Water Conservation Science and Engineering (2023) 8:16

https://doi.org/10.1007/s41101-023-00190-3

RESEARCH

Classification and Prediction of Developed Water Quality Indexes

Using Soft Computing Tools
Deshbhushan Patil1 · Soumya Kar1 · Rajiv Gupta1

Received: 29 September 2022 / Revised: 25 January 2023 / Accepted: 6 April 2023

© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2023

Abstract
The water quality index (WQI) is a coherent method of expressing the state of the water quality, which has become com-
plicated due to subjectivity and ambiguity in the data. The proposed study aims to classify and predict the water quality for
potable water by applying soft computing techniques, namely fuzzy, adaptive-network-based fuzzy inference system (ANFIS)
and artificial neural network (ANN), along with a novel weight-integrated health hazard index (HI). Initially, the water clas-
sification was performed through the fuzzy and HI indexes and was subsequently used for predictive modelling through the
ANFIS and ANN on the 349 water samples for total dissolved solids (TDS), chloride ­(Cl−), hardness (as C ­ aCO3), fluoride
­(F−), nitrates ­(NO3−), iron (Fe), and copper (Cu) parameters collected from the municipality region of Jaipur, India. The
trained ANFIS model was proven satisfactory for fuzzy and HI indexes with the coefficient of determination ­(R2) value of
0.8413 and 0.996, respectively. In contrast, the ANN model failed to provide an adequate result for the fuzzy index with an
­R2 value of 0.5243 and a satisfactory result for HI with an ­R2 value of 0.839. The study’s novelty lies in predicting the water
quality index using ANFIS and ANN for the developed unique HI and fuzzy-based indexes. The results of this study prove
that ANFIS is a trustworthy and reliable approach for predicting WQI for potable water, which serves as a valuable guide
for decision-makers in the field of water resource management.

Keywords Fuzzy Index · Weight-integrated Health Hazard Index · ANFIS · ANN

Introduction characterise water quality through a single number ranging

from 0 to 100, where 100 indicates excellent water quality,
Only 1% of the earth’s surface is covered by freshwater, and a value of 0 indicates very poor water quality, unfit for
despite water constituting 25% of the total surface area [1, household utilization [2, 27].
20]. The mounting global population has resulted in sig- Several advancements in developing WQIs have hap-
nificant water stress situation and deteriorated water quality pened over time, beginning with the initial index developed
which needs to be addressed to enhance the sustainability of by Horton in 1965 )[16]. These advancements are largely
society [3, 19,41]. Continuous water quality monitoring at attributable to the adoption of soft technology applications,
the end user plays an essential role in preserving the sustain- such as data mining algorithms, artificial intelligence (AI),
able water management system. In such a case, water quality and fuzzy modelling systems, which address the escalat-
index (WQI) has proven significant in reducing the complex ing uncertainties that arise in the assessment of traditional
pre-processing of huge amounts of water quality data and WQI approaches [5, 10, 43]. Simultaneously, the prediction
facilitates efficient communication among the engineers, environment was developed by integrating multiple data
regulators, and stakeholders [40, 52]. A WQI integrates mining techniques, such as support vector machines, k near-
the measurements of various water quality parameters to est neighbours, decision trees, and naive Bayes, to classify
water quality based on the total pollution level [5, 8, 32].
Similarly, numerous studies have demonstrated that adopting
* Deshbhushan Patil
[email protected]; AI technologies such as knowledge-based systems, fuzzy
[email protected] logic, neural networks, and evolutionary algorithms has sig-
nificantly benefited the successful management of environ-
1
Department of Civil Engineering, Birla Institute mental concerns, [11, 18]. While dealing with non-linear,
of Technology and Science, Pilani 333031, India

