cox-Box Transform (sometimes referred in statistical literature as Box-Cox)

is a method of removing skewness from a set of data. The transform is:

the Cox-Box transformation parameter, expressing the degree of skewness.

The software will iteratively solve for the best value of lambda down to an

accuracy of 0.1.

A logarithmic transformation of a variable will remove skewness for a variable

following log-normal PDF. This is a special case of Cox-Box where the

coefficient lambda = 0. Be aware that the uncertainty is not transformed

linearly under a logarithmic transform and a Normal score transform may be

preferred.

After a Cox-Box or logarithmic transform to remove skewness the mean of the

variable may not be zero and the standard deviation not equal to one. This

can be corrected with the scale-shift transform.

Scale-Shift Transform will correct the data to have a mean of zero and a

standard deviation of one. This is achieved by subtracting the mean from all

data values (to shift it to a mean of zero) and then scaling by divinding the

shifted samples by the standard deviation, which standardises the variability

to a standard deviation of one. Note that scale-shift does not change the

shape of the histogram.

Normal Score Transform is a complete approach which forces the variable

CDF to a standard normal distribution N[0,1]. Recall that a standard normal

distribution is essential as the input to Sequential Gaussian Simulation (SGS),

All rights reserved

DATA ANALYSIS

the algorithm which should usually be used for generating porosity and

permeability properties within the model.

The normal score transform should be considered as the standard approach

to be used in preference to other methods in most cases. The normal score

transform may be difficult to apply with very small data sets. In this case

consider a simpler method such as Cox-Box followed by scale-shift, or

manually edit the normal score transform. Care should always be taken with

modelling of the tails of the distribution and truncation of the output.

Trend removal in Petrel can be in 1D, 2D or 3D. For a 1D trend removal the

direction X, Y or Z should be specified, although it is most common to remove

vertical trends. The distances used in trend removal can be based on the real

coordinates or the simbox coordinates.

Examples of trends might be:

Horizontal map trends

Spatial volume trends

Trends in 3D such as those derived from seismic inversion cubes should best

be used as a secondary property in petrophysical modelling under a

collocated co-kriging or co-simulation framework.

It is very important to note that deterministic inversion data should NOT be

used to constrain a reservoir model. Deterministic inversion data MAY be

used if the low frequency component is filtered out to remove the influence of

the wells from the deterministic inversion. A seismic attribute or relative

impedance (such as coloured inversion) CAN be used. The ideal

impedance product to constrain a reservoir model would impedance

realisations generated from a stochastic inversion. It is important to note

that some stochastic inversion methods (such as Promise) produce output

similar to deterministic inversion schemes and the same caution is advised

before using these as conditioning for the model.

Input truncations to remove outliers in input data

1D trend removal to account for porosity-depth trend

Normal score transform for input to SGS

All rights reserved

DATA ANALYSIS

Stratton Field Porosity Transforms

In the Stratton

data the

supplied total

porosity log

has been

defined form

the bulk

density log. It

contains

erroneous

high porosity

estimates for

any non-sand

interval. Therefore a first step has been to use input truncation on the

upscaled logs to set the total porosity values to null in the shale facies. For a

reservoir model, the shale facies porosity values would probably be set to a

low or zero value across the model after generating the facies.

Setting the porosity to null in this case allows the complete porosity maps,

only generated using shale values, to be viewed for the entire model. This is

particularly useful for training purposes.

In the Stratton data set, the best approach appears to be work over the whole

of the upscaled and not zone by zone. Working with all the upscaled cells

gives a clear porosity-depth trend which can be modelled and gives a lot of

data for the subsequent normal score transform. Click the Zones button off

and select the sand facies.

1D Trend has been used

to model the porosity-depth

trend across all the zones.

It is usually worth noting

down the depth trend

coefficients in case you

need to use this function in

a later stage of modelling,

or report it.

Remember to select the Z

button to model a vertical

trend.

All rights reserved

DATA ANALYSIS

Normal Score Transform

has been used to convert

the residual porosity values

(after removal of the depth

trend) to a Gaussian or

normal distribution. By

using all zones a large

number of upscaled values

are available and this

makes the normal score

transform more robust.

Use the fit function button.

In some cases you may

wish to edit the resulting

function manually

After performing the

transforms use the copy

button to copy the parameters, select Use Zones and then paste the final

transforms to all zones.

A relation

between porosity

and impedance

has been

investigated using

a function window.

Crossplotting

coloured

impedance with

porosity suggests

that there is a

weak correlation

between the two

and this might

lead to the use of

the seismic

impedance data in

constraining the

porosity

simulation.

However, care should taken to properly understand the impedance

information. The crossplot only considers the upscaled cells (sensibly, since

these will be averaged to be closer to the seismic resolution). Unfortunately,

this means that we are selectively comparing only sands to the coloured

impedance data.

All rights reserved

DATA ANALYSIS

An inspection of the original log data shows that for clean sands there is clear

functional relationship between impedance and porosity (Wyllies relation

this is Gulf of Mexico data), the presence of shales means that it would be

incorrect to apply this relation to model porosity within our model. This is

because our simulated facies will not be constrained and could then be

populated with seismic driven data which is inappropriate at a given location.

The crossplot of sand and shale shows that we must first constrain the facies

distribution in the model using the seismic, the histogram shows there is

discrimination of sand and shale using the seismic data.

13000

12000

11000

10000

Non-clean

9000

Clean sand

8000

7000

6000

5000

0%

5%

10%

15%

20%

25%

30%

35%

Porosity (%)

16%

14%

12%

Frequency

10%

Shale

8%

Sand

6%

4%

2%

0%

5500

6500

7500

8500

9500

10500

All rights reserved

11500

EARTH

WORKS

FACIES MODELLING

So far in this course all of the geostatistical examples we have considered involve

spatial estimation of continuous measurements. The variable can take on any value,

subject to some limitations concerning physical bounds.

Some types of data are not continuous measurements. For example, male and female.

We can convert any continuous variable into a categorical variable by classification.

Many sciences, including geology, are pre-occupied with classification of observations

into classes. An example in geology is classification by lithology such as sand or

shale. The problem with these classification systems is they assume black/white and

there is ambiguity in how to handle shades of grey in between. The problem is usually

solved by defining a cutoff such as < 25% shale volume rocks will be classified as

sand, greater than this shale.

Most reservoir models are constructed by initially defining a series of facies (or

sometimes flow units). This is for two reasons

Different units may have different shape or form in the earth such as channels

or mouth bars and it is important to capture this form in the model

Different units may have different porosity-permeability characterstics and

so the facies distribution is used to provide a template into which poro-prm data

are mapped or simulated.

The facies act as templates for the subsequent simulation of petrophysical properties,

notably porosity and permeability.

We will first consider the simplest problem of mapping classes or indicators by

considering Indicator kriging (and its simulation counterpoart). We will also consider

some of the limitations of the indicator methods and, in the following section, consider

the alternative of object modelling.

The original purpose of indicator Kriging was to provide a least-squares estimate of the

conditional cumulative distribution function at different cutoffs for a continuous variable.

However, indicator Kriging has been most used for estimating the indicator transform of

a variable or, even more familiarly, for estimating facies, and this is the application we

are going to consider here. Be aware that the implementation in Isatis is based on the

original purpose and we have to do some quick calculations in File -> Calculator after

running the indicator kriging in order to obtain the individual facies probabilities. The

Indicator simulation implementation in Isatis is set up for facies simulation in the usual

manner of most reservoir modelling packages.

If the variable is a binary categorical variable, eg set to 1 for sand and 0 if not a sand

then the direct Kriging of the ones and zeros provides a model for the probability that

the rock type represented by 1 prevails at the target location.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

The following figure shows a basemap posting of sample locations and the rock type

present at each location.

0 = Shale

1 = Channel Sand

2 = Crevasse Splay

Indicator kriging is based on the principle of kriging of a binary indicator. If

we use 1 to indicate the presence of a particular rock type and 0 to indicate

the absence of that rock type then kriging of the 1 / 0 data set will give us the

probability (expected value) of the rock type represented by 1.

By repeating the process with several rock types, redefining the 1 / 0 definition

for each, we would obtain a set or probability maps, one for each rock type.

The sum of these maps will be 1.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

Indicator kriging solves the multiple indicator (multiple rock type) problem by kriging the

cumulative distribution function cut by the indicators rather than the individual 1 / 0 data

sets by category. The result is the same from the cumulative we can calculate the

probabilities for the individual facies.

The question of what variogram model to use does arise. For facies based indicator

kriging, we can define a variogram on each category, but for true indicator kriging

where we are estimating the CDF then the variograms are defined on the cutoffs. A

simple alternative is to define a single variogram, either on the original data (ie the

facies classifications) or the cutoff which splits the data into two halves (the median

cutoff. The implementation will be generally be software dependent.

approach. The procedure for sequential indicator simulation (SIS) is as follows:

Krige each of the facies probabilities at the target grid node

Check sum of probabilities is 1, otherwise normalise

Arrange facies probabilities cumulatively from 0 1

Draw a random number on the interval [0,1] and determine facies at target grid

node

Insert facies code into grid

Repeat with new target grid node treating the previously simulated facies point

as additional data points

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

From looking at the indicator simulation realisations the principle drawback of the

method is apparent the shapes of the facies do not look like geological structures.

The problem is the variogram is a 2-point statistic and therefore cannot characterise

the geometry of the facies. Recall also that it is the maximum entropy solution.

In the illustration above, the channel clearly has a complex geometry that cannot be

reproduced in the indicator simulation. Also, the facies proportions may be incorrect

we bias our sampling with wells in order to target the better quality reservoir rock

such as might be found in the channel. In fact, the proportions of shale, sand and

splay lithologies in the above diagram are 0.63, 0.24 and 0.13, different to the 0.45,

0.35 and 0.20 sampled in the wells. Also, be aware that facies proportions are also a

function of the size of the model area.

An improvement to indicator simulation that would allow more complex geometric

shapes to be generated would be to use higher order statistics such as 3-point or multipoint. Direct solutions are not currently available. Existing methods, such as SNESIM,

use a template geometry and solve the problem by a sort of pattern matching.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

of size and shape parameters. This is best achieved by the now widely used objectbased simulations.

Indicator simulation is an example of a pixel-based facies simulation method. For

many years there have been several competing methods for simulating facies which

are pixel-based, in contrast to the object methods.

Each of the pixel-based methods have their strengths and weaknesses (as does the

object method approach).

As we have already met indicator simulation, it is probably a good place to start in

comparing the available methods. Its advantages are:

Sound theory

Handles any facies organisation

Widely available and fast

Entropy too high?

equivalent approach which was invented in Fontainebleau. Its called truncated

Gaussian.

Hierarchical facies succession (may be disadvantage too)

Proportion curves an excellent geologic descriptor

Disadvantages include:

Entropy too high?

Not widely available

It was recently extended with the so-called plurigaussian method which has addressed

the facies hierarchy problem but it still remains a method with limited availability. The

principal packages implementing it are Heresim (from IFP) and Isatis.

The truncated Gaussian and sequential indicator methods are somewhat similar and

both suffer from the same fundamental weakness they are unable to generate the

complex geometric forms we associate with geology owing to their 2-point statistic

basis. They are essentially pixel-based simulation methods.

