Environment International 52 (2013) 1728

Contents lists available at SciVerse ScienceDirect

Environment International
journal homepage: www.elsevier.com/locate/envint

Water quality analysis in rivers with non-parametric probability distributions and

fuzzy inference systems: Application to the Cauca River, Colombia
William Ocampo-Duque a,, Carolina Osorio a, Christian Piamba a, Marta Schuhmacher b, Jos L. Domingo c
a
b
c

Faculty of Engineering, Ponticia Universidad Javeriana, Cll. 18 #118-250, Cali, Colombia

Department of Chemical Engineering, Universitat Rovira i Virgili, Av. Pases Catalans 26, 43007 Tarragona, Spain
Laboratory of Toxicology and Environmental Health, School of Medicine, IISPV, Universitat Rovira i Virgili, Sant Llorens 21, 43201 Reus, Spain

a r t i c l e

i n f o

Article history:
Received 20 April 2012
Accepted 16 November 2012
Available online 23 December 2012
Keywords:
Water quality
Non-parametric density estimators
Uncertainty
Fuzzy inference systems
Monte Carlo simulation
Cauca River (Colombia)

a b s t r a c t
The integration of water quality monitoring variables is essential in environmental decision making. Nowadays,
advanced techniques to manage subjectivity, imprecision, uncertainty, vagueness, and variability are required in
such complex evaluation process. We here propose a probabilistic fuzzy hybrid model to assess river water
quality. Fuzzy logic reasoning has been used to compute a water quality integrative index. By applying a
Monte Carlo technique, based on non-parametric probability distributions, the randomness of model inputs
was estimated. Annual histograms of nine water quality variables were built with monitoring data systematically
collected in the Colombian Cauca River, and probability density estimations using the kernel smoothing method
were applied to t data. Several years were assessed, and river sectors upstream and downstream the city of
Santiago de Cali, a big city with basic wastewater treatment and high industrial activity, were analyzed. The
probabilistic fuzzy water quality index was able to explain the reduction in water quality, as the river receives
a larger number of agriculture, domestic, and industrial efuents. The results of the hybrid model were compared
to traditional water quality indexes. The main advantage of the proposed method is that it considers exible
boundaries between the linguistic qualiers used to dene the water status, being the belongingness of water
quality to the diverse output fuzzy sets or classes provided with percentiles and histograms, which allows
classify better the real water condition. The results of this study show that fuzzy inference systems integrated
to stochastic non-parametric techniques may be used as complementary tools in water quality indexing
methodologies.
2012 Elsevier Ltd. All rights reserved.

1. Introduction
Despite the huge numeric datasets collected nowadays, it is well
known that the assessment of water quality still relies heavily upon
subjective judgments and interpretation. Linguistic computations should
be considered together with numerical scoring systems to give appropriate water quality classications (Ocampo-Duque et al., 2006). There is no
doubt that the introduction of intelligent linguistic operations to analyze
databases is producing self-interpretable water quality indicators for a
better assessment. Moreover, to simplify and improve the understanding
and the interpretation of water quality, methodologies for integration, aggregation, and fusion of data must be developed (Sadiq and Tesfamariam,
2007). Data aggregation is not simply a problem of calculations; rather it
is a problem of judgment. Therefore, it deals not only with uncertainty
or variability related to random phenomena, but also with the subjective
uncertainty related to linguistic, subjective, vague and imprecise concepts
faced in decision-making processes. Consequently, Fuzzy Logic and

Corresponding author. Tel.: +57 2 321 8200.

E-mail address: willocam@javerianacali.edu.co (W. Ocampo-Duque).
0160-4120/$ see front matter 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.envint.2012.11.007

Monte Carlo based methods are highly recommended in water quality

management since they are appropriate tools to deal with all diverse
types of uncertainties (Chowdhury et al., 2009; Darbra et al., 2008).
Fuzzy inference systems (FIS) have recently attracted the attention of environmental scientists as suitable platforms to evaluate
multiple criteria related to water quality, and other environmental
conditions (Marchini et al., 2009). A common application of FIS has
been the integration of water quality variables to design suitable integrative systems, which are successfully compared to traditional indexing
techniques. Water quality is a vague term that cannot be easily described using crisp data or limited indicators. Instead, water quality
should be considered as a fuzzy term appropriately estimated with
linguistic computations (Mahapatra et al., 2011). The amount of linguistic if-then rules, as well as the number of indicators considered, seems
to be denitive for a robust and reliable evaluation (Lermontov et al.,
2009). In a previous study, we developed a structured fuzzy hierarchy
to interconnect various partial inference engines intended to dene
water quality (Ocampo-Duque et al., 2006). Here, the FIS contained an
analytical hierarchy process to deal with the relative weight of the
variables involved in the evaluation process. Adaptive and cooperative
neuro-FIS models have also been implemented to provide water

18

W. Ocampo-Duque et al. / Environment International 52 (2013) 1728

quality management solutions (Ocampo-Duque et al., 2007, 2012).

An integrated risk assessment methodology, based on the weight of
evidence approach, which implemented a FIS in order to hierarchically
aggregate a set of biological indicators following the precepts of the
Water Framework Directive, was recently described (Gottardo et al.,
2011). Also, Bayesian networks and probabilistic neural networks
have been recently used to train a water quality index supported in FIS
(Nikoo et al., 2011).
Probabilistic approaches are commonly applied in environmental
analysis and modeling to control uncertainty propagation. Parameter
uncertainty is a major aspect of the model-based estimation of the
risk of human exposure to pollutants. The Monte Carlo method is extensively applied despite it relies heavily on a statistical representation of
available information. The probability distributions of each variable are
dened according to the Bayesian theory (Ramaswami et al., 2005). For
instance, in human health risk assessment some variables are usually
managed as probability density functions (PDF) (Legay et al., 2011;
Mari et al., 2009). Probabilistic Monte Carlo computations are powerful
tools for water quality modeling (Cardona et al., 2011; Misha 2011).
However, their use in water quality indexing systems is scarce. A new
probabilistic water quality index intended for use in the production of
drinking water is described by Beamonte-Cordoba et al. (2010). In this
approach, each water quality variable is considered random with normal
distribution. Likewise, classical water quality indexes available worldwide
could be computed with Monte Carlo methods, assuming probability
distributions, or estimating them from monitoring data to provide a
most comprehensive evaluation.
Recently, fuzzy-probabilistic methods have emerged to deal with
complex problems related to water management (Chen et al., 2010;
Zhang K. et al., 2009; Zhang X. et al., 2009). Hybrid methods allow address
model parameter uncertainty in situations where available information
is not sufcient to identify statistically representative distributions.
Therefore, they assign fuzzy numbers when the amount of data is short,
or when the information about the condence intervals of variables and
parameters is unknown (Baudrit et al., 2007; Kentel and Aral, 2005).
For example, Faybishenko (2010) showed a recent application of combining probability and possibility theory for simulating a soil water balance.
Moreover, fuzzy-stochastic hybrid methods are currently used to solve
optimization and management issues associated to water pollution
(Guo et al., 2010; Rehana and Mujumdar, 2009; Zhang K. et al., 2009;
Zhang X. et al., 2009). In order to preserve the origin of uncertainties,
some methods partitioning the total variance in risk analysis have
been developed (Kumar et al., 2009). Likewise, current methodologies
are handling both random uncertainty and epistemic uncertainty, because they can combine the fuzzy set theory and Monte Carlo simulations (Li and Zhang, 2010; Li et al., 2007).
The method proposed in the present study is somehow inspired in the
formal concept of fuzzy randomness, which was rst introduced to structural analysis in civil engineering (Mller and Beer, 2004; Mller et al.,
2002). The idea behind such concept is that stochastic as well as
non-stochastic uncertainty is treated on the basis of the super-ordinated
uncertainty model fuzzy randomness. This new uncertainty model contains the special cases of real valued random variables and fuzzy variables,
and permits to take into account both uncertainty characteristics, simultaneously. Hybrid stochastic fuzzy model was also applied for in-ight gas
turbine engine diagnostics, where the random uctuations of performance parameters were modeled with PDF while the complex functional
relationships were dealt with Neural Networks with FIS structure, commonly called ANFIS (Ghiocel and Altmann, 2001). In the present study,
the objective was to model variables with two layers of analysis for uncertainty estimation, one inner layer of FIS using fuzzy membership functions and rules, and one outer layer using Monte Carlo simulation.
Randomness in water quality input variables was dealt with probability
theory. Then, decision about the water quality status was made by integration of these variables with the help of a FIS. In that way, we introduce
a combined stochastic fuzzy model to assess water quality in rivers. The

