Vietnam National University - Ho Chi Minh City

University of Technology
Faculty of Geology & Petroleum Engineering
Department of Drilling - Production Engineering

Course
Geostatistics in Petroleum Engineering
Trần Nguyễn Thiện Tâm
[email protected]
 Chapter 4

Conventional Estimation Techniques

12/4/2024 Geostatistics 2
 References
Mohan Kelkar, Godofredo Perez. Applied Geostatistics for Reservoir Characterization.
Society of Petroleum Engineers, Texas, 2002.

12/4/2024 Geostatistics 3
 Contents
❑ Introduction
❑ Preliminary Considerations
❑ Linear Kriging Procedures

12/4/2024 Geostatistics 4
 Introduction
The term kriging comes from Danny Krige, a South African geoscientist, who first
applied this technique to gold mines. Later, the mathematical validity and foundation
was provided by Matheron.

12/4/2024 Geostatistics 5
 Introduction
The kriging technique uses a linear estimation procedure to estimate a value at
unsampled locations. In principle, the technique assumes that the value at the
unsampled location is estimated by 𝑛

𝑋 ∗ 𝑢𝑜 = ෍ 𝜆 𝑖 𝑋 𝑢𝑖
𝑖=1
where
𝑋 ∗ 𝑢𝑜 = the estimated value at the unsampled location, 𝑢𝑜 ,
𝑋 𝑢𝑖 = the value at the neighboring location 𝑢𝑖 , and
𝜆𝑖 = the weight assigned to the neighboring value 𝑥 𝑢𝑖 .
In other words, the estimated value is a weighted average of the neighboring values.

12/4/2024 Geostatistics 6
 Introduction
The goal, in the estimation procedure, is to calculate the weights assigned to the
individual neighboring points. These weights depend on the spatial relationship
between the unsampled location and the neighboring values, as well as the
relationship among the neighboring values. These relationships are obtained from the
modeling of the variogram.

12/4/2024 Geostatistics 7
 Preliminary Considerations
▪ Search Neighborhood
▪ Cross Validation

12/4/2024 Geostatistics 8
 Search Neighborhood
The search neighborhood defines the neighboring sample points used in estimating
values at the unsampled location.

12/4/2024 Geostatistics 9
 Search Neighborhood
In practice, we restrict the number of samples to a smaller neighborhood for several
reasons.
• Kriging algorithms involve inverting a matrix to obtain the weights (𝜆𝑖 ) described in
Eq. 4.1. The larger the number of data points selected, the larger is the matrix size. As
the matrix size gets larger, both the memory and computation time requirements
increase.

12/4/2024 Geostatistics 10
 Search Neighborhood
• As stated before, kriging algorithms involve the inversion of a matrix. If too many
sample points are used, especially clustered sample points, there is a possibility that the
matrix will become close to singular, making inversion difficult.

12/4/2024 Geostatistics 11
 Search Neighborhood
• We typically restrict variogram estimations to half the maximum distance. This
ensures that we obtain a reliable estimate of the variogram.

12/4/2024 Geostatistics 12
 Search Neighborhood
• When sample points are farther away from an unsampled location, their influence
tends to be screened by closer sample points, which are between the unsampled
location and the sample point farthest away. Therefore, additional sample points,
farther apart in the estimation process, may not necessarily improve the estimate at
the unsampled location.

12/4/2024 Geostatistics 13
 Search Neighborhood
• Kriging algorithms require the implicit assumption of first- and second-order
stationarities. As noted in the previous chapters, this assumption is rarely satisfied in
practice. Local properties can differ from sample properties. In estimating values at
unsampled locations, it is critical to choose samples that reflect the behavior of the
unsampled location. This is especially true when significant local variability exists. By
restricting the search neighborhood to local neighborhoods, we ensure that the
estimated value is much more representative.

12/4/2024 Geostatistics 14
 Search Neighborhood
In selecting an appropriate search neighborhood, the following factors must be
considered: number of data points necessary for a representative sample, size of
search neighborhood, and shape of search neighborhood. These factors determine the
final shape and size of the search neighborhood.

12/4/2024 Geostatistics 15
 Search Neighborhood
• Number of Samples
• Size of Neighborhood
• Shape of Neighborhood

12/4/2024 Geostatistics 16
 Cross Validation
Cross validation involves the estimation of values at the sampled locations so that the
estimation can be compared with the sampled values.

12/4/2024 Geostatistics 17
 Cross Validation
After cross validation is finished, we have two values at every sampled location - an
observed, or true, value and an estimated value. By comparing the true value and the
estimated value, we might be able to detect problems related to the estimation process.
We first define
𝑒 𝑢𝑖 = 𝑥 𝑢𝑖 − 𝑥 ∗ 𝑢𝑖
where
𝑒 𝑢𝑖 = the estimation error at location 𝑢𝑖 ,
𝑥 𝑢𝑖 = the true value, and
𝑥 ∗ 𝑢𝑖 = the estimated value.

