Downloaded from https://iranpaper.

ir
https://www.tarjomano.com https://www.tarjomano.com

Earth Sci Inform (2016) 9:235–245

DOI 10.1007/s12145-015-0240-4

SOFTWARE ARTICLE

FuzzyKrig: a comprehensive matlab toolbox for geostatistical

estimation of imprecise information
Saeed Soltani-Mohammadi 1

Received: 15 July 2014 / Accepted: 1 September 2015 / Published online: 15 October 2015
# Springer-Verlag Berlin Heidelberg 2015

Abstract Using kriging has been accepted today as the most Introduction
common method of estimating spatial data in such different
fields as the geosciences. To be able to apply kriging methods, A very basic step in mine planning, the results of which will have
it is necessary that the data and variogram model parameters considerable effects on other planning phases, is using estimators
be precise. To utilize the imprecise (fuzzy) data and parame- for the accurate prediction of the ore grade (and other key prop-
ters, use is made of fuzzy kriging methods. Although it has erties of a deposit) based on a limited number of core samples.
been 30 years since different fuzzy kriging algorithms were Estimators can be generally divided into 2 groups: 1) linear (k-
proposed, its use has not become as common as other kriging nearest neighbor, inverse distance to a power, feed forward neu-
methods (ordinary, simple, log, universal, etc.); lack of a com- ral network, linear transfer function, etc.), and 2) nonlinear (feed
prehensive software that can perform, based on different fuzzy forward neural network with nonlinear transfer function, neuro-
kriging algorithms, the related calculations in a 3D space can fuzzy-frectal approach, nonlinear static kriging, etc.) (Nelles
be the main reason. This paper describes an open-source soft- 2001). Since kriging is the best unbiased linear estimator that
ware toolbox (developed in Matlab) for running different al- can guarantee a minimum estimation variance (Cressie 1993),
gorithms proposed for fuzzy kriging. It also presents, besides a it is known as the most common estimator for the evaluation of
short presentation of the fuzzy kriging method and introduc- mineral deposits. To use it, it is required that the input data be
tion of the functions provided by the FuzzyKrig toolbox, 3 assumed precise (not tainted with uncertainty), and structural
cases of the software application under the conditions where: analysis and variogram model fitting be possible. But, the uncer-
1) data are hard and variogram model parameters are fuzzy, 2) tainties tainted with assaying, on the one hand, and fitting
data are fuzzy and variogram model parameters are hard, and variogram model tainted with uncertainty (due to insufficient
3) both data and variogram model parameters are fuzzy. data during structural analysis), on the other hand, make the
kriging results unreliable. Geostatisticians have dealt, in 3 differ-
ent ways, with these uncertainties: 1) neglecting them, 2) defin-
Keywords Fuzzy variogram model . Geostatistics . Kriging . ing a unique prior distribution function for every parameter with
Uncertainty uncertainty and using Bayesian kriging method (Handcock and
Stein 1993), and 3) defining the imprecise parameters in the form
of fuzzy interval parameters and using fuzzy kriging methods
(Diamond 1989); in most geostatistical studies, the first method
is used because the epistemic uncertainty of the variogram model
Communicated by: H. A. Babaie parameters and data are insufficient. In the Bayesian kriging
method, presence of epistemic uncertainty is assessed through
* Saeed Soltani-Mohammadi relating the primary subjective probabilities to every possible
[email protected] model. Despite all the merits, it has 2 general objections: 1) it
presents much more subjective probability information than
1
Department of Mining Engineering, University of Kashan, what really exists and 2) since it presents the subjective and
Kashan, Iran objective probability information related to 2 very different
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

236 Earth Sci Inform (2016) 9:235–245

natures, their product (like what happens in the Bayes rule) is parameters determining the variogram model. The ordinary
inconsistent (Loquin and Dubois 2012). Fuzzy kriging can be kriging predictor Z (sp) is displayed as follows:
done based on 2 different algorithms: 1) extending the principles


of random functions to triangular fuzzy random functions Z * sp ¼ λT ðθ; S ÞZ ¼ f ðθ; S; Z Þ; p ¼ 1; …; n ð1Þ
(Diamond 1989), and 2) applying the extension principles to
some appropriate selected operators (Bardossy et al. 1988, where the weights λT(θ,S)={λ1,…,λn} satisfy
1990a, b; Piotrowski et al. 1996). Despite some applications of  T
1−1T Γ−1 γ
the first approach, (Loquin and Dubois 2010) studied fuzzy λ ¼ γ þ 1 T −1
T
Γ−1 ð2Þ
kriging methods and concluded that the second approach was 1 Γ 1
more comprehensive because the first had some ambiguities
where 1T =(1,…,1)∈Rn, γT =(γ(s1,sp),…,γ(sn,sp)) and Γ={γ(si,
(Loquin and Dubois 2010). Although the number of drillholes
sj)}ni,j = 1 (Cressie 1993; Schelin and Luna 2010). The correspond-
in many mineral deposits is low and, as a result, the fitted
ing kriging variance is given as follows:
variogram model is tainted with uncertainty, throughout the years
since fuzzy kriging method was first proposed, only few cases of

