The UX of Automation by AJ King, Senior UX Researcher, Ocado
Deriving and applying direct and cross indicator variograms in SIS (2006)
1. Centre for Computational Geostatistics
School of Mining and Petroleum Engineering
Department of Civil & Environmental Engineering
University of Alberta
Deriving and Applying Direct and Cross
Indicator Variograms for SIS
David F. Machuca Mory and Clayton V. Deutsch
2. 1
Outline
• Introduction and motivation
• Theoretical Framework
• Deriving the indicator variograms
• Modelling the indicator variograms
• The adjacent cut-off’s alternative
• Conclusions
(c)2006 David F. Machuca-Mory
3. 2
Introduction (1/2)
• Indicator based techniques
exhibit unrealistic inter-class
transitions.
• The use of the full matrix of
indicator direct and cross-
variograms could help to
alleviate this problem.
• But, how do the indicator
cross variograms for
continuous variables behave?
• How do they relate to the
multiGaussian assumption?
(c)2006 David F. Machuca-Mory
4. 3
Introduction (2/2)
• Several stochastic simulation techniques for continuous variables are
based in the assumption of multiGaussianity:
– The univariate cumulative distribution functions (cdf) must be normal
– The N-point cdf of the normal score data must be N-normal distributed
too.
• In practice only bivariate Gaussianity is tested.
• The most common test consists of comparing the experimental
indicator direct variograms of the raw variable with the direct
indicator variograms derived from the biGaussian distribution.
• Currently this check is performed only for indicator direct variograms
(c)2006 David F. Machuca-Mory
5. 4
Outline
• Introduction and motivation
• Theoretical Framework
• Deriving the indicator variograms
• Modelling the indicator variograms
• The adjacent cut-off’s alternative
• Conclusions
(c)2006 David F. Machuca-Mory
6. 5
Theoretical Framework (1/3)
• Under the multiGaussianity assumption the biGaussian distribution of
the pairs Y(u) and Y(u+h) is determined by the correlation, ρ(h)=1-γ(h):
ρ(h)
γ(h)
Y(u)
Y(u+h)
(c)2006 David F. Machuca-Mory
7. 6
Theoretical Framework (2/3)
• The biGaussian CDF can be also defined by the correlation function of the
continuous variable:
Where and are the standard normal quantile threshold values
with probabilities p and p’, respectively.
• This is equivalent to the non-centered indicator cross-covariance, :
2 2
arcsin ( )
20
2 sin1
F( , , ( )) Prob (u) , (u h) . exp
2 2cos
Y h p p p p
p p Y p p
y y y y
y y h Y y Y y p p d
)(1
pGyp
1
( )py G p
(h; , )IK p p
),;h();hu();u()hu(,uProb ppIpppp yyKyIyIEyYy)Y(
),;h( ppI yyK ),;h( ppI yyK
(c)2006 David F. Machuca-Mory
8. 7
Theoretical Framework (3/3)
• The BiGaussian derived indicator cross variogram can be understood as
a combination of volumes under the biGaussian distribution surface:
2 (0; , ) ( ; , ) ( ; , )
2min( , ) ( ; , ) ( ; , )
( ; ) ( ; ) ( , ) ( ; )
2 ( ; , )
I p p I p p I p p
I p p I p p
p p p p
I p p
K y y K h y y K h y y
p p K h y y K h y y
E I u y I u h y I u y I u h y
h y y
{ ( ; ) ( ; )}p pE I u y I u h y { ( ; ) ( ; )}p pE I u h y I u y { ( ; ) ( ; )}p pE I u y I u y { ( ; ) ( ; )}p pE I u h y I u h y
(c)2006 David F. Machuca-Mory
9. 8
Outline
• Introduction and motivation
• Theoretical Framework
• Deriving the indicator variograms
• Modelling the indicator variograms
• The adjacent cut-off’s alternative
• Conclusions
(c)2006 David F. Machuca-Mory
10. 9
The biGauss-full program
yp yp’
( ) / ( ) ( ) ( )Y u h Y u h Y u
2 2
( ) / ( ) 1 ( )Y u h Y u h
• Draw a random number: 1 [0,1]p
1
1( ) ( )Y u G p
• Calculate:
• Define the conditional distribution:
N(μY(u+h)/Y(u), σ ² Y(u+h)/Y(u) )
1
( , ) 2( ) ( )Y u h G p
• Calculate:
2 [0,1]p • Draw a random number:
• Repeat several thousand times
Y(u)
• Calculate the proportion of
realizations that:
Which is equivalent to the indicator
cross variogram for the thresholds yp
and yp’
( ) and ( ) , and
( ) and ( )
p p
p p
Y u y Y u h y
Y u y Y u h y
Y(u+h)
yp
yp’
• Repeat the complete Monte
Carlo simulation for all lags
h.
• Repeat the whole process for
all cut-off’s combinations.
( )h
(c)2006 David F. Machuca-Mory
11. 10
Deriving hypothetical indicator
variograms (1/2)
• Gaussian derived indicator variograms from a spherical model of sill and range
equal 1, without nugget effect.
(c)2006 David F. Machuca-Mory
12. 11
Deriving hypothetical indicator
variograms (2/2)
• Gaussian derived indicator variograms and a spherical model of sill and range
equal 1 plus a nugget effect of 0.3.