12/4/2024 Geostatistics 18
 Cross Validation
The following observations, regarding 𝑒 𝑢𝑖 , 𝑥 𝑢𝑖 and 𝑥 ∗ 𝑢𝑖 , must be examined as part
of the cross-validation process.
• Plot the true value vs. the estimated value
• Plot the estimation error, 𝑒 𝑢 , vs. the estimated value

12/4/2024 Geostatistics 19
 Cross Validation
• Plot the true value vs. the estimated value
Some examples of this plot are shown in Fig. 4.2.
In Fig. 4.2a, the spread of error is uniform around the 45° line. An
approximately equal number of values are under- and overestimated
over the entire region. This is an ideal situation.
In contrast, in Fig. 4.2b, the spread of error (deviation from the straight
line) increases as the true value increases. In other words, the error is
dependent on the magnitude of the sample. This type of behavior
reflects the heteroscedasticity of error variance.
In Fig. 4.2c, up to a certain range, the sample value is overestimated;
and over that range, the sample value is underestimated. This type of
estimate is called a conditionally biased estimate.

12/4/2024 Geostatistics 20
 Cross Validation
• Plot the estimation error, 𝑒 𝑢 , vs. the estimated value
Fig. 4.3a shows an example where the error is centered
around a zero line with equal
𝑛
spread. Ideally, for n samples,
1
෍ 𝑒 𝑢𝑖 ≈ 0
𝑛
𝑖=1
indicating that estimation is conditionally unbiased. The equal
spread around the zero line indicates the homoscedasticity of
error variance. That is, the error is independent of the
estimated value.
In contrast, in Fig. 4.3b, although the estimation is
conditionally unbiased, the magnitude of error increases as
the sample value increases. This is a reflection of
heteroscedastic behavior.

12/4/2024 Geostatistics 21
 Linear Kriging Procedures
▪ Simple Kriging
▪ Ordinary Kriging
▪ Cokriging

12/4/2024 Geostatistics 22
 Simple Kriging
Simple kriging starts with the assumption that a value at an unsampled location can be
estimated by 𝑛

𝑋 ∗ 𝑢𝑜 = 𝜆 𝑜 + ෍ 𝜆 𝑖 𝑋 𝑢𝑖
𝑖=1
where
𝑋 ∗ 𝑢𝑜 = the value to be estimated,
𝑋 𝑢𝑖 = a nearby sample value at location 𝑢𝑖 ,
𝜆𝑖 = the weight assigned to each sample, and
𝜆𝑜 = a constant

12/4/2024 Geostatistics 23
 Simple Kriging
To estimate the value of 𝜆𝑖 , we use the MVUE technique. An unbiased condition requires
that
𝐸 𝑋 ∗ 𝑢 𝑜 − 𝑋 𝑢𝑜 = 0
Substituting 𝑋 ∗ 𝑢𝑜 from Eq. 4.4, 𝑛
𝜆 𝑜 + ෍ 𝜆 𝑖 𝐸 𝑋 𝑢𝑖 = 𝐸 𝑋 𝑢𝑜
𝑖=1
If we assume 𝐸 𝑋 𝑢𝑖 = 𝐸 𝑋 𝑢𝑜 , based on the first order stationarity, we can write
𝑛

𝜆𝑜 = m 1 − ෍ 𝜆𝑖
𝑖=1

12/4/2024 Geostatistics 24
 Simple Kriging
In addition to the unbiased requirement, the condition of minimum variance must
also be satisfied. Mathematically, weights should be chosen so
𝜎 2 𝑋 𝑢𝑜 − 𝑋 ∗ 𝑢𝑜
is minimized.
The result of this condition
𝑛
is equation
෍ 𝜆𝑗 𝐶 𝑢𝑖 , 𝑢𝑗 = 𝐶 𝑢𝑖 , 𝑢𝑜 for i=1, …, n
𝑗=1
where
𝐶 𝑢𝑖 , 𝑢𝑗 = the covariance value between points located at 𝑢𝑖 and 𝑢𝑗 , respectively, and
𝐶 𝑢𝑖 , 𝑢𝑜 = the covariance between the sampled location, 𝑢𝑖 , and the unsampled
location, 𝑢𝑜 .
The covariance values are obtained based on the spatial model.

12/4/2024 Geostatistics 25
 Simple Kriging
In matrix form, Eq. 4.9 can be written as
𝐶 𝑢1 , 𝑢1 ⋯ 𝐶 𝑢1 , 𝑢𝑛 𝜆1 𝐶 𝑢1 , 𝑢𝑜
⋮ ⋱ ⋮ . = .
𝐶 𝑢𝑛 , 𝑢1 ⋯ 𝐶 𝑢𝑛 , 𝑢𝑛 𝜆𝑛 𝐶 𝑢𝑛 , 𝑢𝑜
or
𝐶 Λ = 𝑐
To calculate the value of Λ , we write
Λ = 𝐶 −1 𝑐
where
𝐶 −1 =the inverse of matrix 𝐶 .
Once the weights are estimated, Eq. 4.4 estimates the value of 𝑋 ∗ 𝑢𝑜 .