2 .
T −1 
σ2k sp ¼ γ T Γ−1 γ− 1T Γ−1 γ−1 1 Γ 1 ¼ gðθ; S Þ ð3Þ
its application have been reported in mineral deposit evaluations
(evaluation of granite deposits in Spain based on the first approach
and γ are calculated based on the γθ variogram model fitted to
(Taboada et al. 2008) and estimation of the glacial aquitard thick-
the experimental variogram. The variogram model type (linear,
ness in northwestern Germany based on the second approach
gaussian, spherical, cubic and exponential) and its parameters
(Piotrowski et al. 1996)). The main applications have been in such
(nugget effect, sill and range) have important and considerable
cases as the evaluation of underground aquifers and environmen-
effects on the results of kriging (Bardossy et al. 1988; Diamond
tal issues. Lack of the appropriate software for the related calcu-
and Armstrong 1984). Although various methods have been
lations is, perhaps, the main reason why the application of this
proposed regarding the fitting of the variogram model (Webster
method has not become popular yet; the only reported software
and Oliver 2007), fitting a robust model to an experimental
that can perform fuzzy kriging calculations is FUZZEKS (Bartels
variogram and determining the precise values of the model pa-
1997). This software has been codified (developed) based on the
rameters are not easy tasks due to insufficient data or their asym-
second approach and is able to do the required calculations related
metric distribution in the region. Another limitation set forth in
to the data gathered in a 2D space while the software needed for
ordinary, simple and universal kriging methods is that they can
mining issues should have the capability of being used in calcu-
be applied only to hard data (Bardossy et al. 1988). Bardossy
lations related to the data gathered in a 3D environment.
studied these 2 problems in 1988 and 1990 and proposed a
FuzzyKrig, the software proposed in this paper, tries, based on
solution for each based on the fuzzy theory.
all the existing algorithms, to do fuzzy estimation calculations ~ and their components Z~ ðsi Þ
Considering fuzzy data Z~ and θ
related to the data gathered in a 3D space and it can meet the
~
and θ j respectively (i=1,. . , n and j=1, . . ,p) (here the fuzzy
requirements of the mining sector. It is hoped that we may, in near
future, witness progress in the use of fuzzy kriging method in data are regarded as fuzzy numbers), we can obtain the fuzzy
*


geosciences with the help of this useful software. The rest of the kriging estimate Z~ sp and variance σ ~ 2k sp according to
*


paper has been organized as follows: a brief theoretical back- Eqs. 1 & 3 and the extension principle. Z~ sp and σ ~ 2k sp
ground of the subject matter, and calculations related to the fuzzy
membership functions are then as follows:
kriging method, in section 2; a review of the software generalities
including functions, inputs and outputs, and specific observations    n  k !
*

considered in its codification, in section 3; presentation of 3 case μ Z^ sp ¼ sup min μ ðZðsi ÞÞ ; μ ðθi Þ
^ ðsi Þ ^θi
Z;θ:Z * ðsp Þ¼ f ðθ;S;ZÞ Z i¼1 i¼1
studies in section 4; and conclusions in section 5.
    ð4Þ
2

^ k sp
μ σ ¼ sup min μ ðθi Þ ð5Þ
Brief theoretical background j¼1;…;k ~θi
θ:σ2k ðsp Þ¼gðθ;SÞ

*


Problem statement Z~ s
p and σ
 ~ 2k s
p are

fuzzy
 numbers and their α -level sets
Z~ sp
*
and σ ~ 2 sp are then given by:
α k
Kriging is one of the most popular methods used for ore reserve   α n o
*

Z~ sp ~ j ; f or i ¼ 1; …; n; j ¼ 1; …; p
¼ f ðθ; S; ZÞ : Zðsi Þ∈Z~ ðsi Þ; θ j ∈θ
estimations. It allows incorporation of spatial correlation into the α
estimation process. Let ZT ={Z(s1),…,Z(sn)} denote the obser- ð6Þ
vations from a second order stationary process at locations   n o
2

S={s1,…,sn} with the variogram function 2γθ(h)=var[Z(s+h) ~ k sp
σ ~ j ; f or j ¼ 1; …; p
¼ g ðθ; SÞ : θ j ∈θ ð7Þ
−Z(s)] where handθ denote the separation distance and vector of α
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