purpose of this research was to manage both the random nature of

input variables and the linguistic subjectivity present in the water quality
indexing process. A case study, with information from a Colombian River,
was selected to explain the application of the proposed method and its
benets. The results are here reported. Comparison with common indexes is also discussed. Consequently, the simulation outputs involved both
kinds of uncertainty: fuzzy and probabilistic.
2. Methods
2.1. Case study: the Cauca River
The Cauca River is one of the most important water resources in
Colombia. It has a length of 1350 km, with a basin area of approximately
63300 km 2. It goes across the country from south to north through nine
departments and a number of cities and towns without appropriate
wastewater treatment plants (WWTP). In fact, there are municipalities
without any kind of treatment of their sewage. In the Department of
Valle del Cauca there is a notable deterioration of water quality in the
river, especially when it receives discharges from the City of Santiago
de Cali. In this zone, a number of big river releases from domestic, agricultural, and industrial activities are present. The City of Santiago de
Cali, with more than two million inhabitants and several companies
located at Yumbo Industrial Park, is the main source of river pollution.
After crossing these areas, the organic loads are as high as to diminish
dissolved oxygen levels below 1 mg/L, compromising the ecosystems
living downstream and producing a clear reduction in its ecological
status. Although the environmental concerns about water pollution in
the river are commonly expressed by people and expert scientists, little
actions to recover the river to its original good ecological status, are
undertaken.
For the current assessment, a water quality monitoring database
including nineteen sampling sites was used. Data were provided by
the regional environmental protection agency, called the CVC Corporation (www.cvc.gov.co). Data from ten years, considering four sampling campaigns per year, were used. Fig. 1 shows the sampling points
where the data were collected: SP1 (Antes Suarez), SP2 (Antes Ovejas),
SP3 (Antes Timba), SP4 (Paso de La Balsa), SP5 (Paso de La Bolsa), SP6
(Hormiguero), SP7 (Antes Navarro), SP8 (Juanchito), SP9 (Paso del
Comercio), SP10 (Yumbo Puerto Isaacs), SP11 (Paso de la Torre),
SP12 (Vijes), SP13 (Yotoco), SP14 (Mediacanoa), SP15 (Puente Ro fro),
SP16 (Puente Guayabal), SP17 (Puente la Victoria), SP18 (Puente
Anacaro), SP19 (Puente La Virginia) (CVC Corporation, 2004).
2.2. Water quality analysis and traditional indexes
According to the objectives of this study, the Cauca River was
divided into three river sections: Section I (SP1 to SP6), Section II (SP7
to SP14), and Section III (SP15 to SP19). Thereby, the division includes
a relative less impacted area, an area highly impacted because of the
discharges from the city of Santiago de Cali and its industrial parks,
and an area where these impacts should be reduced due to natural
attenuation. Table 1 displays the main statistics of water quality variables
used in this study. These were: dissolved oxygen (DO), fecal coliforms
(FC), biochemical oxygen demand (BOD5), temperature (T), phosphates
(PO4), nitrates (NO3), turbidity (TUR), total solids (TS), and hydrogen
potential (pH). Three years are displayed equally time spaced. Sampling
campaigns included monitoring data in eld (pH, DO, T), and laboratory
measurements of composite samples. The sampling campaigns were
carried out during the same day in all sites. The 19 sites are monitored
in 4 periods: FebruaryMarch, MayJune, JulyAugust, and October
November, seeking stable hydrological conditions which are complex in
tropical regions.
Traditional water quality indexes are designed to integrate water
quality variables or indicators to provide a class or score about physicochemical and biological water quality status. It is intended that

W. Ocampo-Duque et al. / Environment International 52 (2013) 1728

19

RISARALDA
SP19
SP18

CHOCO
QUINDIO

SP17
SP16

SP15

SP14
SP13

CAUCA RIVER
SP12
SP11
SP10

COLOMBIA

SP9
SP8

VALLE DEL CAUCA

SP7
SP6

SP5
SP3

SP4

SP2

16.000

32.000

Meters
64.000

SP1
Fig. 1. Map of the studied area: the Cauca River in the Valle Department (Colombia).

they are useful in environmental decision making. A commonly

referred water quality index was developed by the National Sanitation Foundation of United States (NSF_WQI) (Brown et al., 1970). It
was dened for any use of water by simply determining the specications required by that use. This index included various physical,
chemical and biological characteristics. For each variable, the index included a quality-value function that expressed the equivalence between
the variable and its quality level. The strongly subjective character of the
equivalence functions is a problem with that index (Beamonte-Cordoba
et al., 2010). The NSF_WQI is computed with Eq. (1),
NSF WQI i1 wi Q i
N

variable i. At local level, in the Cauca river basin, the CVC Corporation
also uses the ICAUCA index to evaluate the water status (Torres et al.,
2010). This index is computed according to Eq. (2),
N

wi

ICAICA i1 Ii

where Ii is a special function dened for the variable i to transform the

real value to a normalized quality number. The functions to calculate
both indexes may be consulted in (CVC Corporation, 2004).
2.3. Fuzzy inference systems

where wi is the weight of the variable, usually dened by experts, N is

the number of variables, and Qi is the quality value function of the

It has been recently shown that linguistic computations used in

fuzzy inference systems (FIS) are superior to algebraic common expressions for water quality indexing evaluation (Lermontov et al.,

20

Table 1
Basic statistics of water quality variables involved in the study.
Indicator, abbr., units

Fecal coliforms, FC, CFU/100 mL

Biochemical oxygen demand, BOD5, mg/L

Temperature, T, C

Phosphates, PO4, mg/L

Nitrates, NO3, mg/L

Turbidity, TUR, NTU

Total solids, TS, mg/L

Hydrogen potential, pH, ()

2002
2006
2010
2002
2006
2010
2002
2006
2010
2002
2006
2010
2002
2006
2010
2002
2006
2010
2002
2006
2010
2002
2006
2010
2002
2006
2010

