Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Principle of kriging
Types of kriging
Influence of variogram model parameters on kriging
results
1
Kriging
Principles
Assumption:
Stationarity
Mean does not change laterally “Behave everywhere in the same manner”
No trends
Inputs:
Variogram
Data points (normally well data)
CDF
Kriging is a estimation technique that can be used for creating either a 2D surface or a 3D
property model. Input to the algorithm are normally well data as point data and a variogram.
The variogram defines how to use the well data for populating values in the grid cells
between the wells.
Kriging is a linear equation system where the variogram values are known parameters, and
the kriging weights are unknown parameters.
The principal idea is that the changes are interpreted local fluctations around a constant
mean value.
Stationarity is simply an ASSUMPTION which is made regarding the rules for behavior of
the properties which we analyze, study, or model with geostatistical tools.
The rule is simply that the property must behave consistently within the volume chosen for
analysis, study, or modeling. If it does not, then the geostatistical tools which we use will not
work properly. Stationarity assumes that a property behaves the same way in all locations of
the chosen volume; i.e., that the samples have no inherent trend. If a trend exists, it must be
removed before using certain algorithms e.g. Sequential Gaussian Simulation.
2
Kriging
Principles
The basic idea is to combine known values, after applying a weighting
factor, so that an estimate can be made at an unknown location:
Weight Known value
n
z ( x0 ) = ∑ λi z ( xi )
i
Unknown value at position x0 is calculated from a linear combination of surrounding
data points.
Unknown values is estimated as a weighted sum of the known values
Kriging Weights are calculated from variogram model
If position x0 is outside the Range of the data points the value at x0 gets the Mean
value derived from the data (only approximately true for Gaussian and Exponential
variogram model type)
3
Kriging
Principles
Kriging finds the “best” Least Square estimate of the unknown values
4
Kriging
Principles
n
z ( x0 ) = ∑ λi z ( xi ), what is λi ?
i
5
Kriging
Principles
n
z * ( x ) = ∑ λi z ( xi ), linear estimator
i
Z(x): true but unknown value
X ? Direction of major continuity
6
Kriging
Principles
The estimate that minimizes the
error variance by construction
will be called kriging in
geostatistics.
Kriging uses the variogram to
understand the variability of the
data over a distance.
This knowledge allows kriging to
calculate weights that minimize
the error variance.
7
Kriging
From a Variogram Model to a Final Estimation
n
Variogram Æ ??? Æ z ( x 0 ) = ∑ λi z ( x i )
i
Linear Equation System to be solved:
Unknown Known
X2
⎡γ ( x11 ) γ ( x12 ) ... γ ( x1n ) 1 ⎤ ⎡ λ1 ⎤ ⎡γ ( x10 ) ⎤ (0.22)
⎢γ ( x ) γ ( x ) ... γ ( x 2 n )1 ⎥⎥ ⎢⎢λ 2 ⎥⎥ ⎢⎢γ ( x 20 ) ⎥⎥
⎢ 21 22 60
30
⎢ ... ... ... ... ...⎥ * ⎢ ... ⎥ = ⎢ ... ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ X0 (?)
⎢γ ( x n1 ) γ ( x n 2 ) ... γ ( x nn ) 1 ⎥ ⎢λ n ⎥ ⎢γ ( x n 0 )⎥ 42
40
⎢⎣ 1 1 1 1 0 ⎥⎦ ⎢⎣ μ ⎥⎦ ⎢⎣ 1 ⎥⎦ X1
(0.25)
20
55
X3
Variance (gamma) values have to be found and multiplied with the unkown weight (lambda) (0.30)
To figure out the property value at location x0, we need to solve the linear equation system.
The first matrix represents the variogram values for the distances between the neighborhood
data (blue in the image). The second represents the Kriging weights and the third are the
variogram values representing the distances from the neighborhood points to the point to be
estimated (red in the image).
8
Kriging
A Simple Example to Demistify Kriging
(from: Akin/Siemes: Praktische Geostatistik p. 120ff)
Gam m a
0.60
X0 (?)