Earth Sci Inform (2016) 9:235–245 237

  "    #
and, since they are fuzzy numbers, their α -level sets are the 2
 2
 L 2
 U
following closed intervals: σ~ k sp ¼ σ~ k sp ~ k sp
; σ ð9Þ
α α α
  "    #  *
 L
*
 *
 L *
 U ~ sp
Z~ sp ¼ Z~ sp ; Z~ sp ð8Þ From Eq. (6), the left-endpoint Z
α
and the right-
α α α  *
 U
~ sp
endpoint Z can be displayed as:
α

 
*
 L
Z~ sp ¼ min  f ðθ; S; ZÞ ð10Þ
α
 L  U  L  U
~ ðsi Þ
ðZ;θÞ: Z ~ ðsi Þ
< Z ðsi Þ< Z ~j
; θ ~j
< θ j< 0 θ ; f or i¼1;…;n; j¼1;…:;p
α α α α

 
*
 U
Z~ sp ¼ max  f ðθ; S; ZÞ ð11Þ
α
 L  U  L  U
ðZ;θÞ: Z~ ðsi Þ < Z ðsi Þ< Z~ ðsi Þ ~j
; θ ~j
< θ j< 0 θ ; f or i¼1;…;n; j¼1;…:;p
α α α α

2
L
~ k sp α and the right-
and from
Eq.
7,the left-endpoint σ minn f ðxÞ
U x∈R
~ k sp α can be shown as:
endpoint σ 2

subject to : cðxÞ ¼ 0
 
2
 L
σ~ k sp ¼ min  g ðθ; SÞ ð12Þ where f: Rn →R and c are objective and constraint functions.
α
 L  U
~j
θ: θ ~j
< θ j< 0 θ ; f or j¼1;…:;p
The SQP algorithm generates a search direction, d, at any
α α iteration as follows:
 
2
 U
σ~ k sp ¼ max gðθ; SÞ ð13Þ Minimize 8 ∇ f Tk d þ 0:5 d T Bk d
θ:ðθ
~ j Þ < θ j < 0 ðθ
~ j Þ ; f or j¼1;…:;p
L U
α
α α
< c1t ðλk Þ þ ∇cT1t ðλk Þd ¼ 0; t ¼ 1; …; N
subject to : −c2t ðλk Þ−∇cT2t ðλk Þd ≤ 0; t ¼ 1; …; m
:
Relations 11–13 are constrained nonlinear optimization c3t ðλk Þ þ ∇cT3t ðλk Þd ≤ 0; t ¼ 1; …; N −1
problems, the solutions of which are not possible with the La- ð14Þ
grangian method; to solve them, use has to be made of such
methods as sequential linear programming, sequential quadratic where k is the iteration number, d is the search direction and Bk
programming, feasible directions, genetic algorithm, and so on. is the positive definite approximation to the Hessian matrix of
We made use of the sequential quadratic programming (SQP) the Lagrangian function as follows:
because it is perfect for moderate-size, constrained nonlinear
problems (Hock and Schittkowski 1983; Fletcher 1987). To
X
N
X
m
simplify the solutions of problems 10–13, the authors studied Lðλ; μÞ ¼ f ðλÞ þ μi c1i ðλÞ− μiþN c2i ðλÞ
3 different cases: 1) the measured values were tainted with i¼1 i¼1
epistemic uncertainty, but the variogram model parameters N −1
were precise (Bardossy et al. 1988), 2) the measured values X
þ μiþN þm c3i ðλÞ ð15Þ
were precise, but the model parameters were tainted with epi- i¼1
stemic uncertainty (Bardossy et al. 1990a, b), and 3) both the
measured values and model parameters were tainted with un-
Convexity of this problem is guaranteed by approximating
certainty (Loquin and Dubois 2012).
B using Broyden–Fletcher–Goldfarb–Shanno (BFGS)
updating scheme as follows:
Solution method T
^y ^y Bk ssT Bk
Bkþ1 ¼ Bk þ − ð16Þ
Sequential quadratic programming (SQP) is based on solving S T ^y sT B k s
a series of sub-problems designed to minimize nonlinearly
constrained problem where
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