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

Object Simulation

The alternative to pixel-based methods is too attempt to model facies using objects.

The idea of object simulation was demonstrated as early as the mid 1970s (Bridges

and Leeder, 1976) with a simple 2D cross-section model of shales.

The idea of defining geological facies such as channels as objects came to the fore

with work from Norway in the 1990s. These models are now widely used for reservoir

modelling, relegating the pixel-based methods as second line.

Object models are excellent at reproducing the complex geometries we associate with

geological objects, but this is also their weakness. Having a means to define them still

leaves the problem of determining the parameters. Rarely do we have sufficient data

from wells to reliably parameterise the objects.

The process of object simulation involves trial and error. A neutral background facies

is first set in all cells of the model (this is usually the shale, but sometimes it is another

type). Objects are generated according to probability distributions defining their

characteristics such as width and thickness of channels. As each object is generated,

the software attempts to insert it into the model. Its insertion must not result in passing

through a well where a different facies is coded and ultimately all facies coded in the

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

well as channel should end up with a simulated channel passing through them. We

keep inserting objects until all well with that facies have been included in an object. In

addition, we must ensure the volume proportion is correct. We cannot insert channels

if we end up with too high proportion of sand in the model. Alternatively if we insert

channels and satisfy our well conditioning data, we may have insufficient channel

proportion and have to simulate more non-conditional objects. This is controlled by a

stop criteria, usually the net:gross ratio.

In object modelling the wells are typically honoured first and then the interwell region

is simulated second. Care must be taken by the algorithm in simulating the interwell

region to avoid conflicts with the known sequence of lithologies in the wells.

conflict at wells

stopped when N:G achieved

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

Object methods give excellent reproduction of entropy but they are relatively slow due

to the accept/reject mechanism. The main problem is parameterisation of the objects.

Some of the parameters are simply guessed at, as the following typical example of

object parameters from a reservoir modelling report shows:

Body Type:

Low sinuosity Channel

Shape

Half-cylinder

Thickness

Triangular Distribution

Width

Expression

Orientation

Triangular Distribution

Amplitude

Expression

Wavelength

Triangular Distribution

Stop Criterion Volume (Net:Gross)

Branch Points

Uniform Distribution

Branch Location

Uniform Distibution

2 4 10 m

Thickness * 25

10 50 90 degrees

Width

1000 3000 5000 m

26 %

10 20/1000 m

0 default

The limitations of object modelling are primarily the difficulty of parameterising the

models. To some extent it can be argued that indicator models are more data driven,

although the paucity of wells to produce reliable variograms negates this argument

somewhat. Choice of models and parameters can be assisted by knowledge of the

geological environment and could be supplemented by analogue databases comprised

of outcrop or historical field data. Another source of data could be analysis of the

seismic expression of objects such as channel widths, sinuosity which might come from

analogous shallow intervals where seismic resolution is good.

Objects in Petrel

Object modelling in Petrel requires specification of a series of parameters to be

specified. For a fluvial channel the channel direction (orientation of the channel

axis) must be specified along with the wavelength, amplitude, width and thickness

distributions or relations. Levees can also be defined and these will require additional

width and thickness parameters as will the occurrence of crevasse splays.

In addition Petrel has a number of trend options available for use with fluvial channels.

These include:

Vertical probability curves

Flow lines

Source points

A pair of flow lines can be used to give orientation and an evelope within which

channel objects will be inserted to give bundles of objects. Source points are used

as the start points for channels and can be combined with flow lines (but not used in

conjunction with the probability map).

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

Most reservoir modelling packages, like Petrel, can also offer a wide variety of others

facies bodies geometries including:

Box

Pipe and upper/lower half pipes

Ellipse (full, half and quarter)

Deltaic/Alluvial fan

Aeolian sand dune

Oxbow lakes

The facies bodies are generally defined by orientation, minor width, major/minor axis

ratio and thickness parameters.

Erosion rules allow the next simulated facies types to erode previously simulated

facies. In its simplest form, all subsequent facies erode the background facies, of

course. Typical erosion rules include:

Replace ALL other facies

Replace ONLY the specified facies in the previous bodies

Replace ONLY the following facies in the background property

Other option, which are generally slow to run, including only replacing previously

simulated bodies if they are completely within a new object and replacing a facies with

itself (if this is not allowed only isolated objects of that facies will be present).

The following pages show some examples of indicator kriging, indicator simualation

and object simulation of facies within a zone

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

Layer 1

Layer 6

Layer 2

Layer 7

Layer 3

Layer 8

Layer 4

Layer 9

Layer 5

Layer 10

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

Layer 1

Layer 6

Layer 2

Layer 7

Layer 3

Layer 8

Layer 4

Layer 9

Layer 5

Layer 10

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

FACIES MODELLING

EARTH

WORKS

Layer 1

Layer 6

Layer 2

Layer 7

Layer 3

Layer 8

Layer 4

Layer 9

Layer 5

Layer 10

Introduction to Geostatistics

2002 Earthworks Environment & Resources Ltd.

All rights reserved.

VARIOGRAMS IN PETREL

VARIOGRAMS IN PETREL

In Petrel there is more than one way to construct and model experimental

variograms. None of the schemes provided is ideal and variogram work in

Petrel should be regarded as generally sub-standard compared to other

industry packages.

The two routes for generating variograms in Petrel are:

Data Analysis

The original method of estimating and modelling

variograms in Petrel is through the Property

Settings. This is more reliable for the user to

successfully be able to generate and correctly

model variograms but it is still tedious, with a lot

of repetitious work required by the user.

Variogram modelling using this method is

accessed by double clicking on a property in the

model to open the Property Settings and

scrolling to the last tab.

The first point to note is the check-box option

labelled Isotropic (circled orange, below right).

This is very misleading. The presence or

absence of anisotropy is a property of the

geological data we are trying to detect or

establish through variogram analysis. The

button should properly be labelled

Omnidirectional, to denote ignoring directional

information in the analysis.

To perform indicator simulation we need to

generate experimental variograms by facies

and zone and model them individually. This

can be done using the property modelling

route by double clicking on properties and, on

the last tab, set the appropriate filter for facies

and zone (see figure below). On the

variogram tab (shown right) turn on the Use

property filter option.

The resulting variograms will be stored in the

variograms folder in the Input tab. If making

several variograms remember to deselect the

Overwrite last option at the bottom of the panel.

All rights reserved

VARIOGRAMS IN PETREL

appropriate way in which to calculate

variograms from seismic properties. With

the seismic attribute resampled into the

Petrel model use the index filter to allow

horizontal variograms to be generated by

each required K layer within the model.

Again, remember to turn on the Use

property filter in the Settings dialogue.

For many zones or K layers the Property

Settings route to generate experimental

variograms is slow, requiring the filter to

be amended each time a different zone or

K layer is used for computation.

first display the experimental variogram in a

Function Window. Then use the Define

Variogram button on the Petrel right hand

tool bar. This opens a menu for defining a

variogram model (see right). Define the

appropriate settings and press Ok. The

variogram model is superimposed on the

variogram window and can be edited

interactively. An example is shown below.

All rights reserved

VARIOGRAMS IN PETREL

The most commonly used

approach is the Data

Analysis route. This is

the newest method

implemented in Petrel.

The interface in data

analysis has some clever

graphical gimmicks, such

as implementing the GSLIB variogram diagram as

part of the user interface

to select parameters, but

it basically fails to provide

the essential functionality

for what is the most basic

geostatistical step. In

particular it has great

potential for errors or bad parameter choices, both through user ignorance but

also due to fundamental design flaws in the interface.

All rights reserved

VARIOGRAMS IN PETREL

Some of the key problems with the variogram facilities provided in Data

Analysis are:

Sill rescales for every variogram computation and direction

User has to ensure same parameters in different directions when

recalculating

Cannot compute multiple directions and display together

Cannot compare directions for models, especially horizontal

No line joining variogram points

No ability to mask samples to exclude from variogram computations

Must APPLY to keep settings, even when switching between zones!

The histogram displays the number of pairs used to calculate the experimental

variogram value at each lag useful for determining reliability of the

variogram. Generally, vertical variograms will be reliable whereas horizontal

variograms will be difficult to interpret and model. Seismic derived attributes

may give limited additional information.

To generate an omnidirectional variogram in data analysis set the tolerance

angle to 90.

All rights reserved

FACIES VARIOGRAMS

Stratton Field Facies Variograms

Assuming we want to attempt a Sequential Indicator Simulation for facies

generation within our model, we will require to model a variogram for each

zone and facies modelled.

For a two facies case (as we have here in the Stratton Field data set) we only

need to generate variograms and models for one of the facies. For a two

facies case, the facies codes are a binary complement and so the variograms

will be identical for the two facies. For three or more facies the variograms

must be generated individually and modelled separately for each facies.

Vertical Variograms

With well data it is usually

best to begin with the

vertical variogram. The

vertical variogram is

usually straightforward to

generate, with plenty of

samples available to

generate sufficient pairs

for a reliable variogram.

Check the vertical

variograms in real mode

rather than simbox mode.

Note that the Data

Analysis variogram

functionality always

normalises the total

variance to 1. The true

variance without normalisation would be related to the facies proportions.

Above is an experimental facies

(indicator) variogram for the MFRIO

zone. The variogram model in the

above diagram has the default

parameters, with the sill normalised

to 1. Note how the experimental

variogram rises and dips on the

shorter lags whilst staying below 1

and then rises to a variance above

1 in the later lags before again.

This is possibly indicative of nonstationary behaviour. Examination

of the vertical proportion curve for

this interval supports this

interpretation (see right). The

sands are clearly clustered at the base and top of the interval. Short lags tend

All rights reserved

FACIES VARIOGRAMS

to be in sands or shales. With long lags, pairs tend to be in sand (one sample

of pair in the top sand, one in the base sand) causing the reduction of

variance at long lags.

One possibility in this data set is to

consider the vertical variogram for

all zones simultaneously. Of

course, in deciding to generate and

model just one variogram vertically

we gain an advantage in that a

stationary assumption is quite

reasonable but this must be

contrasted with the possibility that

the vertical variogram range could

change between the zones, for

example due to compaction effects.

Using the data from all zones does

result in a well behaved variogram.

Using the data from all

zones together, a vertical

variogram has been

modelled. Choose the

model type. In this case

an exponential model

provides a god fit to the

data, a common result for

well log based

variograms. The sill is set

to 1. Remove the nugget

we do not interpret the

facies as having a

random noise component.

When satisfied with the

the model, copy and

paste to all facies before

turning on the zones and

pasting to all zones. Remember that with more than 2 facies, each facies

must be modelled separately.

Horizontal Variograms

After establishing the vertical variograms attention should turn to modelling

the horizontal variograms. Experimental horizontal variograms are likely to be

less reliable than vertical variograms unless sufficient wells are available.

Also, the essential short range information is usually missing unless wells are

very close together or have substantial horizontal well section.

All rights reserved

FACIES VARIOGRAMS

The Stratton Field data

contains only vertical

wells although the

interwell spacing is

relatively short, being as

close as several hundreds

of feet. When setting the

experimental variogram

calculation parameters

consider the maximum

pair separation distance.