13
Vol.:(0123456789)
 16 Page 2 of 16 Water Conservation Science and Engineering (2023) 8:16

unpredictable, ambiguous, and subjective data, Zadeh origi- using the qualitative criteria established by the WHO and
nally suggested fuzzy logic, which has since become one the Indian Council of Medical Research (ICMR) [25]. The
of the most widely used AI techniques and is suitable for standard test procedure was applied to analyse water qual-
creating environmental indexes [4, 58]. ity indicators such as pH, dissolved oxygen (DO), total dis-
On the other hand, predictive algorithms such as ANN solved solids (TDS), chloride (­ Cl−), hardness (as CaCO3),
have proven efficient to represent non-linear input–output fluoride ­(F−), nitrates (­ NO3−), iron (Fe), and copper (Cu).
relationships of complex data by providing substantial model Water quality classification was accomplished using fuzzy
flexibility [28, 35]. In some cases, for predicting water qual- and HI indexes using water testing results as input. The
ity based on its parameters, a multi-layer perceptron (MLP) classified water samples were further used as input to train
neural network has proven to be an efficient tool [11]. Now- the ANN and ANFIS models in the MATLAB tool. Out of
adays, an adaptive neuro-fuzzy inference system (ANFIS) 349 classified samples, 70%, 20%, and 10% were split and
is an approach that is becoming more popular for dealing exhausted for training, testing, and validating purposes while
with inadequately uncertain domains involved in forecast- performing ANN modelling. The detailed process performed
ing water quality [34]. ANFIS integrates a fuzzy inference to generate the water indexing is depicted in Fig. 1, and
system into an adaptive network architecture which provides the approach is explained in more detail in the subsequent
the combined advantages of ANN and fuzzy approaches by sections.
eliminating the core difficulty of specifying the member-
ship function constraints and creating a set of fuzzy if–then Development of a novel Weight‑Integrated Hazard
rules [50]. The learning capacity of ANNs has been utilised Index (HI)
in ANFIS to create autonomous fuzzy if–then rules and
optimise model parameters [36, 44]. In numerous studies, In the study, a novel hazard index was created by incorporat-
ANFIS has been successfully used to analyse water quality ing weights of various magnitudes into the health quotient
issues by categorising water quality status and comparing it (HQ) to calculate the risks brought on by each pollutant [40].
to the effectiveness of ANN [22, 56]. The comparison study The applied weighting method prevents erroneous risk esti-
employed a variety of membership function types, includ- mates resulting in moderate to significant underestimates.
ing generalised bell, Gaussian, trapezoidal, and triangular. Each drinking water quality parameter is given a certain
The key aim of the proposed methodology is to offer a weight (Wi) according to its relative significance to the total
comprehensive approach to analyse and predict water quality water quality. The allotted weight was divided by the total
within a specific region to improve decision-making. The sum of all parameter weights to produce relative weight-
proposed research work has been applied in two steps in the age (Wi), which measures the relative weight of grouped
following order: qualitative variables. Each criterion received a maximum
and minimum weight of 5 and 3, respectively, based on its
a) The initial stage is to categorise the water using Fuzzy importance. Expert recommendations and its severity were
and weight-integrated health hazard index (HI) indexes. used to determine the weights to be assigned to each vari-
b) The real-time water quality data and the deduced index able. The new HQ was termed the weighted hazard quotient
datasets have been employed to perform ANFIS and (HQW). The HQW, in conjunction with weighted relevance,
ANN modelling. attempts to turn the massive and challenging information of
raw water quality data into a more straightforward and logi-
After that, the developed models were analysed in terms cal form by using various categories of water quality reflect-
of performance through statistical errors. The following sec- ing the overall water quality state of the chosen region. The
tions explain the specific methodology and its implementa- summation of the HQW gives the weight-integrated hazard
tion in detail. index (HI).