Section I

Section II

Section III

X

Min

Max

X

Min

Max

X

Min

Max

76.09
73.08
72.16
1.51E + 05
1.12E + 04
1.05E + 05
1.55
1.87
8.51
20.4
21.3
22.4
0.062
0.034
0.069
0.30
0.42
0.84
30.8
110.8
79.1
131.33
181.25
163.29
6.93
6.88
7.15

20.71
14.87
18.65
5.13E + 05
2.52E + 04
2.61E + 05
1.12
0.76
3.31
2.8
1.9
0.8
0.008
0.010
0.016
0.20
0.02
0.89
20.0
117.5
95.1
63.25
108.37
145.42
0.47
0.69
0.30

24
24
24
24
24
24
24
24
24
24
24
24
24
24
24
24
24
24
24
24
24
24
24
24
24
24
24

17.11
37.99
22.65
0.00E + 00
2.30E + 01
7.30E + 02
0.30
1.09
5.33
15.0
18.0
20.9
0.060
0.021
0.064
0.11
0.40
0.11
3.0
9.0
2.0
68.00
59.00
58.00
5.76
5.30
6.45

83.26
94.42
94.68
2.40E + 06
1.10E + 05
9.30E + 05
5.30
4.02
16.00
24.2
28.8
24.3
0.099
0.050
0.125
1.05
0.44
2.57
75.0
349.0
344.0
310.00
396.00
721.00
7.98
7.62
7.65

27.90
47.95
35.70
5.97E + 07
2.72E + 05
5.35E + 06
5.28
3.96
19.68
23.8
21.1
24.4
0.076
0.099
0.089
0.26
0.57
0.98
67.3
143.1
131.1
172.94
270.09
233.25
6.98
6.82
7.05

25.45
26.16
23.96
1.01E + 08
4.96E + 05
1.62E + 07
2.92
1.51
6.99
2.0
1.2
1.3
0.039
0.047
0.025
0.25
0.10
1.10
53.4
125.4
135.6
51.33
134.88
121.32
0.19
0.32
0.19

32
32
32
32
32
32
32
32
32
32
32
32
32
32
32
32
32
32
32
32
32
32
32
32
32
32
32

2.76
7.47
7.02
2.40E + 04
7.50E + 03
9.10E + 04
1.30
1.75
9.82
20.0
18.0
22.2
0.060
0.031
0.064
0.04
0.45
0.11
30.0
18.0
17.0
68.00
129.00
116.00
6.58
5.65
6.68

77.34
85.32
75.30
2.40E + 08
2.40E + 06
9.30E + 07
13.80
7.52
36.90
27.0
23.0
27.1
0.216
0.241
0.157
1.53
0.69
3.29
300.0
404.0
670.0
302.00
811.00
621.00
7.27
7.39
7.32

34.64
32.00
37.22
1.82E + 05
1.79E + 04
3.15E + 05
2.79
3.44
20.66
25.2
24.0
25.3
0.065
0.083
0.084
0.43
0.57
1.39
61.2
201.3
107.4
203.55
338.35
256.71
6.90
7.04
7.36

8.84
7.51
15.68
5.30E + 05
2.98E + 04
7.16E + 05
0.83
1.08
24.20
0.9
1.4
1.5
0.015
0.026
0.045
0.20
0.38
1.40
34.5
231.6
75.4
80.53
156.07
139.37
0.31
0.36
0.33

20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20

20.50
18.20
8.81
2.40E + 03
2.40E + 02
2.30E + 03
1.20
2.05
6.73
22.8
21.5
22.6
0.060
0.053
0.064
0.07
0.40
0.11
29.0
21.0
23.0
136.00
191.00
0.08
6.22
6.40
6.72

55.98
45.90
69.32
2.40E + 06
1.10E + 05
2.40E + 06
4.30
5.77
121.00
26.7
26.3
27.5
0.123
0.142
0.271
0.78
2.01
4.28
185.0
892.0
265.0
406.00
901.00
551.00
7.48
7.54
7.90

Note: X
is the median, s is the standard deviation, N is the number of data, Min is the minimum, Max is the maximum. Abr. is the abbreviation of the water quality variable.

W. Ocampo-Duque et al. / Environment International 52 (2013) 1728

Dissolved oxygen, DO, % Sat.

Year

W. Ocampo-Duque et al. / Environment International 52 (2013) 1728

2009; Ocampo-Duque et al., 2006). Water quality assessment is a

subjective task that must be carried out with tools able to manage
such subjectivity and imprecision. Here, linguistic operations in a
FIS frame are proposed to compute water quality by integrating parameters within an inference engine. Thus, a methodology to design
a water quality index is proposed. It could be adapted to diverse purposes with different number of inputs. In this sense, a FIS is a mapping process from given water quality inputs to desired water
quality index. The FIS involves three important parts: membership
functions, fuzzy set operations, and inference rules. The Fuzzy
Logic toolbox of MATLAB (R2010) was used to build and compute
the FIS.
A FIS was parameterized to assess water quality considering nine
input indicators (Table 1), using the same indicators that those included
in the well-known NSF_WQI and the ICAUCA. The FIS output is a fuzzy
water quality (FWQ) index. Table 2 summarizes the parameters of the
membership functions. Five fuzzy sets were dened for input variables:
Very low, low, medium, high, and extreme. In turn, the output
water quality was dened according to ve fuzzy sets (qualiers):
poor, bad, regular, good, and excellent. Gaussian functions
were used at low, medium, high, bad, regular and good fuzzy sets, having
the following expressions:



xc
x; s; c exp
2
2s

where s and c are the parameters shown in Table 2, x is the value of the
input, and is the belongingness (or membership) of the input to the
respective fuzzy set, which is a number between 0 and 1, meaning none
and total membership, respectively. The parameter c represents the
center of the function in the abscissa where the membership value is 1,
and the parameter s denes the width of the function. It is important to
point out that in fuzzy logic reasoning an x value may belong to more
than one fuzzy set. Z-shape functions were used in very low and poor
fuzzy sets, having the following equations to represent them:
9
8
1;
xa
>
>
>
>
xa2
>
>
>
>
a

b
>
>
>
>
; ax
=
< 12
ba
2


2
x; a; b
xb
a

b
>
>
>
xb >
;
>
> 2
>
>
>
>
ba
2
>
>
;
:
0;
xb

where a and b are the parameters displayed in Table 2. These parameters

locate the extremes of the sloped portion of the curve. Finally, S-shape

21

functions were used in extreme and excellent fuzzy sets, having the
following equations to represent them:
8
>

0; 2
>
>
>
xd
>
>
;
< 2
ed


2
x; a; b
>
xd
>
>
12
;
>
>
ed
>
:
1;