For Stratton, the

maximum pairs are

separated about

L=16,000 ft, so we would

expect an experimental

variogram to have reliable

lags to about 8,000 ft

(L/2).

Begin with an omnidirectional variogram. Set the tolerance angle to 90 and a

large bandwidth. It is usually better to work in simbox mode for the horizontal

variograms, the variogram calculation then following the layering. The

thickness is the vertical offset in distance units for a pair to be accepted. To

obtain the maximum information from the variogram, this should be set to the

sample size (either log sample rate or average vertical cell thickness). Set the

number of lags to give a sensible lag spacing, not too many lags or the

statistics in each will poor. To few lags and the experimental variogram will be

averaged to much. Use the same judgement as you would to decide the bin

widths for a histogram plot.

After considering the omnidirectional variogram and establishing the basic

parameters for calculating the experimental variogram, next analyse

directional variograms. As already discussed, at least 4 directions should be

considered. Set a tolerance angle of 5 20 degrees, depending on the data

availability. Remember that decreasing the bandwidth will also restrict the

number of pairs available for inclusion in the experimental variogram.

Although it is awkward to do so in Petrel, you need to compare the horizontal

directions in order to establish any evidence for anisotropy in the horizontal

plane. The only way to do this may be set all parameters and then capture

the image for each direction and compare in PowerPoint or some similar

manner. When switching from major to minor direction it is very important in

Petrel to ensure that you have consistent calculation parameters to ensure

valid comparisons are made.

If you set the variogram calculation parameters for one zone, it is

recommended to copy/paste these calculation parameters to all facies/zones

for convenience in calculating variograms for subsequent zones.

All rights reserved

FACIES VARIOGRAMS

For Stratton Field it is very difficult to see any evidence for anisotropy in the

facies variograms. The horizontal plane has been modelled as isotropic, the

model being fitted to the omnidirectional variogram with a range of 1400 ft.

The vertical range was already modelled as 22 ft, giving a horizontal:vertical

anisotropy of around 60:1.

All rights reserved

FACIES VARIOGRAMS

All rights reserved

FACIES VARIOGRAMS

Omnidirectional horizontal indicator variogram and fitted model for C38 Upper

All rights reserved

FACIES VARIOGRAMS

Omnidirectional horizontal indicator variogram and fitted model for C38 Lower

All rights reserved

FACIES VARIOGRAMS

All rights reserved

SEISMIC VARIOGRAMS

SEISMIC RESAMPLING

Seismic SEGY data is usually loaded into Petrel in the time domain. A

velocity model is then required to depth convert the seismic SEGY data. A

number of depth conversion strategies are available within Petrel. In

preparing this course, an attempt was made to use the V0 + kZ method. This

has a serious error in its implementation within Petrel in that it does not

correctly construct the velocity model and depth convert in the presence of

negative elevation units. For Stratton Field a simple layer cake model has

been used for velocity modelling and depth conversion. Two flat time slices,

one at 1320 ms and the other at 1750 ms have been used to define a model.

Average and interval velocities have been used from the well positions and

simply mapped to provide a smooth model for depth conversion. As Stratton

Field (in the shallower intervals we are considering) has almost no relief, the

depth conversion is very accurate. This is an important consideration as

misalignment of the seismic data with the model will seriously impair any

attempts to find relationships with seismic attributes or constrain the model

with seismic data.

such as during an inversion study, in ensuring a

consistent well tie framework between seismic and well

logs and that this framework is correctly imported into

the model.

After constructing the velocity model and depth

converting the seismic SEGY data the seismic can be

resampled into the model cells. Seismic resampling

can be found in Property Modelling -> Geometrical

Modelling.

All rights reserved

SEISMIC VARIOGRAMS

In the Geometrical Modelling dialogue

select Seismic Resampling. Select the

appropriate seismic volume and its

accompanying property template to ensure

that the colour table is carried through to the

property resampled into the model.

There are a number of options available to

control how the cells are populated.

Closest takes the nearest or most central

seismic sample and assigns it to the cell. If

there is a significant difference in resolution

between the seismic and model grid this is

not the best choice

Interpolate using a weighted interpolation of the four seismic cells closest to

the cell centre. Cannot be used with discrete data types such as a facies

volume.

Intersecting uses all the seismic samples within the cell boundary. No

correction is made for the intersection volume. If bringing in a discrete data

type such as a facies variable then make sure to use the most of or median

averaging options.

Exact is the same as Intersecting but with the addition of a volume correction.

This is the slowest method.

The averaging method chosen should be appropriate for the data type. For

continuous properties this may be arithmetic or other mean calculation. For

discrete data types such as facies then most of (mode) or median should be

used.

There is almost certainly a difference in resolution between the seismic data

and the model. Vertical seismic resolution is likely to be lower than the

resolution of the cells within the model. Horizontal resolution may be similar,

although could be larger or smaller. We call this the support of the

measurement in geostatistics.

The seismic time sample rate and trace spacing do not define the resolution

or support of the seismic measurements. It is the seismic amplitude spectrum

(frequency content and band width) which defines the support. Resampling

data does not change the resolution. Upscaling (averaging) does change the

resolution to a large support. The popular notion of downscaling is nave:

once a process is measured at a larger support a change to a smaller scale of

support can only be made in probabilistic (stochastic) sense.

All rights reserved

SEISMIC VARIOGRAMS

SEISMIC VARIOGRAMS

Given the difficulty of generating horizontal variograms from well data we

often need to consider alternative sources of information on variogram

parameters. One possible source is the use of analogue and outcrop data

bases. Another source, almost always to hand, is to analyse the seismic data.

The most appropriate date for this purpose is a relative impedance volume

such as that produced by coloured inversion. Deterministic or stochastic

inversion data should not be used for estimating variograms.

Having resampled the seismic attribute

such as coloured impedance into the

Petrel model, horizontal directional

variograms can be calculated by K layer

within model and used to confirm or

supplement the variogram analysis made

using the well data. To calculate

experimental variograms by K layer,

double click on properties and turn on

Index Filter and select K index filter and

define the appropriate layer.

click the appropriate property in the model (in

this case, seismic coloured impedance) and

select the Variogram tab (last tab to right).

Make sure the property filter is turned on to

filter only the values for the required K layer.

Choose Sample Variogram and Horizontally

and then set the direction and angular

tolerance. Recommended starting point

would be 4 regular directions with a narrow

angular tolerance of 5 degrees or less.

Ensure the Overwrite last option is

unchecked. Each direction and K Layer

experimental variogram will have to be

calculated manually. If the generation is

systematic and they are grouped properly in

the variogram folder then they will be coloured consistently for subsequent

display and analysis

All rights reserved

SEISMIC VARIOGRAMS

Coloured Impedance Horizontal Variograms

K layer = 1

K Layer = 20

K layer = 40

K Layer = 47

K layer = 60

K Layer = 80

All rights reserved

SEISMIC VARIOGRAMS

Coloured Impedance Horizontal Variograms

K layer = 100

K Layer = 120

K layer = 140

K Layer = 160

K layer = 180

K Layer = 191

All rights reserved

SEISMIC VARIOGRAMS

Inspection of the horizontal variograms from the seismic coloured

impedance suggests little evidence for anisotropy, with the short lag

element of the directional variograms of all layers appearing isotropic.

At longer lags the total variance exceeds 1 and this suggests that there

they may be some small non-stationary component present. This nonstationary element cannot be a function of the mean impedance

though, as this is a seismic attribute and has no low frequency

component present due to the limitations of seismic bandwidth.

As previously

described, a model

can be added to the

experimental

variograms using the

variogram function on

the right hand toolbar.

The generate

variogram dialogue

will be displayed,

allowing selection of

the model. After

pressing Ok the

chosen model will be

overlaid on the

function window.

The variogram displays on the following pages show the model fitted to

each of the K layer horizontal variograms. The model parameters are

given in the table below. There does not appear to be a pattern with

depth. In addition, the ranges are generally larger than those

estimated from horizontal well variograms.

K Layer

1

20

40

47

60

80

100

120

140

160

180

191

Function

Exponential

Exponential

Spherical

Exponential

Spherical

Exponential

Exponential

Exponential

Exponential

Exponential

Exponential

Exponential

Range (ft)

2000

2800

2050

3400

1300

3000

2600

2000

2800

3500

1800

1600

Sill

0.50

0.78

0.91

1.00

0.23

0.94

0.65

0.67

0.75

0.95

0.88

0.62

All rights reserved

SEISMIC VARIOGRAMS

Coloured Impedance Horizontal Variograms and Fitted Models

K layer = 1

K Layer = 20

K layer = 40

K Layer = 47

K layer = 60

K Layer = 80

All rights reserved

SEISMIC VARIOGRAMS

Coloured Impedance Horizontal Variograms and Fitted Models

K layer = 100

K Layer = 120

K layer = 140

K Layer = 160

K layer = 180

K Layer = 191

All rights reserved

FACIES SIMULATION

SEQUENTIAL INDICATOR SIMULATION

First step for performing any facies simulation is

to make a copy of the upscaled Facies object.

Rename the copy Facies SIS. All the data

analysis settings will be retained if you copy the

Facies object for which the data analysis was

performed.

Double click Facies Modelling to open the facies

modelling dialogue.

tab. Of the three buttons, select

on the first button (use variograms

from data analysis) and the third

button (use vertical proportion

curves from data analysis). Make

sure that you did model the

vertical proportion curves for both

sand and shale and that the

variogram models were copied to

all facies.

Choose the modelling method as

Sequential Indicator Simulation.

For this first example we will

ignore any lateral trends.

There is a quality control question

here: there is no way to verify

which variogram models are

actually picked up by the program,

which is disconcerting.

Pressing Ok will run the SIS

simulation.

All rights reserved

FACIES SIMULATION

K=41

K=42

K=43

K=44

K=45

K=46

K=47

K=48

K=49

K=50

K=51

K=52

K=53

K=54

All rights reserved

FACIES SIMULATION

QUALITY CONTROL OF SIS

When the simulation has completed a number of checks should be made of

the results:

Check vertical proportions

Check thickness distribution by zone

Checking thickness distribution between (l-r) original logs, upscaled and SIS cells

All rights reserved

FACIES SIMULATION

SIS WITH SEISMIC CONSTRAINT

The sequential indicator simulation can also be run using a 3D probability

cube as a constraint. In the Stratton Field dataset a stochastic seismic

inversion has been run to generate impedance realisations. These

realisations have then been used to generate a probability cube based on an

impedance cutoff of 8,000 (described in the Stratton Field introduction section)

in Petrel you must specifiy a

probability cube for each facies.

The sum of the cubes should be 1.

The sand probability cube is

turned into a shale probability

cube in calculator where Pshale = 1

Psand. Copy the Facies property

to a new property called Facies

SIS seismic.

Turn off the vertical proportion

from Data Analysis button. For

each facies turn on the Probabiity

from property option and assign

the correct probability property for

each facies. Repeat for all zones

and facies.