12/4/2024 Geostatistics 26
 Simple Kriging
In addition to the estimation, it also estimates the error variance associated with the
estimate. It is written as 𝑛

𝜎ො𝐸2 = 𝐶 𝑢𝑜 , 𝑢𝑜 − ෍ 𝜆𝑖 𝐶 𝑢𝑖 , 𝑢𝑜
𝑖=1

12/4/2024 Geostatistics 27
 Example 4.1
Consider the data configuration, 𝑥 𝑢1 = 10, 𝑥 𝑢2 = 20, and m = 18 (Fig. 4.5). Further,
assume that the isotopic variogram, representing the spatial relationship, is given as
−3𝐿
𝛾 𝐿 = 100 1 − exp for 𝐿 ≥ 0
100
Estimate the value at location 𝑢0 . Also, estimate the error variance.
If the spatial model is given as
−3𝐿
𝛾 𝐿 = 200 1 − exp for 𝐿 ≥ 0
100
how would the estimate and error variance change?

12/4/2024 Geostatistics 28
 Example 4.1 - Solution
Part 1. In this part, we assume the variogram is represented by
−3𝐿
𝛾 𝐿 = 100 1 − exp
100
In this example, we have two sample points, or n = 2. Therefore, the matrix equation can
be written as
𝐶 𝑢1 , 𝑢1 𝐶 𝑢1 , 𝑢2 𝜆1 𝐶 𝑢1 , 𝑢𝑜
=
𝐶 𝑢2 , 𝑢1 𝐶 𝑢2 , 𝑢2 𝜆2 𝐶 𝑢 2 , 𝑢𝑜
We know the relationship between the covariance and the variogram,
𝐶 𝐿 =𝐶 0 −𝛾 𝐿
Therefore, if we calculate 𝛾 𝐿 , we can calculate 𝐶 𝐿 .
The second-order stationarity assumption requires a variogram or covariance to be only
a function of distance and direction. Therefore, if we know the distance and the direction
(only distance for an isotropic model), we can calculate the associated variogram value.

12/4/2024 Geostatistics 29
 Example 4.1 - Solution
In the previous matrix, the only unknowns are 𝜆1 and 𝜆2 .
𝐶 𝑢1 , 𝑢1 = 𝐶 𝑢2 , 𝑢2 = 𝐶 0 = 100
because the distances between 𝑢1 and 𝑢1 , and 𝑢2 and 𝑢2 = 0. 𝐶 0 is the sill of the
variogram.
𝐶 𝑢1 , 𝑢2 = 𝐶 𝑢2 , 𝑢1 = 𝐶 0 − 𝛾 𝑢1 , 𝑢2 = 𝐶 0 − 𝛾 80
−3 × 80
= 100 − 100 1 − exp = 9.07
100

12/4/2024 Geostatistics 30
 Example 4.1 - Solution
Because the covariance and the variogram are symmetric, the variogram between two
given points is the same, irrespective of the directions.
−3 × 30
𝐶 𝑢1 , 𝑢0 = 𝐶 0 − 𝛾 𝑢1 , 𝑢0 = 100 − 𝛾 30 = 100 − 100 1 − exp = 40.66
100
and
−3 × 50
𝐶 𝑢 2 , 𝑢0 = 𝐶 0 − 𝛾 𝑢2 , 𝑢0 = 100 − 𝛾 50 = 100 − 100 1 − exp = 22.31
100
Note, the distance between 𝑢1 and 𝑢0 is 30 ft, and between 𝑢2 and 𝑢0 , it is 50 ft.

12/4/2024 Geostatistics 31
 Example 4.1 - Solution
Substituting all the values,
100 9.07 𝜆1 40.66
=
9.07 100 𝜆2 22.31
or
100𝜆1 + 9.07𝜆2 = 40.66
ቊ
9.07𝜆1 + 100𝜆2 = 22.31
Solving for 𝜆1 and 𝜆2 , we obtain 𝜆1 = 0.389, and 𝜆2 = 0.188.

12/4/2024 Geostatistics 32
 Example 4.1 - Solution
Using Eq. 4.7,
2

𝜆𝑜 = m 1 − ෍ 𝜆𝑖 = 18 1 − 0.389 + 0.188 = 7.614

𝑖=1

Using Eq. 4.4,

2

𝑥 ∗ 𝑢𝑜 = 𝜆𝑜 + ෍ 𝜆𝑖 𝑥 𝑢𝑖 = 𝜆𝑜 + 𝜆1 𝑥 𝑢1 + 𝜆2 𝑥 𝑢2
𝑖=1

= 7.614 + 0.389 × 10 + 0.188 × 20 = 15.264

This is the estimated value at the unsampled location.