238 Earth Sci Inform (2016) 9:235–245

Table 1 Functions related to program implementation

Function Description and note

fuzzy_kriging_bardossy88 Do fuzzy kriging calculations based on Bardossy et al., 1990a algorithm

fuzzy_kriging_bardossy90 Do fuzzy kriging calculations based on Bardossy et al., 1988 algorithm
fuzzy_kriging_Loquin Do fuzzy kriging calculations based on Loquin and Dubois 2012 algorithm
gama3d_com_block Calculate the semivariance between the data points and the target (grid point or block)
gama3dcom_com Calculate the semivariance between the data points.
grade_function_button Create objective function to be used for the lower extreme of the grade at different α levels.
grade_function_up Create objective function to be used for the upper extreme of the grade at different α levels.
variance_function_button Create objective function to be used for the lower extreme of the varianceat different α levels.
variance_function_up Create objective function to be used for the upper extreme of the variance at different α levels.
Int_weith_calculation Calculate interval width of the fuzzy number.



s ¼ λkþ1 −λkþ1 ; y ¼ ∇λ L λkþ1 ; μkþ1 −∇λ L λk ; μkþ1 (Soltani-Mohammadi and Tercan 2012):
^y ¼ l y 8
þ ð1−l Þ Bk s
<1 ; if sT y > 0:2sT Bk s 1. Set up the initial solution λ0 and the Hessian approxima-
l ¼ T
0:8s Bk s tion B0 and let k=0.
: T ; otherwise 2. Solve Eq. (9) and determine d.
s Bk s−sT y
ð17Þ 3. Calculate λk+1 =λk +αd as the new potential solution (α
being the step length).
With ŷ obtained, the damping factor l is used to guarantee that 4. Update Hessian approximation B by the BFGS algorithm.
Bk+1 is sufficiently positive definite (Powell 1978). The com- 5. Stop the calculation when convergence condition is satis-
putational aspects of the SQP procedure are as follows fied, otherwise write k=k+1 and return to step 2.

Fig. 1 Geology map of the copper deposit (Ghaderi et al. 2007)

Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

Earth Sci Inform (2016) 9:235–245 239

Cross validation Table 2 Statistical

parameters of Cu grade Total number of records 127
Maximum 0.8552
Usually, to verify the kriging performance, use is made of the
Minimum 0.06
cross validation method for the optimal case of which two
Mean 0.2558
independent sets of data are required; kriging is done with
Variance 3.77E-02
one set, and validation with the other. But, since data are
Standard deviation 0.194
insufficient, use is usually made of the leave-one-out cross
Mean of logs −1.6004
validation method wherein the variable value is first estimated
Logarithmic variance 0.4522
at known (but temporarily eliminated) positions and, then, the
Log estimate of mean 0.253
kriging quality is verified through a comparison between the
estimated and real (measured) values; this method is common
for the verification of the quality of the universal, simple and Design, development and description
ordinary kriging (Davis 1987; Falivene et al. 2010). But, since
the output of the fuzzy kriging method (contrary to the com- Input parameters are as follows:
mon geostatistical methods) is a fuzzy value characterized by
its mean and width, it is necessary that attention be paid to 1) Method parameter: 3 different choices have been consid-
both when comparing the estimated and real values (similar to ered in the software:
the validation used for the Bayesian kriging (Aelion et al.
2009)). When data are crisp, but the variogram model is fuzzy, a) Epistemic uncertainty in the data (Bardossy et al.
the real values will be compared with the estimated fuzzy 1988).
mean when comparing the mean values. But, if data are fuzzy, b) Epistemic uncertainty in the variogram (Bardossy et
the comparison is made between the measured values and the al. 1990a), and
estimated fuzzy mean values whether the variogram model is c) Epistemic uncertainty in both the data and variogram
crisp or fuzzy. The ideal case is when the correlation coeffi- (Loquin and Dubois 2012).
cient and the slope of the regression line are both equal to 1. If 2) Sample data: it is a matrix, the rows of which are equal to
data are crisp, a good estimator is one that based on the width the number of samples and the columns are 6 if measure-
criterion, its measured values lie in the range of the fuzzy ments are tainted with uncertainty; otherwise, they are 4.
kriged values. But, when data are fuzzy, a good estimator’s The first 3 columns show respectively the longitude, the
fuzzy values have the highest in common with the measured latitude, and the height of the sampling point. If ap-
ones. proaches 1 and 3 are used, the next 3 columns will respec-
tively show the lower, the mode, and the upper extremes
of the samples’ fuzzy numbers and, if data are crisp, col-
Software umn 4 will be the measured value.