9
xd
>
>
>
d e>
>
>
dx
=
2
>
de
>
xe >
>
>
2
>
;
xe

where d, and e are the parameters shown in Table 2. These parameters

locate the extremes of the sloped portion of the curve.
The design and selection of membership functions from intervals of
the input variables is a very subjective task. The main questions arise
from the number of fuzzy sets used to divide the ranges of the variables,
and the own shape of these sets. A division in ve fuzzy sets seems
appropriate. However, the number of rules may considerably increase,
especially if rules with more than one antecedent are desired. In continuous variables the number of fuzzy sets to represent any range could be
selected from three to seven, being ve a reasonable number. The shape
of the functions selected above was considered because of the low number of parameters required. Notwithstanding, other functions could be
also used, perhaps requiring more than two parameters. The Colombian
Decree 1594/1984 and Resolution 2115/2007, the Spanish Decree 927/
1988, the boundaries taken in the Lermontov fuzzy water quality index
(Lermontov et al., 2009) and the limits set by our previous study
(Ocampo-Duque et al., 2006) were used to dene the ranges from
very low to extreme that water quality variables could take. Then, the
division in ve qualiers was given trying to equally divide the universe
of discourse with appropriate fuzzy intersection between sets.
The inference engine is where the linguistic computations are executed. It was created considering two kinds of rules: rules with only
one antecedent, and rules with two antecedents or water quality variables. Forty ve (45) rules were written with one antecedent and one
consequent (9 water quality variables per 5 fuzzy sets or options,
from very low to extreme). Nine hundred (900) rules were written
with two antecedents and one consequent. All the likely combinations
without repetitions were considered ((81 9)/2 = 36 pair combinations and 25 options). In each rule, the most conservative output was
considered, and the importance of the rule was dened according to
the importance of the variables involved. Rules with DO, pH, BOD5
and FC received a weight of 1.0. Rules with NO3, PO4 and T, received a
weight of 0.75. Finally, rules with TUR and TS received a weight of 0.5.
More complex rules with three or more antecedents could be created,

Table 2
Parameters of the fuzzy inference system.
Indicator*

Units

Membership function parameters

Very Low
a

Low
b

Z-shape
DO
FC
BOD5
T
PO4
NO3
TUR
TS
pH
FWQ

% Sat.
CFU/100 mL
mg/L

C
mg/L
mg/L
NTU
mg/L

0.0
58.9
0.0
15.1
0.0
0.0
3.0
25.6
5.0
Poor
a
Z-shape
0.0

Medium
c

Gaussian
27.8
272.0
2.2
19.9
0.15
3.8
30.7
230.4
6.5
b
38.9

15.0
143.3
1.2
2.6
0.07
1.6
15.0
80.0
0.5
Bad
s
Gaussian
10.5

High
c

Gaussian
31.3
337.5
1.5
18.9
0.14
3.2
33.5
150.6
6.4
c
35.5

15.0
143.3
1.2
2.6
0.07
1.6
15.0
80.0
0.5
Regular
s
Gaussian
11.6

Extreme
c

Gaussian
58.2
675.0
3.5
23.0
0.25
6.1
70.7
300.0
7.5
c
60.0

15.0
143.3
1.2
2.6
0.07
1.6
15.0
80.0
0.5
Good
s
Gaussian
9.4

S-shape
84.1
1013.0
5.2
27.2
0.4
9.5
107.4
450.0
8.5
c
81.4

70.0
1078.0
5.0
25.1
0.3
7.2
88.7
395.0
8.0
Excellent
d
S-shape
68.2

110.0
1284.0
6.9
30.0
0.5
12.0
136.8
642.0
9.5
e
100.0

*DO: dissolved oxygen, FC: fecal coliforms, BOD5: biochemical oxygen demand, T: temperature, PO4: phosphates, NO3: nitrates, TUR: turbidity, TS: total solids, FWQ: Fuzzy water
quality index. a, b, s, c, d, and e, are the parameters to build the membership functions according to Eqs. (3)(5).

22

W. Ocampo-Duque et al. / Environment International 52 (2013) 1728

Fig. 2. Conceptual integration of non-parametric Monte Carlo modeling with a Fuzzy Inference System.

although the improvements are not signicant. Rules and ranges were
tested with several environmental experts from the CVC Corporation
and Academia. Some examples of rules are shown:

are dened on the universe X, for a given element x belonging to X,

the following operations can be carried out:
Intersection; AND :

If fecal coliform is very low then water quality is excellent,

If dissolved oxygen is high then water quality is good,
If phosphate is medium then water quality is regular,
If nitrate is high then water quality is bad,
If BOD5 is very high then water quality is poor,
If fecal coliform is very low and dissolved oxygen is very high
then water quality is excellent.
Computations with words within the inference engine followed
standard fuzzy set operations. These are: union (OR), intersection
(AND) and additive complement (NOT). If two fuzzy sets A and B

AB x min A x; B x

Union; OR : AB x max A x; B x

Additive complement; NOT : A x 1 A x:

Vector inputs are fuzzied to enter to the inference engine using the
membership functions. When there are two antecedents, fuzzy logic operations are applied to give a degree of support for these rules. In rules
with one antecedent, their degree of support is the degree of membership. The degree of support for the entire rule is used to shape the output
fuzzy set. The consequent of a fuzzy rule assigns an entire fuzzy set to the

Fig. 3. Propagation of uncertainty when a probabilistic variable is introduced to a fuzzy inference system.

W. Ocampo-Duque et al. / Environment International 52 (2013) 1728

output. This fuzzy set is represented by a membership function, which is

chosen to indicate the qualities of the consequent. If the antecedent is
only partially true (i.e., b 1), the output fuzzy set is truncated at this
value. This procedure is called the minimum implication method. Since
decisions are based on the testing of all the rules in the system, these
must be aggregated to make a nal decision. Therefore, output fuzzy
sets of each rule are aggregated to a single output fuzzy set that may
have a complex geometry. The aggregation procedure used here was
the maximum method, which is the union of all truncated output
fuzzy sets (Mathworks, 2012). The nal step was defuzzication to provide a numerical water quality score. A convenient way to give FIS outputs is also by means of the linguistic fuzzy sets with their respective
membership degrees. In the current study, the centroid method was
used for defuzzication. It delivers a numeric score to water quality, so

FWQ

zzdz
zdz

where FWQ is a fuzzy water quality index which is a score between

0 a 100, and z is the independent variable of the output fuzzy set in
each rule. Fuzzy water quality indexes have recently been proposed
(Lermontov et al., 2009; Ocampo-Duque et al., 2006).
2.4. Monte Carlo simulation of FIS
When the fuzzy water quality index is stochastically computed with
Monte Carlo method, a stochastic fuzzy water quality index is obtained.
The stochastic model used in this study is described below. Obviously,
the building of a FIS for water quality analysis is extremely subjective.
The number of input variables should be considerably higher than
nine, since the number of physicochemical, microbiological, and biological variables measured nowadays in rigorous water protection agencies
may be greater than hundreds. The creation of appropriate fuzzy rules is
an important issue for increasing the preciseness of the simulation. A
considerable number of fuzzy rules may make more accurate the decision from the inference engine. However, if the number of input variables increases, the number of rules would also increase exponentially
to thousands or millions, which would make extremely more complex
the model requiring powerful computation to deliver a single score
under stochastic conditions. Moreover, the number and form of the
rules, as well as the shape of the ranges of the membership functions,
are also subjective complex decisions, which could be designed for specic, regional and/or local requirements. Therefore, we here propose a
convenient method to build a FIS for water quality evaluation purposes