12/4/2024 Geostatistics 33
 Example 4.1 - Solution
The error variance is calculated with Eq. 4.12.
2

𝜎ො𝐸2 = 𝐶 𝑢𝑜 , 𝑢𝑜 − ෍ 𝜆𝑖 𝐶 𝑢𝑖 , 𝑢𝑜 = 𝐶 𝑢𝑜 , 𝑢𝑜 − 𝜆1 𝐶 𝑢1 , 𝑢𝑜 + 𝜆2 𝐶 𝑢2 , 𝑢𝑜
𝑖=1
= 100 − 0.389 × 40.66 + 0.188 × 22.31 = 79.99

12/4/2024 Geostatistics 34
 Example 4.1 - Solution
Part 2. Here, we assume that the variogram is represented by
−3𝐿
𝛾 𝐿 = 200 1 − exp for 𝐿 ≥ 0
100
𝐶 𝑢1 , 𝑢1 = 𝐶 𝑢2 , 𝑢2 = 𝐶 0 = 200
𝐶 𝑢1 , 𝑢2 = 𝐶 𝑢2 , 𝑢1 = 𝐶 0 − 𝛾 𝑢1 , 𝑢2 = 𝐶 0 − 𝛾 80
−3 × 80
= 200 − 200 1 − exp = 18.14
100
−3 × 30
𝐶 𝑢1 , 𝑢0 = 𝐶 0 − 𝛾 𝑢1 , 𝑢0 = 200 − 𝛾 30 = 200 − 200 1 − exp = 81.31
100
and
−3 × 50
𝐶 𝑢 2 , 𝑢0 = 𝐶 0 − 𝛾 𝑢2 , 𝑢0 = 200 − 𝛾 50 = 200 − 200 1 − exp = 44.63
100

12/4/2024 Geostatistics 35
 Example 4.1 - Solution
Substituting in the matrix form,
200 18.14 𝜆1 81.31
=
18.14 200 𝜆2 44.63
Solving for 𝜆1 and 𝜆2 , 𝜆1 = 0.389, and 𝜆2 = 0.188.
Therefore,
𝜆𝑜 = 18 1 − 0.389 + 0.188 = 7.614
𝑥 ∗ 𝑢𝑜 = 7.614 + 0.389 × 10 + 0.188 × 20 = 15.264
The error variance can be estimated as
𝜎ො𝐸2 = 200 − 0.389 × 81.31 + 0.188 × 44.63 = 159.98

12/4/2024 Geostatistics 36
 Example 4.1 - Solution
By comparing answers from Parts 1 and 2, you will notice that the estimate is not
affected by changing the sill value, whereas the error variance doubled as the sill value is
doubled. In other words, the estimated value at the unsampled location is
independent of the sill value. This important characteristic is not restricted to just
simple kriging. It is common to all kriging procedures. This characteristic proves
useful in quantifying uncertainties with respect to kriging estimates.

12/4/2024 Geostatistics 37
 Simple Kriging
• Block Estimation

12/4/2024 Geostatistics 38
 Block Estimation
In the previous example, we considered the estimation of sample values at point
locations. This section extends the estimation procedure for block values, using simple
kriging technique. The difference between the point estimation and the block
estimation is the support of the location. Point estimation allows the estimation of the
value at unsampled locations without any significant volume or area association. In
contrast, block kriging allows the estimation of an unsampled location associated with
a certain area (in two dimensions) or volume (in three dimensions).

12/4/2024 Geostatistics 39
 Block Estimation
In estimating block values, we use an equation similar to point kriging. The equation is
written as 𝑛

𝑋𝑣∗ 𝑢𝑜 = 𝜆𝑜 + ෍ 𝜆𝑖 𝑋 𝑢𝑖
𝑖=1
where
𝑋𝑣∗ 𝑢𝑜 = the estimate of a block value located at 𝑢𝑜 , and
𝑋 𝑢𝑖 = the point values surrounding the unsampled block.

12/4/2024 Geostatistics 40
 Block Estimation
By applying the unbiased condition, we obtain 𝑛

𝜆𝑜 = m 1 − ෍ 𝜆𝑖
𝑖=1
This is the same equation we obtained for simple point kriging.
The second condition of minimum
𝑛
variance results in equation,
෍ 𝜆𝑗 𝐶 𝑢𝑖 , 𝑢𝑗 = 𝐶𝑣 𝑢𝑖 , 𝑢𝑜 for i=1, …, n
𝑗=1
This equation is very similar to Eq. 4.9, except for the right side of the equation. The
right side of the equation represents a covariance between a sampled point location at
and a block located at 𝑢𝑜 . 𝐶𝑣 𝑢𝑖 , 𝑢𝑜 = an arithmetic average of all the point-to-point
covariances between a sample point and all the points located within the block.

12/4/2024 Geostatistics 41
 Block Estimation
In practice, to obtain a representative
arithmetic average between a point
and a block, we select a certain
minimum number of points within a
block and calculate the point-to-
point covariance between the
selected points and the sample
location; then, we take the arithmetic
average. Schematically, this approach
is shown in Fig. 4.17. In this figure, the
covariance between a sampled
location and 16 points selected
within the block is calculated, and the
arithmetic average of these 16 values
represents the covariance between the
point and a block.

12/4/2024 Geostatistics 42
 Block Estimation
Mathematically, 𝑛𝑏
1
𝐶𝑣 𝑢𝑖 , 𝑢𝑜 ≈ ෍ 𝐶 𝑢 𝑖 , 𝑢 𝑜𝑗
𝑛𝑏
𝑗=1
where
𝑛𝑏 = the selected number of points
within a block, and
𝐶 𝑢𝑖 , 𝑢𝑜𝑗 = the covariance between
the point, 𝑢𝑜𝑗 within the block and the
sampled point, 𝑢𝑖 .