FuzzyKrig program was developed with such goals as

0.05
ease of use, time efficiency, accuracy, and flexibility for
every step of the process of data analysis. These were 0.045
accomplished by running the program using windows- 0.04
based graphic user interfaces (GUIs) to make navigation
of the system easy and user-friendly. All programming 0.035

was done using Matlab® (version: 2010a), which is a 0.03

matrix-based mathematical package capable of handling
Gama

0.025
large amounts of data and carrying out time-efficient
handling of mathematical manipulations. We chose to 0.02
develop FuzzyKrig in Matlab for a number of reasons: 0.015
Experimental
adoption by many geostatisticians, portability, and ease Upper bound
of generating graphical outputs. One of its major capa- 0.01 Mode bound
bilities is that it can integrate powerful tools for linear 0.005 Lower bound
algebra and optimization algorithms. FuzzyKrig can be
0
run within the Matlab® programming environment (the 0 50 100 150 200 250 300 350
software requires purchasing) allowing users to modify Lag
or add additional routines to the main program tailored Fig. 2 The upper bound, mode, and lower bounds of a fuzzy spherical
to their study-specific needs. variogram model for the experimental variogram for Cu grade
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

240 Earth Sci Inform (2016) 9:235–245

a b c
10,050 10,050
10,050

10,000 10,000
10,000 0.5

9,950 9,950
9,950
0.4

9,900 9,900 9,900

0.3

9,850 9,850 9,850

0.2
9,800 9,800 9,800

9,750 0.1
9,750 9,750

9,700 9,700
9,950 10,000 10,050 10,100 9,700 0 0
9,950 10,000 10,050 10,100 9,950 10,000 10,050 10,100

Fig. 3 Estimated lower and upper extremes, and fuzzy interval for Cu grade at the height level of 1900 m (using fuzzy kriging method)

3) L-variogram model: the lower extremes of the triangular 4) M-variogram model: the mode extremes of the triangular
fuzzy numbers are for the fitted variogram models which fuzzy numbers are for the fitted variogram model.
enter in the form of a vector that has the nugget effect in 5) U-variogram model: the upper extremes of the triangular
its first column, sill in the second, and range in the third. fuzzy numbers are for the fitted variogram model.

Fig. 4 Lower and upper a b

extremes of estimation variance at 10,050 10,050 0.035
the height level of 1900 m

10,000 10,000 0.03

9,950 9,950 0.025

9,900 9,900 0.02

9,850 9,850 0.015

9,800 9,800 0.01

9,750 9,750 0.005

9,700 9,700 0
9,950 10,000 10,050 10,100 9,950 10,000 10,050 10,100
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

Earth Sci Inform (2016) 9:235–245 241

0.2
2) Program implementation functions perform calculations
related to fuzzy kriging. They are given in Table 1 togeth-
er with their operations and the bases for their
0.15 calculations.
Fuzzy mean

3) “Fmincon” function in Matlab optimization toolbox is

used based on the SQP method, to find the optimum
0.1

y = 0.9854x values in optimization problems.

R = 0.7 4) Plot function plots (at different levels of the deposit): a)
0.05

variations of the estimated grade at different α levels, b)

estimation variance at different α levels, and c) interval
width of the fuzzy number.
0

0 0.05 0.1 0.15 0.2

Real The software presents 2 different plots which are related to
the lower and upper extremes of the fuzzy number. It is worth
Fig. 5 Comparison of the measured copper grade and the fuzzy mean of
the estimated grades mentioning that use has been made of the surf command in the
package to prepare these plots and if the user wishes to use
contour maps, it can be easily done through small modifica-
6) Search ellipsoid parameters: Search Radius and Max. No. tions in the program codes.
are respectively the radius and maximum number of sam-
ples in the search ellipsoid. Program availability and system requirements
7) Grid: it includes grid specifications (blocks’ coordinates
and dimensions) for which estimation is carried out; grid “FuzzyKrig” code has been presented in (http://saeedsoltani.
can be generated in such software environments as kashanu.ac.ir/) and it needs Matlab’s basic modules in (http://
SGeMS or in Matlab itself. www.mathworks.com/products/matlab/) plus its optimization
toolbox to implement it. Program implementation has been
Output parameters are also as follows: tested in the 2010 and 2013 versions of Matlab and it is
believed that it will be implemented equally well in the
1) The estimated value: it is a triangular fuzzy number. future versions too. Users are encouraged to report any
2) Estimation variance: it is a triangular fuzzy number. potentially useful changes or enhancements they may make,
3) Interval width of the estimated fuzzy number. along with reports of difficulties, bugs, and requests for new
features.
Functions used in this software are of four types:

1) Functions related to GUI components (pushbuttons, radio FuzzyKrig applications to 3 case-studies

buttons, edit-box) provide a way to perform a specific
operation when the user changes a component. They are FuzzyKrig can be used in such different cases as mineral de-
known as “callback”. posit evaluation, geotechnics, water resource management,

Fig. 6 Comparison of the 0.2

position of the measured copper Estimated fuzzy value
0.19
grade in the samples with respect × Measured value
to the lower and upper extremes 0.18
of the fuzzy kriged values 0.17
Grade (%)