(c)2006 David F. Machuca-Mory
13. 12
Deriving indicator variograms
from real data (1/2)
• Standardized Gaussian and experimental indicator cross and direct variograms.
(c)2006 David F. Machuca-Mory
14. 13
Deriving indicator variograms
from real data (2/2)
• Non-standardized Gaussian and experimental indicator cross and direct
variograms.
(c)2006 David F. Machuca-Mory
15. 14
The extreme continuity of
indicator cross variograms
• Reasonable if we consider indicator cross variograms as a measure of
inter-class transition.
• As difference between thresholds increase, less interclass transitions
are registered at short distances, and the indicator variogram becomes
more continuous.
• This extreme continuity is also present in the raw data indicator cross
variograms
(c)2006 David F. Machuca-Mory
16. 15
Outline
• Introduction and motivation
• Theoretical Framework
• Deriving the indicator variograms
• Modelling the indicator variograms
• The adjacent cut-off’s alternative
• Conclusions
(c)2006 David F. Machuca-Mory
17. 16
Fitting individually the indicator
variograms
• Individually most (but not all) of the variograms
can be fitted by a stable variogram model:
• But the complete matrix does not fulfill the
requirements of the LMC
P1=0.10 p2=0.10
γ(h)=1-exp(-3h^0.723)
0
0.2
0.4
0.6
0.8
1
1.2
-0.1 0.1 0.3 0.5 0.7 0.9 1.1 1.3 1.5
P1=0.50 p2=0.10
γ(h)=1-exp(-3h^1.875)
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4
P1=0.50 p2=0.50
γ(h)=1-exp(-3h^0.877)
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4
P1=0.90 p2=0.10
γ(h)=1-exp(-3h^3.03)
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4
P1=0.90 p2=0.50
γ(h)=1-exp(-3h^1.877)
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4
P1=0.90 p2=0.90
γ(h)=1-exp(-3h^0.723)
0
0.2
0.4
0.6
0.8
1
1.2
0 0.5 1 1.5
Valid model
Valid model Valid model
Valid model Valid modelNot a Valid model
( ) 1 exp 0, 0 2
h
h a
a
Missed continuity in the
regionalization model
fitting
(c)2006 David F. Machuca-Mory
18. 17
Fitting a LMC to the full matrix
of indicator variograms
Gaussian derived indicator variogram
LMC Model fitted
Missed continuity in the
LMC fitting
(c)2006 David F. Machuca-Mory
19. 18
Outline
• Introduction and motivation
• Theoretical Framework
• Deriving the indicator variograms
• Modelling the indicator variograms
• The adjacent cut-off’s alternative
• Conclusions
(c)2006 David F. Machuca-Mory
20. 19
The adjacent cut-off’s alternative (1/2)
• The idea is not to use the full coregionalization matrix for calculating the
conditional CDF values of each cut-off, but only the matrices defined by
the combination of the previous, the next and the same cut-off itself.
Cut-
off’s y1 y2 y3 y4 y5 y6 y7 y8 y9
y1
γ1,1 γ1,2 γ1,3
y2
γ2,1 γ2,2 γ2,3 γ2,4
y3
γ3,1 γ3,2 γ3,3 γ3,4 γ3,5
y4
γ4,2 γ4,3 γ4,4 γ4,5 γ4,6
y5
γ5,3 γ5,4 γ5,5 γ5,6 γ5,7
y6
γ6,4 γ6,5 γ6,6 γ6,7 γ6,8
y7
γ7,5 γ7,6 γ7,7 γ7,8 γ7,9
y8
γ8,6 γ8,7 γ8,8 γ8,9
y9
γ9,7 γ9,8 γ9,9
y1 y2 y3 y4 y5 y6 y7 y8 y9
Correct order relations!
(Proposed and implemented in cokriging by
Goovaerts, 1994)
(c)2006 David F. Machuca-Mory
21. 20
The adjacent cut-off’s alternative
(2/2)
• Thus only the cross variograms with the closest cut-off’s must be modeled,
those that can be fitted by a LMC.
• The adjacent cokriging equations becomes:
• And the adjacent cokriging estimator is:
0
0 0
0
1
, , 0 ,
1 1
0 0
( ; ) ( ; ) ( ; )
1 to , 1 to 1
p n
p p I p p I p p
p p
u y C u u y y C u u y y
n p p p
0
0 0 0
0
1
*
0 ,
1 1
( ; | ( )) ( ) ( ; ) ( ; ) ( )
p n
acoIK p p p p p p
p p
F u y n F y u y I u y F y
(c)2006 David F. Machuca-Mory
22. 21
Outline
• Introduction and motivation
• Theoretical Framework
• Deriving the indicator variograms
• Modelling the indicator variograms
• The adjacent cut-off’s alternative
• Conclusions
(c)2006 David F. Machuca-Mory
23. 22
Conclusions
• The full matrix of indicator direct and cross variograms can not be
fitted satisfactorily by a classic Linear Model of Coregionalization.
• This affirmation is valid for both Gaussian derived and experimental
indicator variograms.
• Further research is needed to develop an adequate model of
coregionalization in order to consistently use the indicator direct and
cross variograms in indicator cokriging and cosimulation.
• The adjacent cut-off’s approach for SIS could solve the problem of
uncontrolled class transitions only partially.
• This approach is being implemented and tested.
(c)2006 David F. Machuca-Mory