12/4/2024 Geostatistics 43
 Block Estimation
Once the covariances between sampled points and the unsampled block are calculated.
Eq. 4.15, in matrix form, is written as
𝐶 𝑢1 , 𝑢1 ⋯ 𝐶 𝑢1 , 𝑢𝑛 𝜆1 𝐶𝑣 𝑢1 , 𝑢𝑜
⋮ ⋱ ⋮ . = .
𝐶 𝑢𝑛 , 𝑢1 ⋯ 𝐶 𝑢𝑛 , 𝑢𝑛 𝜆𝑛 𝐶𝑣 𝑢𝑛 , 𝑢𝑜
Notice the similarity between Eqs. 4.16 and 4.10. The left sides are identical; the right
sides are different. For point kriging, the right side represents the covariance between
a sampled point and an unsampled location. For block kriging, the right side
represents the covariance between a sampled point and an unsampled block. By
solving Eq. 4.16, we obtain the values of 𝜆𝑖 which can be used to estimate the block
value.

12/4/2024 Geostatistics 44
 Block Estimation
In addition, error variance is calculated as 𝑛

𝜎ො𝐸2 = 𝐶𝑣𝑣 𝑢𝑜 , 𝑢𝑜 − ෍ 𝜆𝑖 𝐶𝑣 𝑢𝑖 , 𝑢𝑜
𝑖=1
where 𝐶𝑣𝑣 𝑢𝑜 , 𝑢𝑜 = block variance. The value of block variance is calculated by
obtaining point-to-point covariance between several representative pairs within the
block and then taking the arithmetic average of the covariances. Similar to point kriging,
we also obtain the error variance for the block estimation.

12/4/2024 Geostatistics 45
 Example 4.2
Estimate the porosity value and the corresponding error
variance for the block shown in Fig. 4.18. Assume that
the variogram is defined as
3
3 𝐿 1 𝐿
0.1 + 0.3 − for 𝐿 ≤ 800
𝛾 𝐿 = 2 800 2 800
0.4 for 𝐿 ≥ 800
The coordinates and the corresponding porosities at the
sampled points are given in Table 4.2.

12/4/2024 Geostatistics 46
 Example 4.2
What is the porosity value of the block located within
the larger block (shown by the dotted line)? Assume
four points, within the gridblock, are used to estimate
the covariance between a point and a block.

12/4/2024 Geostatistics 47
 Example 4.2 - Solution
Part 1. We estimate the block value with equation
3

𝑋𝑣∗ 𝑢𝑜 = 𝜆𝑜 + ෍ 𝜆𝑖 𝑋 𝑢𝑖
𝑖=1
where n = 3 in this particular case. Our goal is to calculate the values of 𝜆𝑖 . The matrix
equation is written as
𝐶 𝑢1 , 𝑢1 𝐶 𝑢1 , 𝑢2 𝐶 𝑢1 , 𝑢3 𝜆1 𝐶𝑣 𝑢1 , 𝑢𝑜
𝐶 𝑢2 , 𝑢1 𝐶 𝑢2 , 𝑢2 𝐶 𝑢2 , 𝑢3 𝜆2 = 𝐶𝑣 𝑢2 , 𝑢𝑜
𝐶 𝑢3 , 𝑢1 𝐶 𝑢3 , 𝑢2 𝐶 𝑢3 , 𝑢3 𝜆3 𝐶𝑣 𝑢3 , 𝑢𝑜

12/4/2024 Geostatistics 48
 Example 4.2 - Solution
𝛾 𝑢1 , 𝑢1 = 𝛾 𝑢2 , 𝑢2 = 𝛾 𝑢3 , 𝑢3 = 0
𝐶 𝑢1 , 𝑢1 = 𝐶 𝑢2 , 𝑢2 = 𝐶 𝑢3 , 𝑢3 = 0.4
The distance between any two points is calculated by knowing the coordinate locations.
For example,
𝐿1,2 = 𝑢2 − 𝑢1 2 + 𝑣2 − 𝑣1 2 = 600 − 250 2 + 400 − 200 2 = 403.1 ft
𝐶 𝑢1 , 𝑢2 = 𝐶 𝑢2 , 𝑢1 = 𝐶 0 − 𝛾 𝑢1 , 𝑢2 = 𝐶 0 − 𝛾 403.1
3
3 403.1 1 403.1
= 0.4 − 0.1 − 0.3 − = 0.0924
2 800 2 800

12/4/2024 Geostatistics 49
 Example 4.2 - Solution
Similarly,
𝐿1,3 = 𝑢3 − 𝑢1 2 + 𝑣3 − 𝑣1 2 = 400 − 250 2 + 900 − 200 2 = 715.9 ft
𝐶 𝑢1 , 𝑢3 = 𝐶 𝑢3 , 𝑢1 = 𝐶 0 − 𝛾 𝑢1 , 𝑢3 = 𝐶 0 − 𝛾 715.9
3
3 715.9 1 715.9
= 0.4 − 0.1 − 0.3 − = 0.0048
2 800 2 800