0.16
0.15
0.14
0.13
0.12
0.11
0.1
s10
s11
s12
s13
s14
s15
s16
s17
s18
s19
s20
s21
s22
s23
s24
s25
s26
s27
s28
s29
s30
s31
s32
s2
s3
s4
s5
s6
s7
s8
s9

Sample Name
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

242 Earth Sci Inform (2016) 9:235–245

Fig. 7 Geological conditions in Azad pumped storage power plant (Aalianvari et al. 2010)

etc. to estimate the spatial imprecise (fuzzy) information. In block dimensions along x, y, and z axes for 1592 blocks in the
this section, 3 brief case studies, from among various applica- block model in columns 4–6. Statistical parameters of Cu grade
tions of FuzzyKrig, have been presented. They have been are given in Table 2. Due to insufficient data, structural stud-
selected to show the application of the program in situations ies were limited to only investigating the directionless
where: 1) data are precise, but variogram model parameters variogram. The experimental directionless variogram and the
are tainted with uncertainties, 2) parameters are precise, but fitted spherical fuzzy variogram model are shown in Fig. 2. The
data are tainted with uncertainties, and 3) both are tainted with parameters of the fitted fuzzy variogram model are ~θ1 = (0.009,
uncertainties. The input data related to these 3 case-studies 0.012, 0.0155), related to the nugget effect, ~θ2 = (0.017, 0.021,
have been distributed together with the program itself.
0.0245), related to the sill, and ~θ3 = (68, 84, 100), related to the
range. Figure 3 shows the plots of the lower and upper extremes
Case I: hard data and fuzzy variogram and the fuzzy interval of the estimated Cu grade, and Fig. 4
shows the lower and upper extremes of the estimation variance
An example has been used from among a data collection of Cu at the 1900 m level. To verify the kriging quality, use was made
grades in one of the porphyry copper deposits located in the
Sanandaj-Sirjan belt (Fig. 1) to show how the FuzzyKrig 50
package is used to estimate the hard data used to fit the fuzzy 45
variogram. This porphyry copper deposit is hosted in Oligo- 40
Miocene stocks that intruded Eocene volcano-sedimentary
35
rocks. The stocks are ranging in composition from monzonite
30
through quartz monzonite. The mineralization in this deposit
Gama

occurs in three mineralization zones include hypogene, oxida- 25

tion and supergene. From among 12 drillholes, only 5 crossed 20

the deposit and we based our modeling and evaluation studies 15

on them. The data obtained from the drillholes and from the 10
Experimental
variogram modeling studies are presented in a file titled 5 Mode bound
(input_Iju.mat). It contains: 1) geographic coordinates in the first
0
3 columns of the matrix, and assaying obtained from 360 0 50 100 150 200

drillhole samples s in column 4, and 2) geographic coordinates Lag

of the block center in the first 3 columns of the grid matrix, and Fig. 8 Experimental and model variogram of fracture length
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

Earth Sci Inform (2016) 9:235–245 243

3,913,600 50
0.9 45

40
3,913,400 0.8
35

30
0.7

Gama
25
3,913,200
20
0.6
15
Experimental
0.5 10 Upper bound
3,913,000 Mode bound
5
Lower bound
0.4 0
0 50 100 150 200
Lag
3,912,800
0.3
Fig. 11 The upper bound, mode, and lower bounds of a fuzzy spherical
variogram model for the experimental variogram in Fig. 5
0.2
3,912,600
Case II: fuzzy data and precise variogram
0.1

We have used fuzzy data gathered from fractures in the site of

3,912,400 0
642,500 642,700 642,900 643,100 a pumped storage power plant being constructed in the west-
Fig. 9 Fuzzy intervals for the estimated fracture length at the height level ern part of Iran to show that the FuzzyKrig toolbox is quite
of 1490 m capable of estimating fuzzy data to fit precise and fuzzy
variogram models. This power plant is located in Sanandaj-
Sirjan formation with alternation of sandstone, schist, phylite
of the cross validation method. Figure 5 shows the dispersion of and conglomerate. Figure 7 shows geological conditions in
the real versus the fuzzy mean; the correlation coefficient (0.7) this pumped storage power plant. The maximum, the mini-
and the slope of the regression line (0.93) are close to 1 (the mum, and the average lengths of the joints in all 12 drillholes
desirable value). Figure 6 shows a comparison of the real (true) in this area have been recorded. Figure 8 shows the experi-
values with the fuzzy width for a number of samples selected mental variogram of the data of the average joint lengths to
for cross validation. As shown, in most cases, the real value lies which a single-structured, spherical (13, 25, 75) variogram
between or close to the lower and upper extremes; in a limited model has been fitted. The data obtained from the drillholes
number of the samples the difference is, of course, considerable. and the structural analyses results are given in a file titled
A study of the position of the samples shows that such samples (input_dam1.mat). It contains: 1) geographic coordinates
are highly different in grade from their adjacent ones (e.g. S9 inthe first 3 columns of the drill matrix, and the minimum,
and S25) (Fig. 10). the average, and the maximum recorded fracture lengths at 62
0.4