23

rather than a standard index for use anywhere. Because of the high random uncertainty in water quality variables, due to experimental measurement, human errors, and propagation of error due to the methods
used to measure the water quality variable, we propose treating the
FIS inputs as stochastic. The conceptual model is depicted in Fig. 2.
The algorithm for Monte Carlo simulation assumes each computation with the FIS as deterministic. A vector with water quality variables is randomly selected according to its probability distribution
over the domain. Then the corresponding water quality score to
that vector is computed with the FIS. The computation is carried out
a consistent number of times to cover the entire range of likely inputs,
and to build a well-dened histogram of the water quality scores.
Random numbers were generated with the inverse transform method. The quantity of random numbers was set at 10 000 in all cases.
Fig. 3 outlines the propagation of uncertainty when a probabilistic
variable is introduced to a FIS. A, and C, are fuzzy sets. Arrows point
out the information ow. Suppose a measured water quality variable
X, continuous, positive, and random, with probability density function, f(X) ~ PDF, as shown in Fig. 3, to be introduced to the fuzzy system. Let X
, Q1, and Q3 be the median, the 25th and 75th percentiles
of the data, respectively. When X is introduced to the fuzzy system,
the probabilistic or random uncertainty is transformed into fuzzy
uncertainty. First, X
is fuzzied to take the membership value A(X
).
A( X
) is the degree of membership of X
to the set A. Then, A( X
) is
transformed to C(y) according to the rule:
If X is A then y is C:

10

Such transformation is computed according to the implication

method in fuzzy reasoning. In the case of the Figure, the reasoning
leads to the horizontal projection line from left to right, or A( X
) =
C(y), as shown in the Fig. 3. The shape and size of the output fuzzy
set is dened by the C(y) value where the output set is truncated.
The area of the output fuzzy set is shown in dark gray. Observe that
U is the uncertainty in the height of the output fuzzy set after
fuzzication of the random variable X when the interquartile range
(IQR) is computed. The area of the output fuzzy set in every rule is required in the centroid method to provide the nal output single water
quality score. The centroid computes the center of area under the
curve resulting after aggregation of all fuzzy sets within the inference
engine. Therefore, the uncertainty in the area of the fuzzy set do affects
the water quality score. The Monte Carlo method allows computing the
nal effect over the propagation of uncertainty when dealing with a
random input variable in a FIS to provide a nal defuzzied water quality score, which also leaves the system with an empirical probability

Fig. 4. Examples of optimized tting of non-parametric versus parametric distributions of two input variables. (Data of 2009, Section II).

24

W. Ocampo-Duque et al. / Environment International 52 (2013) 1728

Fig. 5. Box-and-Whisker plots for assessed water quality with the stochastic fuzzy
water quality index (SFWQI) for different years and the three river sections. Reported
values are the medians.

density function. Thus, the shapes of the output fuzzy sets vary with
each run as a random input is chosen. Propagation of uncertainty is
somewhat expressed in this context as the transformation of probabilistic uncertainty into fuzzy uncertainty through the every membership
function and rule evaluation. Such propagation is graphically represented as the uncertainty in the area of the output fuzzy set (U) when the
random input takes a number between Q1 and Q3. To compute such uncertainty, deterministic computations of the FWQ index are performed
depending on the probability of water quality inputs randomly chosen
within the statistic range of the water quality variables. Therefore, two
layers of uncertainty may clearly be identied. The fuzzy uncertainty
is self-contained in the FWQ number as long as probabilistic uncertainty
is observed through the output FWQ histogram.

parameters of the assumed distribution from the data. This is the most
common way to apply the PDF in environmental uncertainty analysis,
with multiple tools available. The main disadvantage of the parametric
approach is the lack of exibility. Each parametric family of distributions imposes restrictions on the shapes that f(x) may have. For example, the density function of the normal distribution is symmetrical and
bell-shaped, and therefore, it is unsuitable for representing skewed
densities or bimodal densities, which may appear in real water quality
datasets. The idea of the non-parametric approach is to avoid restrictive
assumptions about the form of f(x), and to estimate it directly from the
water quality monitoring data (Qin et al., 2011). It could be especially
useful if data are limited. A well-known non-parametric estimator of
the PDF is the histogram, when classes are properly well dened. Likewise, the kernel density estimation method is a widely used method for
density estimation.
The most attractive feature of non-parametric kernel density estimation is that it directly makes use of sample data without a need of
estimating characteristic parameters in a theoretical distribution. In
other words, there is no error caused by assumption of a theoretical
distribution for data and by mismatch between estimated parameters
and actual behaviors of water quality indicators. Let X1, X2,, Xn
denote n water quality variable samples. The real probability density
function f(x) of a water quality variable can be estimated by the
following density function:



n
1 X
xX i
f^n x
K
nh i1
h

where h is the bandwidth, K is called the kernel function and n is the

sample size. Gaussian functions are commonly selected as kernel
functions:

2.5. Non-parametric kernel density estimator

The use of probability distributions to assess water quality, when the
integration of variables is required, may provide a better estimation,
since the outputs of the fuzzy water quality index will also have probability density rather than point estimation. Consequently, stochastic
fuzzy water quality indexes are estimated. Thus, the nal classication
will be more realistic. The probability distribution of a continuousvalued random variable X is conventionally described in terms of its probability density function (PDF), from which probabilities associated with X
can be determined using the relationship
b

P abXbb f xdx:

11

The parametric approach for estimating f(x) is to assume some parametric family of probability distributions, and then to estimate the

12

!



xX i
1
xX i 2
p exp
:
h
2h2
2

13

The determination of the bandwidth h is crucial for accurate estimation of water quality variable distributions. There are many ways
to estimate an optimal bandwidth (hopt). An approximation, known
as the Silverman's rule (Silverman, 1998) has been proposed:


hopt


4 1=5
3n


where min s;

14
IQR
20:6745


n
2
1
, s2 n1
i1 xi x
, and IQR is the

interquartile range of the data. Therefore, parametric or non-parametric

PDF should be estimated for annual data sets to each input water quality

Table 3
Classication of the water quality according to the membership degree of the fuzzy sets.
Year

2002

2006

2008

2009

2010

Section

I
II
III
I
II
III
I
II
III
I
II
III
I
II
III

Lower Quartile (0.25)

Median

Upper Quartile (0.75)

Bad

Regular

Good

Bad

Regular

Good

Bad

Regular

Good

0.344
0.408
0.444
0.336
0.431
0.459
0.322
0.379
0.395
0.348
0.364
0.419
0.379
0.376
0.534

0.741
0.678
0.642
0.750
0.654
0.628
0.766
0.707
0.691
0.737
0.722
0.667
0.707
0.709
0.553

0.005
0.003
0.002
0.005
0.003
0.002
0.006
0.004
0.035
0.005
0.004
0.003
0.004
0.004
0.001

0.310
0.355
0.406
0.302
0.382
0.418
0.292
0.347
0.358
0.326
0.331
0.377
0.343
0.315
0.413

0.777
0.730
0.679
0.786
0.703
0.667
0.797
0.739
0.727
0.760
0.755
0.708
0.743
0.745
0.672

0.006
0.005
0.003
0.007
0.004
0.003
0.007
0.005
0.005
0.006
0.005
0.004
0.005
0.005
0.006

0.266
0.317
0.360
0.268
0.333
0.369
0.268
0.314
0.323
0.301
0.377
0.339
0.314
0.313
0.354

0.823
0.769
0.725
0.822
0.753
0.716
0.822
0.773
0.764
0.787
0.708
0.747
0.773
0.775
0.732

0.009
0.006
0.004
0.009
0.005
0.004
0.009
0.088
0.006
0.007
0.004
0.005
0.006
0.006
0.005