All rights reserved

FACIES SIMULATION

K=41

K=42

K=43

K=44

K=45

K=46

K=47

K=48

K=49

K=50

K=51

K=52

K=53

K=54

Sand Probability from stochastic seismic inversion for K Layers 41 54 Zone B46

All rights reserved

FACIES SIMULATION

K=41

K=42

K=43

K=44

K=45

K=46

K=47

K=48

K=49

K=50

K=51

K=52

K=53

K=54

All rights reserved

FACIES SIMULATION

OBJECT MODELLING

For the Stratton Field we have just two facies to model. The depositional

environment is predominantly shale (88 - 95%, depending on zone) with fluvial

sands only comprising 5 12% of the facies. The first step in object

modelling is to select the background facies, clearly for Stratton Field the

shale is the appropriate background facies.

Next we should consider the objects to be modelled. We can summarise the

sands as:

Fluvial Channels

East-west orientation

5 30 ft thickness

Composite widths of up to 2,500 ft

For the object modelling we have to set distributions (PDFs) describing the

channels. For thickness, we can check the actual sand thicknesses.

Thickness distribution of sands in B46 Original logs (left) and upscaled cells (right)

distribution for sand thickness. Remember that we need to consider the

thickness for our objects. Do the thicknesses refer to composite,

amalgamated channels or individual channels? Which are we modelling?

Petrel does not offer an exponential distribution for modelling, but we can

approximate this distribution by using a truncated normal distribution.

All rights reserved

FACIES SIMULATION

Although we have a geological description of channel complexes having

widths up to 2,500 ft, we should also consider any other evidence such as

seismic data to support this. An example channel is delineated in a stochastic

inversion realisation at K layer 47 in the display below. The channel appears

to vary in width between 400 1200 ft.

Again, copy the original property Facies to a new property Facies Objects and

double click to edit the facies simulation. Select object modelling. Next, on

the Background tab set the background facies to shale. Select the Facies

Bodies tab. Under the Settings (lower tab) select sand and use the estimated

fraction. Note other settings.

All rights reserved

FACIES SIMULATION

Next select the Layout (lower tab). Here we can set distributions describing

the orientation, amplitude and wavelength of the channel. The example

shown has selected normal distributions for each. Orientation is East-West

with 95% of the channels expected to be oriented between 80 and 110

degrees. The Drift parameter allows for change in the channel characteristics

along the channel.

On the Channel (lower tab) set the width and thickness distributions.

ensure that you select to use the

vertical proportion curves from data

analysis.

For the first run, the parameters

selected have been copied to all

zones. Select Ok (or Apply) to run the

object modelling.

All rights reserved

FACIES SIMULATION

QUALITY CONTROL OF OBJECT MODELLING

Check the results of object modelling with

care. Because the method is based on

accept/reject methods and the parameters

such as channel widths and thicknesses are

often difficult to estimate, the object

modelling may not converge or result in a

model with poor reproduction of sand facies

and not correctly honour the well data.

The first output from object modelling is a

listing showing the progress of the algorithm

in each layer. Inspect this carefully noting

the following:

Well match

Facies Fractions

Number objects inserted/rejected

The well match shows how many channel cells in the wells have been

modelled with a channel. In this example it is 40%, a bad result. Compare

the final facies proportions with the observed proportions or your desired

proportions (remember, this may be determined by you and related to the size

of model). Here we have a sand fraction of under 6% but the log data has a

proportion of nearly 10%. Check the number of objects inserted and rejected.

The results above indicate we have problem and will need to reconsider our

parameters.

As for SIS you should also check the reproduction of the vertical proportions

and thickness distributions using Data Analysis. In addition, inspect the

results visually. Do the channels fit with your geological concept?

Because object methods involve filling large amounts of model space with

each object inserted, note there is a relation between the choice of

parameters for channel dimensions, facies proportion and the ability of the

algorithm to condition to both the facies proportions and to the well control

points. Any disparity in the results will suggest that the model parameters are

incompatible with the data in the wells and revision of the parameters is

required.

Object Modelling Example Run 1

We will demonstrate object modelling and the requirement to update the

parameters in an iterative manner. We will begin with the following initial

parameters in all zones:

All rights reserved

FACIES SIMULATION

PDF

Layout Parameters

Normal

Orientation

Normal

Amplitude

Normal

Wavelength

Channel Parameters

Normal

Width

Thickness

Truncated Normal

Drift

Min

Mean

Max

0.1

0.1

0.1

90

1000

4000

5

200

500

800

500

0.1

For the purposes of this exercise we will only investigate the first two layers,

MFRIO and B46. In order to provide a realistic practical, we force an

incompatibility in the parameters by falsely setting the target proportion in the

B46 to be 5.5% (actually the proportion for the MFRIO) when the true

proportion is 9.6%. This can be done by first setting the MFRIO zone and

then copying parameters to all zones an easy mistake to make!

After running the object modelling, we first inspect the algorithm listing output.

For the MFRIO interval the results look excellent, with a 100% match to the

wells and a facies proportion of 5.8 %, close to the target proportion of 5.5 %.

Object model listing for MFRIO (left) and B46 (right) after Run 1

All rights reserved

FACIES SIMULATION

For the B46 zone there is clearly a problem. Although we have reached the

target proportion, the well match is very poor with only 52% of well channel

cells intersected by a channel object. This suggests that the inserted objects

are either too large, causing the algorithm to stop when the facies proportion

is reached in the zone but without correctly conditioning to all the wells, or that

the facies proportion is wrong (which it is, of course).

Checks should also be made of the reproduction of the vertical proportion

curves and the thickness distribution. The various comparisons are shown

below. The results for the MFRIO are generally good, although the thickness

distribution might be improved.

Checking thickness distribution between original logs, upscaled and object cells for MFRIO

All rights reserved

FACIES SIMULATION

For the B46 the vertical proportion curve is not properly reproduced, being

rather patchy. The thickness distribution is also poorly reproduced, with an

underestimate of the thicker part of the distribution.

Checking thickness distribution between original logs, upscaled and object cells for B46

We will proceed by attempting to improve the well match in the B46 zone by

modifying the channel parameters. For our second run the following

parameters have been used in the B46, changing the width and thickness

parameters:

Layout Parameters

Normal

Orientation

Normal

Amplitude

Normal

Wavelength

Channel Parameters

Normal

Width

Thickness

Truncated Normal

Drift

Min

Mean

Max

0.1

0.1

0.1

90

1000

4000

5

200

500

300

100

0.1

With the modified parameters we obtain a better fit to the wells in B46, Run 2

resulting in a match of 78%.

All rights reserved

FACIES SIMULATION

The vertical proportion is

improved compared to Run1,

partly because more channel

has been inserted, intersecting

with the wells. The thickness

distribution is now much

improved, although thicknesses

could still be reduced.

To try and improve the match to the thickness distribution, the thickness range

has been reduced still further.

Layout Parameters

Normal

Orientation

Normal

Amplitude

Normal

Wavelength

Channel Parameters

Normal

Width

Thickness

Truncated Normal

Drift

Min

Mean

Max

0.1

0.1

0.1

90

1000

4000

5

200

500

300

100

0.1

All rights reserved

FACIES SIMULATION

The result is a 94% match to

the well data. The vertical

proportion curve reproduction is

improved and the thickness

distribution shifted as expected.

layers 41 54 in the B46 interval for the three different. Although the final

Run 3 improves the numerical fit, it clearly results in thin channels that are not

consistent with the geological model. A good moment to check everything

carefully and realise that the target facies proportion in the B46 zone is set too

low!

All rights reserved

FACIES SIMULATION

K=41

K=42

K=43

K=44

K=45

K=46

K=47

K=48

K=49

K=50

K=51

K=52

K=53

K=54

All rights reserved

FACIES SIMULATION

K=41

K=42

K=43

K=44

K=45

K=46

K=47

K=48

K=49

K=50

K=51

K=52

K=53

K=54

All rights reserved

FACIES SIMULATION

K=41

K=42

K=43

K=44

K=45

K=46

K=47

K=48

K=49

K=50

K=51

K=52

K=53

K=54

All rights reserved

PETROPHYSICAL MODELLING

PETROPHYSICAL MODELLING

After simulating the facies we need to populate each facies with appropriate

estimates of porosity and permeability. The most common approaches to

used for porosity are:

Simulate porosity (SGS)

Incorporate trends such as those from seismic inversion or attributes

estimation is more problematic and includes

Linear transform of logs or upscaled cells and then mapping

Simulation or co-simulation of above

For the purposes of training, we will run each algorithm using only the porosity

values defined fort he sands, but view the results across all cells in the model.

We can then see the results of each method without any clipping by the facies

simulation. Of course in reservoir modelling, the porosity is estimated or

simulated for each facies and the cells in the model are populated with the

appropriate porosity for the facies simulated in that cell. In this case, a typical

approach would be to select an estimation method to populate the sand facies

and to assign a constant value of porosity to the shale facies.

Typical modelling by facies: Estimation within sands (left) and assign value to shales (right)

All rights reserved

PETROPHYSICAL MODELLING

The methods for filling the cells are referred to as interpolation in Petrel and

include:

Kriging

Sequential Gaussian Simulation

Although there are a whole range of methods within Petrel for creating

surfaces we will make comparisons between a subset comprising the

following algorithms:

Parabaloid Surface

Bilinear Surface

Kriging

Kriging (GS-LIB)

Sequential Gaussian Simulation (SGS)

It should be noted that there are two kriging implementations available within

Petrel. There is an internal implementation and an external implementation

which uses the GS-LIB code executable directly. The manual describes the

differences as:

Petrel internal Kriging implementation

Works in XYZ

Only considers data within the variogram range

Faster because no transfer of data to external algorithms

The hints menu for the internal kriging implementation states the following:

Since kriging in Petrel uses all input observations, the number of

observations should limited. Algorithm will be slow with too many

observations (> 100).

This appears to directly contradict the claim that the internal algorithm is

faster. Also the statement that it only considers data within the variogram

range is a matter for concern the manual notes that this may lead to strange

effects in areas with no data when trends have not been removed correctly. If

the algorithm really only uses data out to the range of the variogram, this is

likely to be undesirable behaviour that will give problems with the estimation of

the mean.

GS-LIB Kriging implementation

Works in IJK

Control of advanced settings especially neighbourhood search

Co-located cokriging available

All rights reserved

PETROPHYSICAL MODELLING

The GS-LIB implementation gives

access to the choice between simple or

ordinary kriging, which is not usually too

critical. Recommend simple kriging as

best default as it will ensure the correct

mean. The search parameters should

ideally use all the samples available

within each K layer plus some vertical

sample inclusion across the K layers

(but not the zones). This is not

straightforward to achieve with the GSLIB program.

The GS-LIB implementation is consistent

with the SGS implementation. It makes

sense to use the algorithms in

conjunction. There are some bad

interface bugs and the program does not

always allow the estimation to proceed

even though the parameters are

reasonable.

Before proceeding to

make a comparison of the

various algorithms, the

geostatistical options will

require the preparation of

variograms as a

prerequisite.

In Data Analysis we have

already described and

applied a 1D vertical

depth trend and normal

score transform to the

porosity data. In Data

Analysis we chose to

work with the porosity

from all zones together

which is a good choice for

this data but is not a

general recommendation

for all data sets.