𝐿2,3 = 𝑢3 − 𝑢2 2 + 𝑣3 − 𝑣2 2 = 400 − 600 2 + 900 − 400 2 = 538.5 ft

𝐶 𝑢2 , 𝑢3 = 𝐶 𝑢3 , 𝑢2 = 𝐶 0 − 𝛾 𝑢2 , 𝑢3 = 𝐶 0 − 𝛾 538.5
3
3 538.5 1 538.5
= 0.4 − 0.1 − 0.3 − = 0.0428
2 800 2 800

12/4/2024 Geostatistics 50
 Example 4.2 - Solution
The right side of the matrix requires covariance
between a point and a block. We assume the average of
four point-to-point covariances will provide a
representative covariance between a point and a block.
If we select four points within the larger block, as
shown in Fig. 4.19, we can calculate the covariance
between the sample points and these four points, and
calculate the arithmetic average.

12/4/2024 Geostatistics 51
 Example 4.2 - Solution
𝐿1,01 = 𝑢1 − 𝑢01 2 + 𝑣1 − 𝑣01 2
= 250 − 250 2 + 200 − 550 2 = 350 ft
𝐶 𝑢1 , 𝑢01 = 𝐶 0 − 𝛾 𝑢1 , 𝑢01 = 𝐶 0 − 𝛾 350
3
3 350 1 350
= 0.4 − 0.1 − 0.3 − = 0.1157
2 800 2 800

𝐿1,02 = 𝑢1 − 𝑢02 2 + 𝑣1 − 𝑣02 2

= 250 − 350 2 + 200 − 550 2 = 364 ft
𝐶 𝑢1 , 𝑢02 = 𝐶 0 − 𝛾 𝑢1 , 𝑢02 = 𝐶 0 − 𝛾 364
3
3 364 1 364
= 0.4 − 0.1 − 0.3 − = 0.1094
2 800 2 800

12/4/2024 Geostatistics 52
 Example 4.2 - Solution
𝐿1,03 = 𝑢1 − 𝑢03 2 + 𝑣1 − 𝑣03 2
= 250 − 250 2 + 200 − 450 2 = 250 ft
𝐶 𝑢1 , 𝑢03 = 𝐶 0 − 𝛾 𝑢1 , 𝑢03 = 𝐶 0 − 𝛾 250
3
3 250 1 250
= 0.4 − 0.1 − 0.3 − = 0.1640
2 800 2 800

𝐿1,04 = 𝑢1 − 𝑢04 2 + 𝑣1 − 𝑣04 2

= 250 − 350 2 + 200 − 450 2 = 269.3 ft
𝐶 𝑢1 , 𝑢04 = 𝐶 0 − 𝛾 𝑢1 , 𝑢04 = 𝐶 0 − 𝛾 269.3
3
3 269.3 1 269.3
= 0.4 − 0.1 − 0.3 − = 0.1542
2 800 2 800

12/4/2024 Geostatistics 53
 Example 4.2 - Solution
We then calculate,
1
𝐶𝑣 𝑢1 , 𝑢𝑜 = 0.1157 + 0.1904 + 0.1640 + 0.1542 = 0.1358
4
Similarly,
1
𝐶𝑣 𝑢2 , 𝑢𝑜 = 0.1020 + 0.1433 + 0.1141 + 0.1614 = 0.1302
4

1
𝐶𝑣 𝑢3 , 𝑢𝑜 = 0.1020 + 0.1141 + 0.0645 + 0.0725 = 0.0883
4

12/4/2024 Geostatistics 54
 Example 4.2 - Solution
Substituting, we write,
0.4 0.0924 0.0048 𝜆1 0.1358
0.0924 0.4 0.0428 𝜆2 = 0.1302
0.0048 0.0428 0.4 𝜆3 0.0883
Solving for 𝜆𝑖 , 𝜆1 = 0.2818, 𝜆2 = 0.24, and 𝜆3 = 0.1917.
Therefore,
3

𝜆𝑜 = m 1 − ෍ 𝜆𝑖 = 0.28 1 − 0.2818 − 0.24 − 0.1917 = 0.0803

𝑖=1

with Mean Porosity = 28%

3

𝑥 ∗ 𝑢𝑜 = 𝜆𝑜 + ෍ 𝜆𝑖 𝑥 𝑢𝑖 = 0.0803 + 0.2818 × 0.13 + 0.24 × 0.25 + 0.1917 × 0.34

𝑖=1
= 0.242
with 𝑥 𝑢𝑖 are ϕ values of data points 1, 2, 3
12/4/2024 Geostatistics 55
 Example 4.2 - Solution
1
𝐶𝑣𝑣 𝑢0 , 𝑢0 = 𝐶 𝑢01 , 𝑢01 + 𝐶 𝑢01 , 𝑢02 + 𝐶 𝑢01 , 𝑢03 + 𝐶 𝑢01 , 𝑢04
4
Knowing the distances, the covariances are computed,
1
𝐶𝑣𝑣 𝑢0 , 𝑢0 = 0.4 + 0.244 + 0.244 + 0.2214 = 0.2774
4
The error variance is calculated as
3