W. Ocampo-Duque et al. / Environment International 52 (2013) 1728

25

2010:
Water Quality

600

700
600

500

800

500

400

600

Frequency

1000

Frequency

Frequency

Water Quality

Water Quality

1200

400
300

400

300

200
200

200

100

100

0
47.25

49.50

51.75

54.00

56.25

58.50

60.75

63.00

48

49

50

2010 Section I

51

52

53

54

45.6

46.8

48.0

2010 Section II

49.2

50.4

51.6

52.8

54.0

2010 Section III

2009:
Water Quality

Water Quality

Water Quality

700

500
900

600

800

400

700

400
300

Frequency

Frequency

Frequency

500
300

200

600
500
400
300

200

200

100
100

100

0
48.8

49.6

50.4

51.2

52.0

52.8

53.6

0
46.2

54.4

47.3

48.4

2009 Section I

49.5

50.6

51.7

52.8

0
44.8

53.9

46.2

47.6

2009 Section II

49.0

50.4

51.8

53.2

2009 Section III

2008:
Water Quality

Water Quality

Water Quality
800

600
400

700
600

400
300

300

Frequency

Frequency

Frequency

500

200

500
400
300

200
200

100
100

100
0

0
49.6

50.4

51.2

52.0

52.8

2008 Section I

53.6

54.4

0
47.7

48.6

49.5

50.4

51.3

52.2

53.1

54.0

2008 Section II

45.6

46.8

48.0

49.2

50.4

51.6

52.8

54.0

2008 Section III

Fig. 6. Non parametric distributions of the stochastic fuzzy water quality index in the Cauca River for some selected years.

variable prior to the FIS calculations. Fig. 4 depicts two examples of

distribution ttings carried out to estimate the best probability distributions better representing the variables BOD5 and total solids, for 2009.
Non-parametric versus parametric distributions are compared, corresponding the best t to the kernel method in both cases.
The KolmogorovSmirnov (KS) test was used to assess the
goodness-of-t of the entire input PDF variables. It is a nonparametric
test for the equality of continuous, one-dimensional probability distributions that can be used to compare a sample with a reference probability distribution. The most attractive characteristic of KS test is that it is
applicable for any continuous variable distribution, and any sample size.
A smaller statistic represents the better goodness-of-t between assumed theoretical distribution, and actual variable samples. Therefore,
the statistics can be used to rank the performances of all the water quality variable distributions including the proposed non-parametric kernel
distribution estimation (Qin et al., 2011). The KS test evaluates if a
sample comes from a continuous distribution with specied parameters, against the alternative that it does not come from that distribution.
The test rejects the null hypothesis at the 5% signicance level (p b 0.05).
All statistic calculations were performed with the statistics toolbox of
MATLAB (R2010). The best goodness of t for 84% of data was obtained
for non-parametric kernel density estimators. The remaining 16% of
data presented KS statistics for parametric ttings similar to that for
non-parametric estimators. That was a good reason to choose the

nonparametric method to build all the probability distributions. The

ksdensity (Kernel smoothing density) algorithm was used to compute
the probability density estimates of the input variables (Mathworks,
2012). The estimate is based on a normal kernel function, using a window parameter (bandwidth) that is a function of the number of points
(Eq. (14)). The density is evaluated at equally spaced points that cover
the range of the data. The ksdensity algorithm makes no assumptions
about the mechanism producing the data or the form of the underlying
distribution. Therefore, no parameter estimates are made. In other
words, it produces a nonparametric density estimate that adapts itself
to the data. Likewise, the ProbDistUnivKernel object, which represents
a nonparametric probability distribution based on a normal kernel
smoothing function, was used to deal with all the PDFs (Mathworks,
2012).
3. Results and discussion
The water condition for the Cauca River when crossing the Valle
Department in Colombia has been here assessed with a water quality
index built on a FIS. Input data have been provided by the CVC Corporation, a regional environmental protection agency. We assessed various years using stochastic modeling with non- parametric kernel
estimators of inputs. Hence, integration between fuzzy and stochastic
models was carried out to manage both random and linguistic

26

W. Ocampo-Duque et al. / Environment International 52 (2013) 1728

Table 4
Comparison of the fuzzy water quality index versus other indexes after Monte Carlo simulations (medians are provided). Membership values in linguistic scores, computed with
fuzzy modeling, are provided.
Index

Stochastic NSF_WQI

Stochastic ICAUCA

Stochastic FWQ

Year

Section I

Section II

Section III

Numeric score

Linguistic score

Numeric score

Linguistic score

Numeric score

Linguistic score

2002
2006
2008
2009
2010
2002
2006
2008
2009
2010
2002

63
56
58
61
56
42.65
63.74
63.03
74.96
40.49
51.58

48
45
50
49
40
27.49
54.95
28.80
30.64
26.70
50.59

51.81

2008

51.98

2009

51.18

2010

50.85

Bad
Bad
Bad
Bad
Bad
Bad
Good
Bad
Bad
Bad
0.35 bad
0.73 regular
0.38 bad
0.70 regular
0.34 bad
0.73 Regular
0.33 bad
0.75 regular
0.32 bad
0.75 regular

49
47
46
49
41
30.09
44.03
42.54
55.26
26.77
49.74

2006

Regular
Regular
Regular
Regular
Regular
Regular
Good
Good
Good
Regular
0.31 bad
0.77 regular
0.30 bad
0.78 regular
0.29 bad
0.79 regular
0.32 bad
0.76 regular
0.34 bad
0.74 regular

Bad
Bad
Bad
Bad
Bad
Bad
Regular
Regular
Good
Bad
0.40 bad
0.67 regular
0.41 bad
0.66 regular
0.35 bad
0.72 regular
0.37 bad
0.70 regular
0.41 bad
0.67 regular

uncertainty in the analysis. Consequently, a stochastic fuzzy water

quality index was developed. Fig. 5 shows the behavior of the stochastic FWQ index for ve years and the three river sections assessed
using a Box-and-Whiskers plot. Most of the time, the stochastic FWQ
index in river Section I was higher than that in river Section II, and in
river Section III, respectively. Therefore, it is noted with the index that
water quality decreases downstream. Generally, there is a bigger dispersion in river Section III, as it is shown by their box heights. Moreover, it is
noticeable the symmetry of the median, as well as the close distance
between estimated boundary values. Such numeric scores agree well
with expert and non-expert opinions, since water quality control is
minimal in such area, and water pollution is considerably increased
downstream. Despite the limited monitored variables considered in the
current study, and included in the index, the FIS model adequately
describes the observed condition. The low water quality scores should
be inferred from a brief inspection to Table 1. Low dissolved oxygen concentrations, specically in Sections II and III, were observed with very low
saturation percentages. Also, high fecal coliform concentrations were
available. Moderate (medium to high) concentration of total solids and
turbidity are also common features in the area. BOD5 was high in Section
III with observed increase in time, since in 2010, the values were
considerably higher than those in 2006 and 2002.
With the aid of the fuzzy stochastic analysis, it is possible to map
fuzzy random input parameter into fuzzy random responses. The
stochastic fuzzy behavior of the FWQ index and some of its advantages
are shown in Table 3. Here, the membership functions () to the diverse
output fuzzy sets are calculated. As mentioned above, the membership
to each fuzzy set is a number between zero (0) and one (1), meaning
none and total membership, respectively. Partial memberships are
also possible, which is one of the advantages of fuzzy logic for environmental decision making. It must be remarked that the sum of specic
set membership values could be higher than 1. The membership degrees may be stated as possibility values to not confuse them with probabilistic computations. Table 3 presents the calculated membership
values to the sets bad, regular and good. Such score was zero in all
years, reaching to poor and excellent water quality fuzzy sets. In all
cases, the belongingness estimation to the good water quality fuzzy
sets was really low. The fuzzy sets with the higher membership values
were the ones to regular water quality classication during all years,
and through the three river sections with median variations between
0.797 and 0.667. Likewise, the membership value to bad water quality
classication during all years and through the three river sections had