The vertical variogram for all zones is shown above. The variogram shows a

large peak at the intermediate lags, with the variance above the standardised

All rights reserved

PETROPHYSICAL MODELLING

1.0 variance. This

corresponds to the low

porosity values in the mid

zones of the data set (see

1D trend to right).

As previously described,

variogram analysis should

include both

omnidirectional and

directional variograms.

Firstly, we begin here

looking at these

variograms for all zones

combined. The

directional variograms are

not straightforward to

analyse.

Vertical 1D trend in porosity

All rights reserved

PETROPHYSICAL MODELLING

Further analysis looks at omnidirectional variogram, for all zones and by each

individual zone. The results are inconclusive. The vertical variogram is well

defined but not the horizontal. One alternative is to consider the seismic

properties as previously. But is this a proxy for facies, porosity or both?

All rights reserved

PETROPHYSICAL MODELLING

Parabaloid Surface

Select Functional (Interpolation) and

choose to follow layers (this would

be a usual practice). Select

parabaloid and use the inverse

distance squared weighting. The

accompanying figures show a 3D

view and layer view of the resulting

interpolation. The key data to

consider, however, is what happens

to the histogram. All estimation

methods are smoothers and cause

the histogram to be squeezed

towards the mean. This poses a

problem both for volumetric

calculations and permeability

prediction.

All rights reserved

PETROPHYSICAL MODELLING

Bilinear Surface

Select Functional (Interpolation) and

use the follow layers option. Select

Bilinear surface and inverse

distance squared weighting. Result

is shown as 3D view to right and

layer map and histograms for

comparison are shown below.

All rights reserved

PETROPHYSICAL MODELLING

Kriging (Petrel Implementation)

Select Kriging (Interpolation). There

is no follow layers option as this

implementation works in XYZ.

Select to use the variograms from

data analysis.

Note how the kriged surface

appears as a series of halos or

bulls-eyes around the well data

control points. We have discussed

the objectives of estimation

previously and described this

behaviour.

3D View of Kriging (Petrel) interpolation

All rights reserved

PETROPHYSICAL MODELLING

Kriging (GS-LIB Implementation)

Select Kriging by Gslib

(Interpolation). There is no follow

layers option as this implementation

works in IJK (Simbox). Select to

use the variograms from data

analysis. The expert tab allows the

choice of simple or ordinary kriging

and options to edit the number of

cells to use.

The results seem to show some

artefacts which may be related to

XYZ domain for calculation. Note

also that the mean is different to that

from the previous example. This is

due to neighbourhood search

differences.

All rights reserved

PETROPHYSICAL MODELLING

Sequential Gaussian Simulation

Select SGS. There is no follow

layers option as this implementation

works in IJK (Simbox). Select to

use the variograms from data

analysis. The expert tab allows the

choice of simple or ordinary kriging

and options to edit the number of

cells to use. This algorithm is

consistent with the Kriging (GS-LIB )

algorithm.

Remember simulation will reproduce

variograms and the histogram

correctly (see below), so the colour

range is larger. It is shown in full in the 3D view and on the same range

View of Kriging (GS-LIB) interpolation

facilitate comparison.

All rights reserved

3D

PETROPHYSICAL MODELLING

For display surfaces, the user might consider any acceptable functional or

gridded surface. For the purposes of reservoir modelling and the resulting

fluid flow simulation, the only algorithm which should be considered is SGS.

As we have seen, this is the only approach which correctly reproduces the

variability (in the form of the histogram) and the spatial connectivity through

the variogram. Both of these are essential properties of an attribute to

reproduce if we intend to use the model for fluid flow simulation.

PERMEABILITY ESTIMATION

So far the analysis has considered only porosity estimation. In common with

many reservoir models, the Stratton Field data set does not have permeability

data loaded. The permeability data is generally only defined on core data and

is usually analysed through the use of permeability-porosity crossplot to

establish a relation between permeability and porosity.

Having established a relation, a permeability field can be estimated by

applying the transform at one of several points in the typical work flow.

Transform mapped porosity to permeability

Transform simulated porosity to permeability

None of these methods are suitable for constructing reservoir models suitable

for fluid flow simulation. Permeability should be co-simulated with porosity.

Permeability-Porosity Crossplots

Consider the crossplot to the

right. The line has been fitted

using least squares and is a

good estimator for predicting

permeability given porosity.

For a net reservoir calculation,

we might apply this relation to

the porosity well logs to

estimate a permeability log,

truncate at our cutoff of 1 mD

and estimate net pay.

This action is identical to using

the regression line to convert

our permeability cutoff to a

porosity cutoff (9.5%) and

truncate the porosity logs. The

net predicted from the well logs would be the same with either approach.

Unfortunately although this is a widespread and almost universal practise it is

All rights reserved

PETROPHYSICAL MODELLING

wrong and the net prediction will be biased. If we apply the porosity cutoff of

9.5% to the crossplot we would count samples in quadrants B+D as

contributing to net, a proportion of 92.5% of the total samples. This is what

we are doing when we apply the transform to the well logs (whether converted

to permeability or remaining as porosity). If we were to apply the permeability

cutoff to the crossplot we would include samples in quadrants A+B as net pay.

This gives 82.5% net, a significant change. Cutoff calculations cannot be

applied to the output from an estimation procedure because the estimator

causes the histogram to change and leads to bias. In the diagram above,

using the transform to predict permeability will result in the lowest estimated

peremeability being predicted from the lowest porosity. Similarly our highest

predicted permeability will be from the highest porosity. A glance at the graph

shows that the predicted permeability range must be narrower than the true

permeability range hence the predicted permeability will have a narrower

historgram about the mean.

The problem is exacerbated because the function is fitted with symmetric error

terms under a log transform of porosity. This results in asymmetric errors on

permeability, with large permeability values significantly underpredicted as

compared to a slight overprediction of low permeability values.

The correct solution is to use a simulation method which, while following the

form of the relation defined by the transform, reproduces the variability and

hence the histogram of permeability. In addition, we need to preserve the

correlation between porosity and permeability.

10000

0.3007x

y = 0.0204e

2

R = 0.6087

1000

Permeability (mD)

used for Stratton Field are

shown here. Note that

permeability is not available for

the Stratton data set in the

public domain and so these

data have been borrowed

from another field. The mean

porosity and permeability have

been shifted to be consistent

with the Stratton Field well log

porosity data. The correlation

is good with R=0.78.

100

10

0.1

5

10

15

20

25

30

35

40

Porosity (%)

The following layer displays

show the result of applying the

transform defined by the regression line shown on the crossplot to two of the

porosity estimations previously described:

SGS

All rights reserved

PETROPHYSICAL MODELLING

These displays will be identical to the porosity displays but the colour table

annotations have changed in to permeability units.

From the cells in the model we can view the poro-perm relation for the above

two examples. Note that the scatter is not reproduced and the permeability

range is highly restricted, although there is more dynamic range in the SGS

result as it correctly reproduces the porosity histogram.

All rights reserved

PETROPHYSICAL MODELLING

A proper approach to generate a suitable permeability field would be to cosimulate with an SGS simulated porosity. True co-simluation is not available

in Petrel, but the co-located co-simulation is an adequate alternative.

Simulating in this way would require permeability logs to be available at the

wells. It would not be correct to substitute linear transforms of the porosity

logs as this forces an exact relation between porosity and permeability at the

wells.

Correlated Permeability

An alternative compromise, using just the permeability-porosity crossplot

information is to simulate a separate but correlated permeability field (not

based on well logs ie non-conditional). This is then coupled to a porosity field

produced using SGS via a joint relation. This method reproduces the form of

the linear transform as well as the correlation coefficient and allows the

permeability field to include a separate spatial correlation function.

The procedure is as follows:

Make a copy of the porosity called X2

Petrophysical Modelling select Common and tick Use Filter

All rights reserved

PETROPHYSICAL MODELLING

analysis

Turn off PDF and depth trend

from data analysis

Simulate a normal distribution

with a mean of 0 and standard

deviation of 0.0385 (This is the

variability of the residual porosity

without the depth trend).

Set min and max to absolute and

use values of 5

filter and then use the calculator to

combine the non-conditional porosity

with the SGS porosity. Note we also

have to correctly account for the depth

trend.

Recall from the statistics section that the

correlated part is a combination, weighted by R2, of the residual parts only,

without the mean. Here we are modelling a non-stationary mean in the form

of a vertical depth trend.

All rights reserved

PETROPHYSICAL MODELLING

The result has been called Porosity_X3, which is a porosity field correlated

with the SGS porosity field. Porosity_X2 and Porosity_X3 can be compared

to SGS porosity to ensure they have consistent histograms.

permeability (Perm_X3) using our linear transform. A crossplot shows that we

now have a correlated permeability field rather than a linear functional

dependence on our simulated SGS porosity. We may not be entirely happy

with the reproduction of the permeability range or the correlation coefficient,

so these can be adjusted empirically by adding a porosity correction, this

could be to the mean in the calculator option, by changing the standard

deviation in the non-conditional simulation or a combination of both.

All rights reserved

PETROPHYSICAL MODELLING

All rights reserved

PETROPHYSICAL MODELLING

Permeability histograms by transform: Kriged (Petrel) (left) and Kriged (GS-LIB) (right)

Permeability histograms by transform: SGS (left), Perm_X3 (centre) and Perm_X4 (right)

The final problem with reservoir modelling which we should mention is what to

do after simulating, say, 100 realsiations of a million cell model? With current

technology we cannot run fluid-flow simulation on such a large problem and

certainly not on so many realisations. After undertaking stochastic reservoir

modelling there is still a tendency to choose just one realisation for upscaling

and simulation (the first one generated, sometimes), or just a few realisations.

This poses a real problem is there any point in generating multiple

realsaitions if we can only afford to perform fluid-flow calculations on one?

One solution is to select a few realisations on the basis of static properties

such as total volume or a connectivity index, choosing realisations with low,

intermediate or high values on these criteria. This gives some confidence that

a range of possible models is being tested. However, it should be

remembered that the reason for fluid-flow simulation is partly because statis

properties are not good proxies for dynamic properties.

A strategy for reducing the number of realisations whilst testing many

parameters is to use the latin hypercube. By crossplotting static properties we

can reduce the number of tests if the statis properties are correlated. This

principal is illustrated below. On the left the two properties extracted from the

realisations are crossplotted. Clearly a low value of reservoir volume could

correspond to low, intermediate or high connectivity in the realisations. This

shows that we would have to test more realisations in order to investigate the

two parameters, perhaps 9 realisations.

On the right we have a clear correlation between the parameters and this

reduces the number of realisations we might test. If we fluid-flow through a

low reservoir volume realisation we are also testing a low connectivity index

realisation.

All rights reserved

PETROPHYSICAL MODELLING

All rights reserved

SEISMIC CONDITIONING

CO-KRIGING AND CO-SIMULATION

Petrel offers both co-kriging and cosimulation based on the GS-LIB

algorithm. The co-kriging or cosimulation is constrained by a secondary

variable. The weight placed on this

constraint is controlled by the correlation

coefficient under a Markov-Bayes model.

The secondary variable can be either a

property, horizontal surface or a vertical

function. Petrel offers two choices of

algorithms:

Co-located

coefficient is used to calculate the

influence of the secondary variable. No

scalar information is used, so the

secondary variable can be in any units.