𝜎ො𝐸2 = 𝐶𝑣𝑣 𝑢𝑜 , 𝑢𝑜 − ෍ 𝜆𝑖 𝐶𝑣 𝑢𝑖 , 𝑢𝑜
𝑖=1
= 0.2774 − 0.2818 × 0.1358 + 0.24 × 0.1302 + 0.1917 × 0.0883 = 0.191

12/4/2024 Geostatistics 56
 Ordinary Kriging
In a simple kriging procedure, we assume the mean value, 𝑚 𝑢 , is known. We must
know the value of 𝒎 before we use a simple kriging expression.
In practice, however, the true global mean is rarely known, unless we assume the
sample mean is the same as the global mean. Further, as seen before, the local mean
within the search neighborhood may vary over the region of interest. As a result, the
assumption of first-order stationarity may not be strictly valid.
An ordinary kriging procedure overcomes this problem by redefining the equation for
estimation.

12/4/2024 Geostatistics 57
 Ordinary Kriging
Consider Eq. 4.4 used for simple kriging. 𝑛

𝑋 ∗ 𝑢𝑜 = 𝜆 𝑜 + ෍ 𝜆 𝑖 𝑋 𝑢𝑖
𝑖=1
However, the unbiased condition requires
𝐸 𝑋 ∗ 𝑢 𝑜 − 𝑋 𝑢𝑜 = 0
If we assume
𝐸 𝑋 ∗ 𝑢𝑜 = 𝐸 𝑋 𝑢𝑖 = 𝑚 𝑢𝑜
where 𝑚 𝑢𝑜 represents the mean within the search neighborhood of location 𝑢𝑜 , we
write 𝑛

𝜆 𝑜 = 𝑚 𝑢𝑜 1 − ෍ 𝜆𝑖
𝑖=1

12/4/2024 Geostatistics 58
 Ordinary Kriging
However, we force 𝜆𝑜 to be zero, if we assume
𝑛

෍ 𝜆𝑖 = 1
𝑖=1
The estimate is written as 𝑛

𝑋 ∗ 𝑢𝑜 = ෍ 𝜆 𝑖 𝑋 𝑢𝑖
𝑖=1
with a constraint defined by Eq. 4.19. By forcing 𝝀𝒐 to be zero, the necessity for
requiring the mean value is eliminated. Further, we relax the assumption that first-
order stationarity be strictly satisfied, by assuming the local mean is dependent on the
location.

12/4/2024 Geostatistics 59
 Ordinary Kriging
In addition to the unbiased condition, we also are required to satisfy the minimum
variance condition. Minimizing
𝑛
the variance with a constraint in Eq. 4.19 results in
෍ 𝜆𝑗 𝐶 𝑢𝑖 , 𝑢𝑗 + 𝜇 = 𝐶 𝑢𝑖 , 𝑢𝑜 for i = 1, …, n
𝑗=1
where 𝜇 is called a Lagrange parameter, and 𝐶 represents the covariance.
In matrix form, Eq. 4.21 is written as
𝐶 𝑢1 , 𝑢1 ⋯ 𝐶 𝑢1 , 𝑢𝑛 1 𝜆1 𝐶 𝑢1 , 𝑢𝑜
⋮ ⋱ ⋮ . . .
=
𝐶 𝑢𝑛 , 𝑢1 ⋯ 𝐶 𝑢𝑛 , 𝑢𝑛 1 𝜆𝑛 𝐶 𝑢𝑛 , 𝑢𝑜
1 … 1 0 𝜇 1
To solve 𝜆𝑖 , we write, in matrix form,
Λ = 𝐶 −1 𝑐
Once 𝜆𝑖 is calculated, the estimate, 𝑋 ∗ 𝑢𝑜 , is obtained with Eq. 4.20.

12/4/2024 Geostatistics 60
 Ordinary Kriging
We also estimate the error variance as
𝑛

𝜎ො𝐸2 = 𝐶 𝑢𝑜 , 𝑢𝑜 − ෍ 𝜆𝑖 𝐶 𝑢𝑖 , 𝑢𝑜 − 𝜇
𝑖=1

12/4/2024 Geostatistics 61
 Example 4.3
Consider the configurations in Fig. 4.23:
𝑥 𝑢1 = 20, 𝑥 𝑢2 = 50, 𝑥 𝑢3 = 30, and 𝑥 𝑢4
= 100
Assume the following variogram models:
Isotropic Variogram
𝛾 𝐿 = 100𝑀𝑆500 𝐿
Estimate the value and the error variance at
location 𝑢𝑜 for both configurations.