50.00
50.78
51.05
50.89

49.47
50.54
50.10
49.45

median variations between 0.413 and 0.292, which is a considerable

belongingness. Moreover, it is noticeable that the membership values
to the regular water quality fuzzy sets decreased downstream while
they increased for bad water quality fuzzy sets. This result may be associated to the higher number of domestic, agricultural and industrial
loads available downstream. In Table 3 can be seen that the membership degrees for bad water quality are higher in the lower quartile
than in the median, which is due to the lower quartile of the index is
the worst condition. For example, in 2002 and Section I, the membership degrees to the bad fuzzy set were 0.344, 0.310, and 0.266 for the
lower quartile, median and upper quartiles, respectively.
Output histograms are necessary in indexing computations since
with point estimates the results could be limited. Fig. 6 shows the frequency histograms (or the nonparametric density estimations) for
the stochastic FWQ index of some assessed years through the three
sections. They point out the randomness of water quality integration
outputs when random inputs are provided to the FIS. So, the output
spread is easily observed in such gures. As it can be noted, diverse
shapes are possible. Some histograms showed peak shapes with relative symmetrical variability. Likewise, some histograms with wide
peaks (2008, Sections II and III) were calculated. The biggest dispersion was generally observed for river Section III. Moreover, in some
cases appeared bimodal histograms, especially in Section III, although
the closeness between peaks was enough to get appropriate classication and unambiguous outputs.
In order to validate the performance of the stochastic fuzzy water
quality index, similar stochastic computations were carried out for
the indexes: NSF_WQI and ICAUCA. Results of the medians are given
in Table 4. From the NSF_WQI calculations, it can be observed that
they always provided a consistent output, classifying as regular the
river Section I, and bad the river Sections II and III, for all the assessed
years. In this case, numeric scores ranged between 63 and 41, with a
spread of 22. The assessment with the ICAUCA index was less strict,
delivering good water quality classication in some cases. ICAUCA
index outputs were between 74.93 and 26.70, with a range of 48.26.
The outputs from stochastic FWQ index were similar to the other indexes. The dispersion of the stochastic FWQ index results was lower
than the other indexes, being within a maximum of 51.98 and a minimum of 49.45, with a range of 2.53. Although the numeric score of the
defuzzied stochastic FWQ index is important, the advantage of the
hybrid probabilistic fuzzy index over the others is that the last one
provides a classication with more information related to the

W. Ocampo-Duque et al. / Environment International 52 (2013) 1728

belongingness to diverse possible classications. In all cases, the stochastic FWQ index outputs have classied water quality in the studied
area as partially bad and partially regular. Lower possibility has the
water quality to belong to the good class (in Table 3, observe that
b 0.01). As above stated, the belongingness to bad water quality sets
increased downstream from the river Section I to Section III. It agrees
with the results from the NSF_WQI. Consequently, the belongingness
to the regular class decreases downstream.
Water quality indexes based on fuzzy systems have been recently
proposed in scientic literature with relative success. The fuzzy frame
clearly improves the conceptual design of the indexes, because they
are computed with expert rules and sets to provide nal numerical/
linguistic scores which include a convenient treatment of linguistic
uncertainty and subjectivity. However, the computation of water
quality index scores is clearly deterministic even within the fuzzy
method. A vector of water quality variables is given to the FIS, and
a unique water quality score is obtained. The challenge now is how
to deal with computation in non-deterministic real world scenarios.
Water quality variables collected in rivers are essentially stochastic,
and density probability functions may easily be computed. Then, the
key question is how to perform computations of water quality indexes
when sufcient data have been collected, and the statistics are dependable. Currently, the easiest way to deal with stochastic computations is
through Monte Carlo methods. Moreover, fuzzy alpha-cuts to deal with
uncertainty in inputs could also be considered (Kumar et al., 2009). In
this paper, we used Monte Carlo simulation to calculate the fuzzy
index to analyze historic and geographic trends in water quality. The
method was powerful because provided better water quality classication, and we observed graphically the consistency in fuzzy classication.
However, the use of combined probabilistic and fuzzy methods is still
under development, and a generalized theory of uncertainty is required
(Zadeh, 2005). Moreover, mathematical foundations about propagation
of probabilistic uncertainty through fuzzy systems may also require further research. Finally, we found that the method was powerful not only
by providing consistent histograms of defuzzied water quality scores
but also delivering the membership values to more than one water
quality class. The value of the membership function of the output
fuzzy sets resulted highly sensitive to input conditions. With this tool,
the decision makers may be able to relax the boundaries between two
or more likely water quality classes. Moreover, a consistent classication in water quality after stochastic simulations was observed which
showed that the fuzzy index was stable in providing appropriate classication. Finally, the use of fuzzy systems avoids using crisp values to
water quality classication which is the most important fact in applying
this methodology. With the Monte Carlo and FIS approach, the strongly
subjective character of the equivalence functions of traditional water
quality indexes is avoided, and the assessment is closer to human reasoning, becoming the technique very useful under many similar environmental assessment problems.
4. Conclusion
We have implemented stochastic simulation to a fuzzy water
quality index in order to improve the water quality assessment provided with deterministic indexes. The hybrid stochastic fuzzy method
combined the benets of Mont Carlo simulations with the advantages
of fuzzy inference. The proposed method updated the design of indexing
techniques to integrate water quality variables available to date. Nonparametric kernel density estimators resulted appropriate tools to
build empirical probability density functions from raw data since normal
and other parametric distributions did not t well the real data, especially
when number of data was limited. The Monte Carlo simulation improved
the results from point estimate of fuzzy water quality indexes since the
dispersion of the nal indexes was estimated. The water quality classication preserved the linguistic uncertainty of the subjective index and the
randomness from real measurements. The main advantage of the