The only requirement is that there is a linear relation between the property

estimated and the secondary variable.

In the locally varying mean method, the constant mean assumed for a simple

kriging is replaced by the mean of the secondary variable.

The implementations in Petrel are directly from GS-LIB and both work

correctly. Problems usually arise because of misunderstanding over the

objectives of these methods.

Stratton Field Example

As an example, we will use colocated co-simulation to map

porosity using seismic coloured

impedance as a constraint. We

have previously observed a

correlation between porosity and

seismic coloured impedance.

The observed correlation

coefficient is R=-0.291.

However, Petrel correctly

requires the correlation

coefficient after both variables

are normal score transformed, which can be found using the estimate option.

The correct correlation coefficient is R=-0.275.

All rights reserved

SEISMIC CONDITIONING

All rights reserved

SEISMIC CONDITIONING

The influence of the seismic property is clearly seen. Remember the

correlation is negative. The porosity around wells 01 and 16 is clearly

increased and the porosity west of wells 12 and 17 reduced by introducing the

seismic constraint. A check of the histograms shows that the distribution is

modified by the seismic constraint.

The use of deterministic seismic inversion to generate quantitative seismic

impedance estimates is becoming common as an input to reservoir modelling.

Determinisitic seismic inversion combines data from wells and seismic to

create a broad bandwidth impedance model of the earth. However, we must

always remember that the seismic resolution is at a coarser scale than the cell

size of most reservoir models and so the seismic is usually informing the

model of variations in the average properties over a zone, zones or part zone,

depending on the actual zone thickness and seismic resolution.

In the vertical direction, the typical averaging from seismic inversion is over a

thickness of perhaps 20 m (10 50m, depending on resolution). In the

horizontal direction, seismic averages over a typical 50 x 50 m are. It is

important to note that the seismic time sample rate and trace spacing do NOT

define the resolution. Resolution is determined by frequency content and

bandwidth.

Seismic data is bandlimited. In particular, it

does not contain low frequencies and

therefore absolute impedances cannot be

recovered directly from the seismic trace.

All inversion schemes with an absolute

impedance output require a low frequency

model or constraint. The low frequency

scalar is usually obtained from interpolation

of well data, stacking velocities or a

combination of these. After inversion the

low frequency model is embedded in the deterministic inversion. Artefacts in

the low frequency model manifest themselves as equivalent artefacts in the

deterministic inversion. Conditioning a reservoir model using this type of data

All rights reserved

SEISMIC CONDITIONING

is equivalent to conditioning to a map of the wells. For this reason

deterministic inversions should NOT be used to condition reservoir models.

amount to little more than a

relative impedance (from the

seismic) added to a low

frequency model. This can be

easily demonstrated by filtering

the low frequency out of the

inversion result and comparing

with relative or coloured

impedance impedance.

An example is shown in the

accompanying figures. A

commercial sparse spike

inversion algorithm has been

used to compute absolute

elastic impedance.

By applying a low pass filter,

the low frequency model is

obtained. Artifacts

(bullseyes) around the well

locations can be clearly seen.

These arise because the low

frequency is simply an

interpolation of the wells, just

like a 3D property estimated

using kriging or bilinear

interpolation in Petrel.

A further concern in this case

is revealed by the application

of a high pass filter. This

removes the well interpolation

element and shows the

contribution to the inversion

obtained from the seismic.

In this example, this seismic

contribution has been

compared to a simple relative

impedance estimate obtained

by simple processing of the

seismic data (similar to the

coloured impedance used in

All rights reserved

SEISMIC CONDITIONING

the Stratton Field example.

Apart from a slight smoothing

caused by the application of a

trace mix in the full

deterministic inversion, the two

displays are indistinguishable,

suggesting that the

deterministic inversion has

made no contribution in this

case.

deconvolution to zero phase. By integrating the seismic traces output from

these processes we obtain relative impedance which is not contaminated with

the artefacts often caused by efforts to obtain absolute impedance by

including a low frequency constraint.

It is very important to note that deterministic inversion data should NOT be

used to constrain a reservoir model. Deterministic inversion data MAY be

used if the low frequency component is filtered out to remove the influence of

the wells from the deterministic inversion. A seismic attribute or relative

impedance (such as coloured inversion) CAN be used. The ideal

impedance product to constrain a reservoir model would impedance

realisations generated from a stochastic inversion. It is important to note

that some stochastic inversion methods (such as Promise) produce output

similar to deterministic inversion schemes and the same caution is advised

before using these as conditioning for the model.

Our ideal for constraining a reservoir model with 3D seismic data would be to

have the model fully consistent with pre-stack seismic amplitude information.

Currently there are several possible approaches considered for solving this

problem.

Probability transforms from seismic to rock properties are simple to apply

and valid but do not take into consideration the spatial elements of the

property. This is essentially a 1D PDF method.

Stochastic seismic inversion, either pre- or post-stack can be used to

generate impedance realisations which correctly include the uncertainty in

seismic inversion. These impedance realisations can be included directly into

a Petrel model using the co-located option or converted to probability

volumes. With current technology this is the most practical method.

All rights reserved

SEISMIC CONDITIONING

seismic properties using a rock physics model and then forward modelled to

produce a synthetic seismic response. This response is then compared to the

observed seismic data. Difference will show where the model requires

updating. Currently, updating the Petrel model is very difficult to do.

Algorithmically, to update the model comprehensively and consistently require

huge computer processing capacity. These methods, while ultimately our

objective, are not yet practical.

The following examples were prepared by Chris Townsend and show the

problems of attempting to condition a reservoir model using seismic

impedance data. In this instance the constraining data comes from Promise.

Although Promise uses a stochastic methodology in the inversion, the results

input to Petrel are equivalent to a layer average map as would be obtained

from deterministic inversion.

Aside from the fact that the embedded well models contained in a

deterministic inversion mean this data should not be used to constrain a Petrel

model, attempts to map using a seismic constraint such as from deterministic

inversion (and this includes the Promise output) will be also compromised due

to the inversion data being a layer average property. The problem should

properly be addressed by accounting for the scale change between the model

cell estimates and the layer average from seismic. Petrel is not able to do

this.

The first action is to copy the seismic

constraint into the model. With a 2D

surface such as from Promise, this

should be copied to each layer within

the zone.

Using the co-kriging option, the SGS

porosity modelling is constrained by the

seismic data.

In map form, looking at the zone

average, the results appear quite

reasonable with the map following the

seismic trends. However, note that

when using the co-located option the

porosity magnitudes from the seismic

data are ignored, the porosity scaling

coming from the well data only.

All rights reserved

SEISMIC CONDITIONING

Porosity average compared to seismic trend: zone average map (left), section (right)

do we see the problem, with the

values across all layers following the

same trend. The algorithm is working

correctly. What we have is a classic

problem of support. We really want

only the SGS average for the whole

zone to be following the seismic

average, not for every layer

We should also note that the well data does not have the same histogram as

the porosity from the seismic property (see right). This is because the seismic

porosity map contains a smooth estimate due to kriging the prior constraint

(the low frequency part mapped from the wells).

The alternative might be to use the locally varying mean (LVM) option in

Petrel. If the well data, cells and seismic constraint were all on the same

support this would be the preferred option. Because they are not the results

are unsatisfactory.

Three LVM options are compared

LVM only

LVM plus Transform using input distribution

LVM plus variance reduction

All rights reserved

SEISMIC CONDITIONING

Using LVM only gives a large range of values. Using the transform option

gives a good reproduction of the spatial trends from the seismic but with a

much reduced dynamic range observed in the histogram. The variance

reduction option allows some control over the final histogram but the

reproduction of the spatial trends is poor. The figures below compare the

histograms for the transform and variance reduction methods to the upscaled

cells.

Comparison of histograms: LVM with Transform (left) and with Variance reduction (right)

The following displays compare the LVM and co-located options with SGS

only and the seismic porosity constraint in both map and cross-section form.

All rights reserved

SEISMIC CONDITIONING

All rights reserved

SEISMIC CONDITIONING

APPROXIMATE SUPPORT CORRECTION

There is currently no algorithm available to correctly constrain SGS or kriging

to a seismic property where there is a large change of support. Current

methods are designed to work where the well samples and seismic data are

at the same support. They work correctly where a layer average in the wells

is mapped using the seismic property as a constraint where both are defined

over the same interval. The thickness of that interval is determined by the

seismic frequency content and bandwidth.

It is possible to devise an approximate work around to this problem as follows:

At each well compute the vertical well average over the interval

At each well subtract the vertical well average from the individual

samples to obtain a residual in each cell

Use SGS to simulate the residuals into the 3D grid for the layer

Calculate the vertical average map of the simulated residuals

Krige a map of the well average property for the interval constrained by

the seismic property using co-located co-kriging or locally varying mean

For each cell subtract the value of the vertical average map in the

same vertical position and add the value of the kriged well average

map (from the previous step)

Result is a simulated property with the correct histogram of the original well

properties but with a vertical average corresponding to the trends from the

seismic property. This method assumes that the residuals are strictly

stationary and structured. The method can be run manually in Petrel.

It is very important to note that this method is a support correction only. In no

way to change the previous comments concerning the use of deterministic

inversion data as a constraint to a reservoir model. Deterministic inversion

data (and this includes Promise output) should NOT be used to constrain a

reservoir model. The following seismic data CAN be used to constrain a

reservoir model, with an appropriate support correction:

Seismic attributes

Relative or coloured impedance

Stochastic inversion impedance realisations

Stochastic inversion probability volumes

All rights reserved

SEISMIC CONDITIONING

The final set of displays

clearly illustrates the

danger of using

Promise data as a

constraint to a reservoir

model. First map

shows the porosity

estimate for an interval

output from Promise.

Second display shows

the prior model (low

frequency) which is

input to Promise. It is a

map of the wells only

that gives the expected

value at each cell. This

cannot be updated in

the inversion because

the seismic does not

contain any low

frequencies and

therefore carries no

information about

variation in the mean.

This surface is already

estimated when we run

co-kriging in Petrel. To

include it in the seismic

constraint counts it

twice and substantially

underestimates the true

uncertainty

The third map shows

the difference between

the Promise output and

the prior model. This is

the information which

the seismic contributes

and it is this that can be

used to constrain the

model.

Note that the residuals

are the prior output,

which gives them the

wrong sign.

All rights reserved

SEISMIC CONDITIONING

A check of the crossplots between the well data and the maps at each stage

is quite revealing. The first crossplot (left) is between the well average and

the Promise output which is what would be observed in Petrel if using the

Promise data. The correlation is 0.935, giving the impression that the seismic

is a good predictor of porosity. The second crossplot (centre) is between the

prior model in Promise and the well average, demonstrating that the strong

correlation of the previous crossplot arises because we are largely comparing

the same information, not because the seismic is an excellent porosity

predictor. The final crossplot (right) shows the correlation between the well

average and the seismic property. The correlation is 0.494. With 29 wells this

is a statistically significant correlation and so there is justification to use this

residual as a constraint in our reservoir (with a suitable support correction, of

course).

All rights reserved

Overview of Seismic Inversion Workflow

A typical seismic inversion workflow, illustrated opposite, is based on a phased

approach. Subsequent phases depend on completion of previous phases of work.