Fig. 10 Comparison of the

estimated (black arrows) and
recorded (blue lines) fuzzy values
for the length of the fracture
0.3
Fracture length
0.2
0.1
0

s20
s10
s11
s12
s13
s14
s15
s16
s17
s18
s19
s1
s2
s3
s4
s5
s6
s7
s8
s9

s21
s22
s23
s24
s25
s26
s27
s28
s29
s30
s31
s32

Sample Name
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

244 Earth Sci Inform (2016) 9:235–245

3,913,600
fuzzy kriging are acceptable. It is worth noting that, similar
0.9
to the previous case study, in some samples, the intersection is
zero between the estimated and real values due to the high
3,913,400 0.8
fluctuations in the value of the variable.

0.7
Case III: fuzzy data and variogram
3,913,200
0.6
Since it was possible to fit the fuzzy variogram model to the
data related to Case study II, the same data set was used to
0.5
3,913,000 investigate the capability of the proposed program with re-
spect to Bardossky algorithm (1990). Figure 11 shows the
0.4 spherical fuzzy variogram model fitted to the experimental
variogram. Parameters of the fitted fuzzy variogram model
3,912,800
0.3
a r e ~θ1 ¼ ð10; 13; 16Þ, r e l a t e d t o t h e n u g g e t , ~θ2 ¼
ð30; 35; 40Þ; related to the sill and ~θ3 ¼ ð55; 75; 95Þ, related
0.2
to the range. Figure 12 shows the plot of the fuzzy interval
3,912,600
calculated at the 1450 m level. A comparison of the obtained
0.1 results shows that despite an increase in the volume of calcu-
lations (compared with the second case study), the values
3,912,400 0 estimated for the fuzzy mean have not had much variations,
642,500 642,700 642,900 643,100
and the major changes have been those of the width of the
Fig. 12 Fuzzy intervals for the estimated fracture length at the height
level of 1490 m
fuzzy value. As shown in Fig. 13, the definition of the fuzzy
variogram model has caused an increase in the width of the
points of these 12 drillholes incolumns 4–6, and 2) geographic fuzzy kriged value.
coordinates of the block center inthe first 3 columns of the drill
matrix, and the block dimensions along x, y, and z axes for 55,
326 blocks situated in the block model of the dam site Conclusions
incolumns 4–6. Figure 9 shows the plot of fuzzy interval cal-
culated at the1450 m level. When data are fuzzy, a good esti- Using kriging has been accepted today as the most common
mator’s fuzzy values have the highest in common with the method of estimating spatial data in such different fields as the
measured ones. Results from cross validation show that de- geosciences. To be able to apply kriging methods, it is neces-
spite very low correlation coefficient (0.12) between the mean sary that the data and variogram model parameters be precise.
of real and estimated fuzzy values, the range of real fuzzy To utilize the imprecise (fuzzy) data and parameters, use is
values, in most of the samples overlap well with the range of made of fuzzy kriging methods. Although it has been 30 years
estimated values (Fig. 10) and, therefore, results from the since different fuzzy kriging algorithms were proposed, its use
0.4

Fig. 13 Comparison of the width

of the fuzzy kriged value for the Based on the Fuzzy
fracture’s length based on the veriogram model
fuzzy and crisp variogram models Based on the crisp
0.3