42 0.40
40
0.20
X1 (0.25) 20 0.00
0 20 40 60 80 100
55
Distance
X3 (0.30)
c0 Nugget Effect
⎡ ⎛ 3 h ⎞ 1 ⎛ h ⎞3 ⎤ c1 Sill
Spherical Model γ ( h ) = c 0 + ⎢c1 ⎜⎜ * ⎟⎟ − * ⎜⎜ ⎟⎟ ⎥ h Distance
⎢ ⎝2 a ⎠ 2 ⎝ a ⎠ ⎥
⎣ ⎦ a Range
Kriging can be defined as an estimation technique based on the interaction between the
variogram parameter and the local neighborhood data. Starting from the variogram analysis
and the existing data Kriging combines the information in the following way:
The distance to the point to be estimated defines the data configuration. The (Ordinary
Kriging) example is derived from Akin & Siemes (Praktische Geostatistik). In this example a
variogram model has been set up and the model have a known variogram Range (see figure
above). The point x0 should be calculated using the information from the points x1, x2 and x3.
The distances from point x1, x2 and x3 to x0 are in red. The distances between the points x1,
x2 and x3 are in blue. The variogram was determined as shown in the figure. Every distance is
representing a variogram value (e.g. 40 m around 0.85). The Kriging equation matrix will be
set up as shown in the previous slide.
9
Kriging
And the Result is:
The parameter µ helps to solve the equation system but is not necessary for the final
calculation of the estimation value Z(x0):
**( equation only in the book)
Finally the value at the position x0 can be calculated:
10
Kriging
Simple Kriging
n n
z ( x0 ) = ∑ λi z ( xi ) + [1 − ∑ λi ]m
i i
11
Kriging
Ordinary Kriging
n
z ( x 0 ) = ∑ λi z ( xi )
i
Ordinary Kriging re-estimates the mean at each location, therefore it is a local mean.
It is more unstable than Simple Kriging, but can be suitable to use when there are many
wells. Can be suitable for data points that show a trend.
In Petrel there is an option to use Ordinary Kriging instead of Simple Kriging in the
Petrophysical Modeling Process.
12
Kriging
Exemplified – Variogram Range and Search Range
If there are no points as close to the node as the Variogram Range, then the algorithm will use the
selected alternate computational method, for example, a trend. This condition typically occurs in the
corners or at the edges of the map.
Map
Area
ch
Points inside this circle
ear e
Are weighted with the S ang
R Points outside of the
Variogram Variogram Search range are not
Range used for the node value
13
Kriging
Exemplified – Variogram Range
This example illustrates the condition where there are no points as close as the Variogram Range. In
this situation, the selected alternate computational method is used for the node. For example, if the
chosen algorithm was Simple Kriging, then the node in this example would be assigned a value equal to
the average value of all the points (Global Mean).
If the chosen algorithm had been Ordinary Kriging, then the node value would be some local average
of a subset of the total data.
Map
Area
ch
ar
Se ange
R
Variogram
Range
Data points
14
Kriging
Properties of Kriging
Note: The kriging algorithm is slow compared to other algorithms like the Convergent
Interpolation.
15
Kriging
Influence of Model Type - Range: 5000m
In general the Exponential and Spherical models gives a more similar result than the
gaussian model. The gaussian models are more smoothed than both the Exponential and
Spherical models.
16
Kriging
Influence of Range
Isotropic Isotropic
Range: 5.000m Range: 50.000m
Hint/recommendation – Range should always be larger than the average spacing of data
points.
17
Kriging
Influence of Azimuth
Good Bad
18
Kriging
Influence of Nugget
Note: The example to the right shows a nugget of 0.99, meaning that Kriging the value at any
points get the value of the global mean (simple kriging) or a local mean (ordinary kriging).
Using a 100% nugget model will create maps that look continuous, except at the data
locations where discontinuous jumps to the actual data values occur (at the well positions).
This is because kriging honors the input data points exactly. The continuous appearance is
due to the fact that any kriged point will get the global mean (simple kriging) of the data when
using a nugget of 0.99.
19
Comparison of Gridding Algorithms
20
Comparison of Gridding Algorithms
Moving Average Interpolation Kriging (Azimuth –40)
+ Anisotropy handling
+ Suitable for any grid or data size
- Limited control over “bull’s eyes”
- Bad extrapolator
- No fault handling
21
Kriging Option in Petrel Make/Edit Surface:
“Kriging with external Drift/Trend”
Note: This trend option is available for most of the other gridding
algorithms.
22
“Kriging with external Drift”
Example
23
Kriging vs. Non-kriging algorithms
24
Exercise
25