12/4/2024 Geostatistics 62
 Example 4.3 - Solution
We assume that the variogram model is given by
3
3 𝐿 1 𝐿
100 − if 𝐿 ≤ 500
𝛾 𝐿 = 2 500 2 500
100 if 𝐿 ≥ 500
Configuration 1. Because there are four sample points, the covariance matrix is 5 x 5.
We know
𝐶 1, 1 = 𝐶 2, 2 = 𝐶 3, 3 = 𝐶 4, 4 = 𝐶 0 = 100
We calculate
𝐶 1, 2 = 𝐶 2, 1 = 𝐶 0 − 𝛾 1, 2 = 100 − 𝛾 200
3
3 200 1 200
= 100 − 100 − = 43.2
2 500 2 500
Similarly, we compute values of other terms in the covariance matrix.

12/4/2024 Geostatistics 63
 Example 4.3 - Solution
Configuration 1. The right side of the equation requires the covariance between the
unsampled location and the sample point. For example,
𝐶 1, 0 = 𝐶 0 − 𝛾 1, 0 = 100 − 𝛾 100
3
3 100 1 100
= 100 − 100 − = 70.4
2 500 2 500
In matrix form, we write

𝐶 1, 1 𝐶 1, 2 𝐶 1, 3 𝐶 1, 4 1 𝜆1 𝐶 1, 0
𝐶 2, 1 𝐶 2, 2 𝐶 2, 3 𝐶 2, 4 1 𝜆2 𝐶 2, 0
𝐶 3, 1 𝐶 3, 2 𝐶 3, 3 𝐶 3, 4 1 𝜆3 = 𝐶 3, 0
𝐶 4, 1 𝐶 4, 2 𝐶 4, 3 𝐶 4, 4 1 𝜆4 𝐶 4, 0
1 1 1 1 0 𝜇 1

12/4/2024 Geostatistics 64
 Example 4.3 - Solution
Configuration 1. By substituting all the values, we obtain

100 43.2 37.4 37.4 1 𝜆1 70.4

43.2 100 37.4 37.4 1 𝜆2 70.4
37.4 37.4 100 5.6 1 𝜆3 = 43.2
37.4 37.4 5.6 100 1 𝜆4 43.2
1 1 1 1 0 𝜇 1
Solving for 𝜆𝑖 and 𝜇, by inverting the matrix, we obtain
𝜆1 = 0.429, 𝜆2 = 0.429, 𝜆3 = 0.071, 𝜆4 = 0.071 and 𝜇 = 3.632
Once the weights are obtained, notice that the sum of all weights equals one. We
estimate the value at 𝑥 ∗ 𝑢𝑜 as
4

𝑥 ∗ 𝑢𝑜 = ෍ 𝜆𝑖 𝑥 𝑢𝑖 = 0.429 × 20 + 0.429 × 50 + 0.071 × 30 + 0.071 × 100 = 39.24

𝑖=1

12/4/2024 Geostatistics 65
 Example 4.3 - Solution
Configuration 1. The error variance is estimated
𝑛
as
𝜎ො𝐸2 = 𝐶 𝑢𝑜 , 𝑢𝑜 − ෍ 𝜆𝑖 𝐶 𝑢𝑖 , 𝑢𝑜 − 𝜇
𝑖=1
= 100 − 0.429 × 70.4 + 0.429 × 70.4 + 0.071 × 43.2 + 0.071 × 43.2 − 3.632 = 29.81

12/4/2024 Geostatistics 66
 Example 4.3 - Solution
Configuration 2. Following similar computations as before, we write the equation in the
matrix form as
100 43.2 70.4 20.8 1 𝜆1 70.4
43.2 100 20.8 70.4 1 𝜆2 70.4
70.4 20.8 100 5.6 1 𝜆3 = 43.2
20.8 70.4 5.6 100 1 𝜆4 43.2
1 1 1 1 0 𝜇 1
By solving these equations, we obtain
𝜆1 = 0.518, 𝜆2 = 0.518, 𝜆3 = −0.018, 𝜆4 = −0.018 and 𝜇 = −2.14
We estimate the value at 𝑥 ∗ 𝑢𝑜 as
4

𝑥 ∗ 𝑢𝑜 = ෍ 𝜆𝑖 𝑥 𝑢𝑖 = 0.518 × 20 + 0.518 × 50 − 0.018 × 30 − 0.018 × 100 = 33.92

𝑖=1

12/4/2024 Geostatistics 67
 Example 4.3 - Solution
Configuration 2. The error variance is estimated
𝑛
as
𝜎ො𝐸2 = 𝐶 𝑢𝑜 , 𝑢𝑜 − ෍ 𝜆𝑖 𝐶 𝑢𝑖 , 𝑢𝑜 − 𝜇
𝑖=1
= 100 − 0.518 × 70.4 + 0.518 × 70.4 − 0.018 × 43.2 − 0.018 × 43.2 + 2.14 = 30.8
The error variance for this configuration is higher, compared to the first configuration.
This is expected, considering Configuration 2 does not provide as much information
as the first configuration.
As Fig. 4.23 shows, in Configuration 1, the sample data are well distributed
suưounding the unsampled location. In contrast, in Configuration 2, all samples are
in one line. One would expect more uncertainty when estimating the value for
Configuration 2, compared to Configuration 1 because of less information. This is
reflected in the estimated error variance.

12/4/2024 Geostatistics 68