The end of each phase is a convenient break-point at which the work and results can

be evaluated with the client and a decision made to proceed to more sophisticated

analysis in subsequent phases.

Phase 1 involves the most time consuming aspects of any inversion study.

Objectives of Phase 1 include preparing the well logs, investigating relationships

between impedance and reservoir properties and tying the well logs to the seismic.

After tying to the seismic, the well log data is used to estimate a seismic wavelet. By

application of zero phase deconvolution a broad-band zero-phase dataset is obtained

which forms the input to coloured inversion (Lancaster and Whitcombe, 2000).

Coloured inversion converts the seismic data to a relative impedance data set. The

advantages of coloured inversion are the speed of calculation and avoidance of

artefacts that may be introduced by a model. Coloured inversion, whether acoustic

or elastic impedance (Connolly, 1999), is an excellent qualitative interpretation tool.

Phase 2 attempts to obtain more resolution from the seismic and to provide absolute

impedances through the procedure often referred to as deterministic inversion. A

model of impedance is built from the wells and seismic horizon interpretation and this

model used to constrain the subsequent inversion. Our model building method is

geostatistical and involves 3D anisotropic variogram analysis and kriging. We offer

two deterministic inversion approaches, both of which are delivered as standard to

the client. The first is band-limited inversion, in which a scaled relative impedance

from the seismic data is mixed with a low frequency element from the model. The

second approach is model-based inversion technique (Russell and Hampson, 1991).

Phase 3 uses the geostatistical model constructed in phase 2 but instead of using

deterministic inversion to obtain the expected value of impedance we compute

alternative realisations of the impedance through our ultra-fast stochastic inversion

technology (Francis, 2001; 2002). The set of impedance cubes represent the

uncertainty or non-uniqueness in the inversion. The realisations can be calculated at

any resolution and reproduce the well log impedance distributions. This allows

relationships defined on the well logs to be applied directly to the impedance

realisations, something which cannot be done with deterministic inversion. In

addition, the forward convolution of each impedance realisation will match the

seismic traces. Application of classifiers or transforms from impedance to reservoir

properties, identified in Phase 1, across the impedance realisations give the typical

output from the stochastic inversion in the form of a single seismic cube representing

the probability of sand, porosity or saturation distribution through the reservoir.

All rights reserved

Phase 1

Well Log

Preparation

Wavelet Estimation

Lithology, porosity,

saturation prediction

Zero-phase

Deconvolution

Coloured

Inversion

Phase 2

Resolution Analysis

Deterministic

Inversion

Phase 3

Stochastic

Inversion

100+ Impedance

Realisations

Lithology / Porosity /

Saturation Classification

Probability

Cubes

Volume

Distribution Curves

All rights reserved

Modelbased

Inversion

Phase 1 Coloured Inversion

Well Log Preparation

After data loading and checking, well log preparation comprises a series of

procedures designed to prepare the log data for comparison with the seismic.

Typical procedures applied include log editing, application of Gassman fluid

substitution, checkshot calibration and calculation of acoustic and elastic responses.

Impedance Analysis

In order to understand the predictive capabilities of impedance, forward modelling will

be used to investigate resolution issues and sensitivity of seismic to changes in

lithology and reservoir parameters. Cross-plotting of impedances and reservoir

parameters such as lithology, saturation or porosity allow quantitative relationships

and reservoir property predictors to be established. Predictors might be defined by

impedance cutoffs, probability density functions or fuzzy classifications and may

include near and far offset information. In the example (page 7, top left) porosity is

poorly discriminated, but an impedance cutoff (> 8,150 m s-1 * g cm-3) or fuzzy

classification can discriminate clean sands, as shown by the separation on the

histograms (page 7, top right) (Francis, 1997). It is important to remember that these

relationships are defined at well log scale and can only be used for prediction based

on stochastic inversion, which reproduces the impedance distribution through

geostatistical conditional simulation.

Well Ties

After preparing the well logs and reaching an understanding of the reliabilty of

impedance for predicting reservoir properties the wells are converted to the time

domain and tied to the seismic. Our well tie procedure includes phase independent

methods based on amplitude envelope and Backus upscaling. Backus upscaling

attempts to account for wave propagation through thin-layered formations and can be

a useful additional tool in improving the tie of the well to the seismic.

At Earthworks we never apply stretch or squeeze in order to arbitrarily improve well

ties. In our experience, these practices are unnecessary and tend to propagate

noise into the wavelet estimation. This in turn tends to result in inconsistent wavelets

estimated at the different wells.

An example well tie and some of the individual estimated wavelets at different wells

are shown on page 7 (centre).

Wavelet Estimation & Zero Phase Deconvolution

Wavelet estimation proceeds through a choice of three methods:

Constant Phase

Wiener-Levinson full phase

Roy White method (White, 1980; Walden and White, 1984)

These are tested in turn to select the most suitable given the seismic data quality.

Examples are shown on page 7 (bottom), the results of de-phasing using constant

phase (CP), full phase (FP) and Roy White (RW) are compared to the raw seismic.

All rights reserved

13000

16%

12000

14%

11000

12%

10%

10000

Non-clean

9000

Clean sand

Frequency

8000

6%

7000

4%

6000

2%

5000

0%

Shale

8%

Sand

0%

5%

10%

15%

20%

25%

30%

5500

35%

Raw

6500

7500

8500

9500

10500

11500

CP

FP

All rights reserved

RW

Phase 1 (continued)

Relative Impedance & Coloured Inversion

After obtaining a wavelet estimate, the wavelet is used to design an inverse operator

to zero-phase deconvolve the seismic. The deconvolution may be a de-phase only

or a full-inverse procedure to correct both the phase and the amplitude spectra. Preand post-deconvolution seismic is compared on page 9.

After zero phase deconvolution, a simple approach to estimate relative impedance is

by trace integration. The result obtained from this straightforward method is also

shown on page 9. Generally, we do not deliver this product to clients, preferring

instead to supply the coloured inversion result shown at the bottom of page 9.

Coloured inversion is reliable, quick to produce and may be generated for acoustic

and elastic impedances, thus conveniently capturing AVO effects.

In the Stratton Field data set used here for illustration (see notes below), there is

sand characterised by higher impedances in Well-08 at a time of 1275 1280 ms.

The sand can be qualitatively interpreted in the coloured inversion display as the

mainly red interval, tracking down into the 1280 1290 ms interval either side of the

well. The same interpretation can be made on the relative impedance, but the

resolution and clarity are poorer.

At Earthworks we consider coloured inversion to be the most cost-effective,

qualitative impedance product that we can deliver to our clients. Its advantages are

ease of interpretation and, being a seismic attribute, it avoids artefacts which may be

introduced by models used to constrain deterministic inversions (Francis and Syed,

2001). However, coloured impedance is still a relative measure of impedance

changes and therefore is not suitable for use in quantitative estimation of reservoir

properties, as may have been indicated by analysis of the relationships between

lithology or reservoir properties and acoustic or elastic impedance.

It may be that for the target reservoir under study impedance cannot predict reservoir

properties quantitatively. If this is the case then consultation with the client and a

decision to stop further analysis may be taken. The coloured impedance cube would

then be the final deliverable for qualitative interpretation by the client.

Notes on Stratton Field 3D Seismic Data

The data-set used here to demonstrate stochastic inversion is the Stratton Field 3D seismic

and well log data package, prepared by the Bureau of Economic Geology, Austin, Texas,

USA (Levey et al, 1994).

Stratton Field is an on-shore gas field producing from the Oligocene Frio Formation in the

NW Gulf Coast Basin. The top of the middle Frio formation is about 1200 ms, the start of the

data shown here. There is little faulting in this interval. Depositional environment is multiple

amalgamated fluvial channel-fill and splay sandstones. The composite channel fill deposits

range from 10 to 30 ft thickness and up to 2,500 ft width. Splay deposits shown typical

thicknesses of 5 to 20 ft and are proximal to the channel systems.. Sands are generally

indicated by high impedances and have typical velocities of 12,000 ft s-1, a 30 ft sand thus

being around 5 ms thick.

Each seismic cross-section is cross-line 154 from the 3D cube. The sand maps on pages 14

and 15 show the location of this section (as a red N-S line) and its relationship to the wells.

All rights reserved

Original Seismic

Zero-phase deconvolved

+1.0

-1.0

Relative Impedance

+1.0

-1.0

Coloured Inversion

All rights reserved

Phase 2 Deterministic Inversion

Model Building

Earthworks has developed its own geostatistical model-building scheme, writing

software specifically for this purpose. The basic framework is provided by the

seismic horizons. Between the horizons, we interpolate the well impedance values

using kriging, either proportional to the seismic horizons or conformal top or base to

the horizons.

As our method is geostatistical and the same model will be used subsequently in

stochastic inversion, we analyse the data using variograms. Each layer (bounded by

the seismic horizons) is able to have its own 3D anisotropic variogram model

definition.

Vertical variogram analysis is made directly from the well log data (page 11, top left).

Horizontal variogram analysis is made from horizon slices through the relative

impedance data obtained from the coloured inversion completed in Phase 1 (page

11, top centre). If sufficient wells are available, the variogram model from the

horizontal slice analysis will be compared to the horizontal well variogram (page 11,

top right). The directional variograms define the model shape and range, the vertical

direction also defines the sill.

Deterministic Inversion

Conventional seismic inversion to absolute impedance is commonly referred to as

deterministic. We usually compute these cubes as a quality control step before

proceeding to stochastic inversion, but for some clients the deterministic impedance

cubes may be the final delivered product from an inversion study.

Band-limited inversion is a method in which the relative impedance component from

seismic is scaled and added to a low pass filtered (low frequency) model. The result

is an absolute impedance cube. Band-limited inversion is useful as a qualitative

interpretation tool, but suffers from some drawbacks. In particular, the wavelet is not

removed and resolution is not improved

In model-based inversion, the initial model is iteratively updated using generalised

linear inversion in order to obtain an optimal impedance solution whose forward

convolution is a good match to the seismic. It is a robust and reliable method, able to

remove the wavelet and, given a good initial model, improve resolution and remove

tuning effects.

There are two significant limitations of all deterministic inversion schemes (including

sparse-spike inversion). The first is that the model is embedded in the result and this

may introduce artefacts which can be misleading in interpretation (see Francis and

Syed, 2001; Francis, 2002). In order to reduce the risk of mis-interpretation we

always deliver the model in addition to the inversion results to the client and

recommend that horizon slices are compared between model and inversion.

The other limitation of deterministic schemes is that, because they produce optimal

solutions, they are unable to reproduce the full range of impedance observed in the

wells. This means that cutoffs or classifiers determined from well logs should not be

applied to the absolute impedances obtained from deterministic inversion schemes.

All rights reserved

10,000

5,000

7,600

6,100

Band-Limited Inversion

10,000

5,000

Model-based Inversion

All rights reserved

Phase 3 Stochastic Inversion

The stochastic inversion technique developed at Earthworks is a hybrid approach

which enables the computation of stochastic impedance realisations using a

conventional inversion algorithm.