27

proposed method is that membership to two or more classes is also possible which gives to decision makers a better conceptual assessment.
When the developed method was applied to the Cauca River, the results
for several years showed that water quality was possibly regular with a
membership degree of approximately 0.7, and possibly bad with a
membership degree of approximately 0.4. The index also predicted that
water quality decreased downstream. The results have complex origins,
since the river is plainly affected by the presence of towns and cities
without adequate treatment for wastewater. We observed that the environmental impact was not reduced downstream. Intense sugarcane agriculture and some industrial plants could also be responsible of surface
water pollution. An intensive environmental protection program from
regional and national government is suggested if ecosystem restoration
and biodiversity conservation is desired in the area.
Acknowledgments
The authors thank the Agencia Espaola de Cooperacin Internacional
para el Desarrollo (AECID) for nancial support (Projects D/026977/09,
and D/031370/10). We also thank the CVC Corporation for providing
water quality monitoring data.
References
Baudrit C, Guyonnet D, Dubois D. Joint propagation of variability and imprecision in
assessing the risk of groundwater contamination. J Contam Hydrol 2007;93:7284.
Beamonte-Cordoba E, Casino Martinez A, Veres-Ferrer E. Water quality indicators:
comparison of a probabilistic index and a general quality index. The case of the
Confederacin Hidrogrca del Jcar (Spain). Ecol Indic 2010;10:104954.
Brown RM, McClelland NI, Deininger RA, Tozer RG. A water quality index: do we dare?
Water Sew Works 1970;117:33943.
Cardona CM, Martin C, Salterain A, Castro A, San Martn D, Ayesa E. CALHIDRA 3.0
new software application for river water quality prediction based on RWQM1. Environ Model Softw 2011;26:9739.
Chen Z, Zhao L, Lee K. Environmental risk assessment of offshore produced water discharges using a hybrid fuzzy-stochastic modeling approach. Environ Modell
Softw 2010;25:78292.
Chowdhury S, Champagne P, McLellan PJ. Uncertainty characterization approaches for
risk assessment of DBPs in drinking water: a review. J Environ Manage 2009;90:
168091.
CVC Corporation. Estudio de la calidad del agua del ro cauca y sus principales
tributarios mediante la aplicacin de ndices de calidad y contaminacin. Project
Report 0168, Oct 2004. Available at: http://190.97.204.39/cvc/Mosaic/dpdf2/
Volumen10/1-ECARCpag1-158.pdf2004. (Accessed 1/9/2012).
Darbra RM, Eljarrat E, Barcelo D. How to measure uncertainties in environmental risk
assessment. Trends Anal Chem 2008;27:37785.
Faybishenko B. Fuzzy-probabilistic calculations of water-balance uncertainty. Stoch
Environ Res Risk A 2010;24:93952.
Ghiocel DM, Altmann J. Hybrid stochastic-neuro-fuzzy model-based system for in-ight
gas turbine engine diagnostics. In: Pusey HC, Pusey SC, Hobbs WR, editors. New frontiers in integrated diagnostics and prognosticsProceedings of the 55th meeting of the
Society for Machinery Failure Prevention Technology, Virginia; 2001.
Gottardo S, Semenzin E, Giove S, Zabeo A, Critto A, de Zwart D, et al. Integrated risk
assessment for WFD ecological status classication applied to Llobregat river
basin (Spain). Part Ifuzzy approach to aggregate biological indicators. Sci
Total Environ 2011;409:470112.
Guo P, Huang GH, Zhu H, Wang XL. A two-stage programming approach for water resources management under randomness and fuzziness. Environ Modell Softw
2010;25:157381.
Kentel E, Aral M. 2D Monte Carlo and Monte Carlo-fuzzy health risk assessment. Stoch
Environ Res Risk A 2005;19:8696.
Kumar V, Mari M, Schuhmacher M, Domingo JL. Partitioning total variance in risk
assessment: application to a municipal solid waste incinerator. Environ Modell
Softw 2009;24:24761.
Legay C, Rodriguez MJ, Sadiq R, Srodes JB, Levallois P, Proulx F. Spatial variations of
human health risk associated with exposure to chlorination by-products occurring
in drinking water. J Environ Manage 2011;92:892901.
Lermontov A, Yokoyama L, Lermontov M, Soares-Machado MA. River quality analysis
using fuzzy water quality index: Ribeira do Iguape river watershed, Brazil. Ecol
Indic 2009;9:118897.
Li H, Zhang K. Development of a fuzzy-stochastic nonlinear model to incorporate aleatoric and epistemic uncertainty. J Contam Hydrol 2010;111:1-12.
Li J, Huang GH, Zeng G, Maqsood I, Huang Y. An integrated fuzzy-stochastic modeling
approach for risk assessment of groundwater contamination. J Environ Manage
2007;82:17388.
Mahapatra SS, Nanda SK, Panigrahy BK. A cascaded fuzzy inference system for Indian
River water quality prediction. Adv Eng Softw 2011;42:78796.
Marchini A, Facchinetti T, Mistri M. F-IND: a framework to design fuzzy indices of environmental conditions. Ecol Indic 2009;9:48596.

28

W. Ocampo-Duque et al. / Environment International 52 (2013) 1728

Mari M, Nadal M, Schuhmacher M, Domingo JL. Exposure to heavy metals and PCDD/Fs
by the population living in the vicinity of a hazardous waste landll in Catalonia,
Spain: health risk assessment. Environ Int 2009;35:10349.
Mathworks. Product Documentation Matlab R2012a. Available at: http://www.mathworks.
com/help/2012. Accessed 29/08/2012.
Misha A. Estimating uncertainty in HSPF based water quality model: Application of
Monte-Carlo based techniques. PhD Thesis at Virginia Polytechnic Institute and
State University, USA, 2011.
Mller B, Beer M. Fuzzy randomness uncertainty in civil engineering and computational
mechanics. Berlin Heidelberg New York: Springer-Verlag; 2004.
Mller B, Graf W, Beer M, Sickert J. Fuzzy randomness towards a new modeling of
uncertainty. In: Mang AH, Rammerstorfer FG, Eberhardsteiner J, editors. The Fifth
World Congress on Computational Mechanics, Vienna; 2002.
Nikoo MR, Kerachian R, Malakpour-Estalaki S, Bashi-Azghadi SN, Azimi-Ghadikolaee
MM. A probabilistic water quality index for river water quality assessment: a
case study. Environ Monit Assess 2011;181:46578.
Ocampo-Duque W, Ferr-Huguet N, Domingo JL, Schuhmacher M. Assessing water
quality in rivers with fuzzy inference systems: a case study. Environ Int 2006;32:
73342.
Ocampo-Duque W, Schuhmacher M, Domingo JL. A neural-fuzzy approach to classify
the ecological status in surface waters. Environ Pollut 2007;148:63441.
Ocampo-Duque W, Juraske R, Kumar V, Nadal M, Domingo JL, Schuhmacher M. A concurrent
neuro-fuzzy inference system for screening the ecological risk in rivers. Environ Sci
Pollut Res 2012;19:98399.

Qin Z, Li W, Xiong X. Estimating wind speed probability distribution using kernel density
method. Electr Power Syst Res 2011;81:213946.
Ramaswami A, Milford JB, Small MJ. Integrated environmental modeling pollutant
transport, fate, and risk in the environment. John Wiley & Sons; 2005.
Rehana S, Mujumdar PP. An imprecise fuzzy risk approach for water quality management
of a river system. J Environ Manage 2009;90:365364.
Sadiq R, Tesfamariam S. Probability density functions based weights for ordered weighted
averaging (OWA) operators: an example of water quality indices. Eur J Oper Res
2007;182:135068.
Silverman BW. Density estimation for statistics and data analysis. London:
Chapman&Hall/CRC, ISBN: 0-412-24620-1; 1998.
Torres P, Cruz C, Patio P, Escobar JC, Prez A. Aplicacin de ndices de calidad de agua
ICA orientados al uso de la fuente para consumo humano. Ing Investig 2010;30:
8695.
Zadeh LA. Toward a generalized theory of uncertainty (GTU) an outline. Inf. Sci.
2005;172:1-40.
Zhang K, Li H, Achari G. Fuzzy-stochastic characterization of site uncertainty and variability
in groundwater ow and contaminant transport through a heterogeneous aquifer.
J Contam Hydrol 2009;106:7382.
Zhang X, Huang GH, Nie X. Robust stochastic fuzzy possibilistic programming for environmental decision making under uncertainty. Sci Total Environ 2009;408:
192201.