veriogram model
Fracture length
0.2
0.1
0

s25
s1
s2

s10
s11
s12
s13
s14
s15
s16
s17
s18
s19
s20
s21
s22
s23
s24
s3
s4
s5
s6
s7
s8
s9

s26
s27
s28
s29
s30
s31
s32

Sample Name
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

Earth Sci Inform (2016) 9:235–245 245

has not become as common as other kriging methods (ordi- Bardossy A, Bogardi I, Kelly WE (1988) Imprecise (fuzzy) information
in geostatistics. Math Geol 20(4):287–311
nary, simple, log, universal, etc.); lack of a comprehensive
Bardossy A, Bogardi I, Kelly WE (1990a) Kriging with imprecise (fuzzy)
software that can perform, based on different fuzzy kriging variograms. I: theory. Math Geol 22(1):63–79
algorithms, the related calculations in a 3D space can be the Bardossy A, Bogardi I, Kelly WE (1990b) Kriging with imprecise (fuzzy)
main reason. FuzzyKrig is a software tool for performing cal- variograms. II: application. Math Geol 22(1):81–94
culations related to fuzzy kriging estimations under conditions Bartels F (1997) Ein Fuzzy-Auswertungs-und Krigingsystem für
raumbezogene Daten, in Informatik und Praktische Mathematik.
where the data or the variogram model or both are tainted with Universität Kiel
uncertainty. It has been presented in the Matlab software in an Cressie NAC (1993) Statistics for spatial data. John Wiley & Sons, New
open source form for scientific purposes. To enhance its user- York
friendliness, GUI has been used in its implementation so that it Davis B (1987) Uses and abuses of cross-validation in geostatistics. Math
Geol 19(3):241–248
can: 1) manage the required input data, 2) receive program Diamond P (1989) Fuzzy kriging. Fuzzy Sets Syst 33(3):315–332
commands, 3) manage the required outputs, and 4) draw the Diamond P, Armstrong M (1984) Robustness of variograms and condi-
required plots. Its other outstanding peculiarity is the compre- tioning of kriging matrices. J Int Assoc Math Geol 16(8):809–822
hensiveness in implementing different fuzzy kriging algo- Falivene O et al (2010) Interpolation algorithm ranking using cross-
validation and the role of smoothing effect. A coal zone example.
rithms; it enables the user to select the appropriate algorithm Comput Geosci 36(4):512–519
and implement it with due consideration to the nature of the Fletcher R (1987) Practical methods of optimization (2nd ed). John Wiley
data and the fitted variogram model. What encouraged me to & Sons, New York
prepare it were the limitations in finding access to a compre- Ghaderi M, Hezarkhani A, Talebi M (2007) The use of litho-geochemical
data and fluid inclusions in the study of Iju porphyry copper deposit,
hensive software to carry out fuzzy kriging calculations which
Northwest of Shahr-e-Babak, AmirKabir. J Sci Technol 67(3):51–63
is, perhaps, the main reason why using fuzzy kriging methods Handcock MS, Stein ML (1993) A Bayesian analysis of kriging.
is not welcome in the mine planning phase. Although it has Technometrics 35(4):403–410
been 30 years since these algorithms were proposed, the re- Hock W, Schittkowski K (1983) A comparative performance evaluation
of 27 nonlinear programming codes. Computing 30(4):335–358
ported number of their applications is quite small. Most of the
Loquin K, Dubois D (2010) Kriging and epistemic uncertainty: a critical
work is in the form of scientific papers that will finally lead to discussion. In: Jeansoulin R et al (eds) Methods for handling imper-
estimation; using estimated fuzzy numbers has not been re- fect spatial information. Springer, Berlin Heidelberg, pp 269–305
ported much. It is suggested here that future researches regard- Loquin K, Dubois D (2012) A fuzzy interval analysis approach to kriging
ing this issue be guided towards the application of fuzzy num- with ill-known variogram and data. Soft Comput 16(5):769–784
Nelles O (2001) Nonlinear system identification: from classical ap-
bers estimated by this software. proaches to neural networks and fuzzy models. Springer, Berlin;
New York
Acknowledgments The author is indebted to Professor H. A. Babaie Piotrowski JA et al (1996) Geostatistical regionalization of glacial
and the anonymous reviewers for their valuable comments on an earlier aquitard thickness in northwestern Germany, based on fuzzy
draft of this paper. kriging. Math Geol 28(4):437–452
Powell MJD (1978) A fast algorithm for nonlinearly constrained optimi-
zation calculations, in Numerical Analysis. Springer, Berlin
Heidelberg, pp 144–157
References Schelin L, Luna S (2010) Kriging prediction intervals based on
semiparametric bootstrap. Math Geosci 42(8):985–1000
Aalianvari A, Katibeh H, Sharifzadeh M (2010) A new approach for Soltani-Mohammadi S, Tercan E (2012) Constrained multiple indicator
computing permeability of fault zones case study: the upper reser- kriging using sequential quadratic programming. Comput Geosci
voir of Azad pumped-storage power station in Iran. Arch Min Sci 48:211–219
55(3):605–621 Taboada J et al (2008) Evaluation of the reserve of a granite deposit by
Aelion CM et al (2009) Validation of Bayesian kriging of arsenic, chro- fuzzy kriging. Eng Geol 99(1-2):23–30
mium, lead, and mercury surface soil concentrations based on inter- Webster R, Oliver MA (2007) Geostatistics for Environmental Scientists.
node sampling. Environ Sci Technol 43(12):4432–4438 John Wiley & Sons, New York