Using the same seismic horizon framework, well logs and 3D anisotropic variograms

defined for kriging the initial model for deterministic inversion in Phase 2, we use a

very fast FFT-based spectral simulation method to generate impedance realisations,

conditional to the well impedance data. As necessary we can also generate coupled

conditional realisations as may be required in joint inversion for near / far offset

impedance or time-lapse studies. The initial impedance realisations are then

updated by application of the generalised linear inversion (GLI) algorithm in order to

make them conditional to the observed seismic.

The algorithm is ultra-fast, allowing routine calculation of 100+ impedance

realisations of large 3D seismic datasets without any special computer hardware

requirements. Using a conventional, high specification two processor PC the dataset

shown here takes under 5 hours to compute one hundred 3D realisations.

Some impedance realisations are shown on page 13. At the top is the band-limited

deterministic inversion from Phase 2 and to its right the original seismic data. Below

are shown four impedance realisations (left) and their forward convolution to a

synthetic seismic section (right). Comparison of the synthetic sections with the real

seismic at top confirms each realisation is conditional to the seismic.

Comparison between the impedance realisations shows how significant is nonuniqueness in seismic inversion. All realisations share a common colour table, with

the high impedance sand previously described shown as the red to blue colours

around 1280 ms. The continuity and thickness variations between just these four

realisations are highly significant and there magnitude may be surprising.

Using the impedance criteria of > 8,150 m s-1 * g cm-3 to indicate clean sands, the net

sand in the entire impedance cube has been computed from the model-based

deterministic inversion from Phase 2 and for each of the 100 stochastic impedance

realisations. The deterministic inversion gives a net sand of 8.5 % whereas the

impedance realisations range from 11.6 to 15.5 % with a mean value of 13.5 % net

sand. The wells show an average net sand of 13.2 %. The cumulative distribution

function of net sand is shown below (red curve).

100%

90%

80%

Wells = 13.2%

70%

60%

50%

40%

30%

Deterministic = 8.5%

20%

10%

0%

0.080

0.090

0.100

0.110

0.120

0.130

0.140

0.150

0.160

All rights reserved

Deterministic

Model-based Inversion

5,000

Original Seismic

10,000

Stochastic Realisations

Impedance Section

4,500

Forward Convolution

11,500

2006 Earthworks Environment & Resources Ltd

All rights reserved

Quantitative Analysis and Probabilistic Sand Prediction

The significant under-prediction of net sand from deterministic inversion is expected

from geostatistical theory. It was mentioned in Phase 2 (see page 10) that

deterministic inversion schemes are unable to reproduce the full range of impedance

as their output is optimal and therefore smoother than the impedance observed at the

wells. If we truncate the distribution and integrate samples above the cutoff (in this

case > 8,150 m s-1 * g cm-3) we will find too few samples and hence systematically

underestimate net sand. The stochastic inversions, whilst uncertain and non-unique,

do reproduce the impedance distribution and therefore there is no bias when we

apply the cutoff. By looking at many realisations, the uncertainty in net sand variation

is quantified, as shown by the distribution curve on the page 12.

We can make further predictive use of the impedance realisations. For each

stochastic inversion realisation we check if each time sample is classified as clean

sand. By counting the proportion of realisations which show the sample as clean

sand, we obtain the probability of sand at this sample. This resultant cube is

sometimes referred to as an iso-probability cube. Cross-line 154 from the cube

calculated from this data is shown on page 15 (top). The colour table shows a

chance of 50% or better of being clean sand. Using a voxel system to pick the

envelope of the sand around 1280 ms from this volume we can then compute the

isochron and hence thickness of the sand. As a comparison this has also been done

for the deterministic inversion and the two sand thickness maps are shown on page

15. The deterministic net sand map has less sand predicted and is generally thinner.

The channel in the top left (NW) corner of the P50 stochastic net sand map is nicely

defined but not evident at all in the deterministic net sand map.

The maximum sand probability map shown on page 15 is the peak probability within

the P50 sand thickness envelope. This is similar to a standard deviation map. There

is a very high probability of sand around wells with sand. From the isochron map

note the thick sand SE of Well-08. The maximum probability map shows this to be

high probability of sand too: a clear candidate for infill drilling. To summarise, the

display below is the interpreted P50 stochastic net sand map. The channel at the top

is nicely defined, together with the thick depositional trend across the centre of the

area. A possible splay is interpreted around Well-17 and an abandoned meander

(ox-bow) adjacent to Well-13. The infill drilling target is clearly indicated.

Thickest Predicted

Sand for Infill Drilling

All rights reserved

Iso-probability sand prediction P50 Net Sand

100 %

50 %

48.0 ft

0.0 ft

100 %

50 %

All rights reserved

References

Connolly, P., 1999, Elastic Impedance. The Leading Edge, April 1999, pp 438-452

Francis, A. M., 1997, Acoustic impedance inversion pitfalls and some fuzzy analysis.

The Leading Edge, March 1997, pp275-278

Francis, A. M., 2001, Understanding and improving acoustic impedance inversion.

Presented at SEG Development & Production Forum Taos, New Mexico.

Francis, A. M. and Syed, F.H., 2001, Application of relative acoustic impedance

inversion to constrain extent of E sand reservoir on Kadanwari Field. Presented

at SPE/PAPG Annual Technical conference, 7-8 November 2001, Islamabad,

Pakistan.

Francis, A. M., 2002, Deterministic Inversion: Overdue for Retirement? Presented at

PETEX 2002 Conference and Exhibition, London, UK.

Russell, B. and Hampson, D., 1991, A comparison of post-stack seismic inversion.

61st Ann. Internat. Mtg. Soc. of Expl. Geophys., pp 876-878

Lancaster, A. and Whitcombe, D., 2000, Fast-track coloured inversion. Presented

at SEG 2000 meeting Expanded Abstracts

Levey, R. A., Hardage, R. A., Edson, R. and Pedleton, V., 1994, 3-D Seismic and

well log data set: Fluvial Reservoir Systems, Stratton Field, South Texas, Bureau

of Economic Geology, University of Texas, Austin, Texas, 78713-7509, USA

White, R. E., 1980, Partial coherence matching of synthetic seismograms with

seismic traces. Geophysical Prospecting 28 pp 333-358

Walden, A.T. and White, R.E., 1984, On errors of fit and accuracy in matching

synthetic seismograms and seismic traces. Geophysical Prospecting 32 pp 871891

NB: Copies of papers by Francis may be obtained either from our website at

http://www.sorviodvnvm.co.uk or by email to ashley.francis@sorviodvnvm.co.uk

All rights reserved

- MULTI-ECHELON SUPPLY CHAIN SIMULATION USING METAMODELUploaded bybujji777
- Refference for mineral reserves and resourcesUploaded byAleksandra
- UNDERGROUND MININGUploaded byIqbal Dachi
- Manual Mine Sight GeoestadisticaUploaded byEfrain Mosqueyra
- Normal Calculations WorksheetUploaded byKrista Parnell
- Formula SheetUploaded bynblzhr
- exercises.pdfUploaded byDon Teretto
- Chapter 7 Notes(1)Uploaded byMary Zurlo Blancha
- 3360nUploaded byWilson
- SolidsHandouts3.pdfUploaded bymahreza189
- 12 Probability and StatisticsUploaded byJeenu Babu
- Statistics FormulaUploaded byUnmilan Kalita
- Mtech CS PCB 2015Uploaded bySandeep Gupta
- 5990-8882ENUploaded byAbdo Nfeniso
- THE NEUTROSOPHIC STATISTICAL DISTRIBUTION, MORE PROBLEMS, MORE SOLUTIONSUploaded byAnonymous 0U9j6BLllB
- Lecture 4 Disc r Prob f 2016Uploaded byMobasher Messi
- Explaining Sigma Shift v2Uploaded bypedroalvarez01
- Tutorial_3Q.pdfUploaded byLi Yu
- Quality ControlUploaded byKofi Daniel
- Student 1 PPT MAT Manuscript Science Achievement in Trece Martinez Final EditedUploaded byDemetrio Barrozo Dela Rosa Jr.
- Normal Distribution bUploaded byKarim Qasim
- STAT1520 Ch8 Lab Problems(1)Uploaded byAhmed Suliman
- STAT 230 - SyllabusUploaded byRakan
- Exercise Chapter4Uploaded byEunice Wong
- For math 181Uploaded byuchiha_rhenzaki
- Porfolio Construction.ppt.pptxUploaded byFiaz Ul Hassan
- First_I MTE Even Semester 2018-19_R0Uploaded byKartik Modi
- Correlation-coefficient.docUploaded bybernadeth villanueva
- Problem Set VUploaded byᗾᖺᕠᖆᘜᕬᐺ ᕼᕠᖆᓌᓍᘘ
- Mtech Cs Pcb 2015Uploaded byNilanjan Sarkar

- Sat TrackUploaded byDeni John
- Statistical Inference and Hypothesis TestingUploaded byhusain
- mpxv5004dpUploaded byJuly Andrea Cubillos Romero
- Experimental Study on Optimization of Thermal Properties of Groundnut Shell Particle Reinforced Polymer CompositesUploaded byRamya Mandava
- NONLINUploaded byPf Neiza
- Population EcologyUploaded bygenetik52
- Rr059210501 Probability and StatisticsUploaded byandhracolleges
- About 6.PDF BbbbbbbUploaded byhabte
- tmp992A.tmpUploaded byFrontiers
- Environmetrics IUploaded bydulcechic
- EA-10-16rev00 - Guidelines on the Estimation of Uncertainty in Hardness Measurements .pdfUploaded byTiago Guimarães
- Comparative analysis of hemp fibre grown in year 2010th.pdfUploaded byTony Ananda
- Seners Stroke 2016Uploaded bytripujiagustine
- Actuarial CT3 Probability & Mathematical Statistics Sample Paper 2011 by ActuarialAnswersUploaded byActuarialAnswers
- ANNALS-2013-1-30Uploaded byPuji Rahayucelalu Cimuya
- Fixed Income > Yield Curve Building With BondsUploaded byapi-27174321
- weldele et al gambusia vittata final versionpdf 1Uploaded byapi-326851466
- ch21.srmUploaded byamisha2562585
- Auditing Gray 2015 Ch 11 Sampling and MaterialityUploaded byAdzhana Aprilla
- How to Read a Systematic Review and Meta-Analysis and Apply the Results to Patient CareUploaded byTemesgen Endalew
- 2006 Planning EvalUploaded byblitzkrieg_im
- A process optimization toolbox for MRO SME Aviation OrganizationsUploaded byEmbreon
- Modeling Extreme ASA AHT Values Using Extreme Value TheoryUploaded byA Pandey
- Chapter 4 in managerial economicUploaded bymyra
- 4. Chapter 4 Demand Estimation (1)Uploaded bydelisyaaamily
- IEC 62539 - IEEE 930 - Guide for the Statistical Analysis of Electrical Insulation Breakdown DataUploaded bydankoranko
- cop04Uploaded byPinto Carlos
- Enhancing Quality of Data using Data Mining MethodUploaded byJournal of Computing
- Sample Size CalculatorUploaded bypriyanka0809
- Diagnostic performance of Body Mass Index, Waist Circumference and the Waist-to-Height Ratio for identifying cardiometabolic risk in Scottish preadolescents..pdfUploaded byMateus Carvalho