Development of Fuzzy Index

Methodology
The fuzzy inference system (FIS) consists of membership
In the present study, area-based classification of water sam- functions, fuzzy set operations, and inference rules while
ples was done using fuzzy-based and novel HI indexes. The generating an index. Each chosen input set has a subset-
initial phase comprised real-time sampling and testing of based domain known as the universe of discourse articulated
water collected across the study region in accordance with using linguistic terms. If–then rules and fuzzy set operators
the World Health Organization (WHO). Following the col- define the relationships between the subsets of inputs and
lection of the samples, the water quality must be assessed outputs [58]. The FIS generates a non-linear mapping of an

13
Water Conservation Science and Engineering (2023) 8:16 Page 3 of 16 16

Fig. 1  Detailed workflow meth-

odology

input data vector into a scalar output using fuzzy rules. The shows a general model of a FIS that turns sharp inputs into
mapping process includes input/output membership func- crisp outputs.
tions, FL operators, fuzzy if–then rules, output set aggrega- Generally, the FIS works on four principles: fuzzifier, infer-
tion, and defuzzification. A multi-output FIS is comprised ence engine, rule base, and defuzzifier. In a fuzzifier, the input
of independent multi-input, single-output systems. Fig. 2 data identifies the extent to which the values belong to each
fuzzy set using the established membership function using
Eq. 1 [23, 26, 53]. The interference engine maps the input
data into fuzzy output sets and assesses the extent to which
the antecedent satisfies the rule. The rule base includes lin-
guistic rules offered by experts, which were extracted from
numerical data. Fuzzy operators provide a unique number rep-
resenting the antecedent’s outcome for a given rule when the
antecedent has many clauses where one or more rules could
activate simultaneously. Once the rules are set, the FIS can be
described as a system that maps an input vector to an output
vector [54]. The outputs from each rule are then combined.
During aggregation, the fuzzy sets representing the output of
each rule are combined into a single fuzzy set. After receiving
a fuzzy set with various output values, the defuzzifier returns
one number, converting the fuzzy set into a crisp number
(Eq. 2) [14, 38, 42].
{( ) }
mA(x) = x1 , 𝜇A (x) |x ∈ X 0 ≤ 𝜇A (x) = 1 (1)

∫ 𝜇(z)zdz
Z= (2)
𝜇(z)dz

Fig. 2  Fuzzy interference system

13
 16 Page 4 of 16 Water Conservation Science and Engineering (2023) 8:16

Adaptive Neuro‑Fuzzy Inference System (ANFIS) O1,i = 𝜇Ai (x) for i = 1, 2or (5)

Jang proposed an ANFIS algorithm, an integration of ANN

and FIS, in 1993 [6]. To map and simulate input–output
O1,i = 𝜇Bi−2 (y) for i = 3, 4 (6)
interactions, ANFIS employs both neural network learning where x and y represent input nodes, Ai and Bi represent
techniques and a fuzzy approach as a multi-layer network fuzzy sets, and (x) and (y) represent membership functions.
[13]. ANFIS uses a Takagi–Sugeno type fuzzy inference
system (FIS), with each fuzzy rule’s output being a linear Layer 2 This layer is referred to as the fire layer. It is the pre-
combination of input variables plus a constant term. ANFIS ceding component of a fuzzy rule. At this point, each node
applies two types of learning algorithms: backpropagation multiplies the input signals and passes the result to the node
and hybrid learning. The hybrid learning strategy combines of the following layer. As an example (Eq. 7).
backpropagation with the least squares method, whereas
backpropagation learning adheres to ANN methodology. wi = 𝜇Ai (x)x ∗ 𝜇Bi (y) for i = 1, 2 (7)
The least squares method (LSM) is applied for forward
where wi denotes the firing strength of a rule at the ith node.
passing and gradient descent is applied to backpropagation
[33]. The detailed architecture of ANFIS is shown in Fig. 3.
Layer 3 This phase offers normalised firing strengths. At this
Assuming that FIS has two input variables, x and y, and a
step, each node computes the firing strength of the i­th rule
single output variable, the first-order Sugeno fuzzy model
using the formula below (Eq. 8).
can provide the following if–then rules (Eqs. 3 and 4):
wi
R1. If x is A1 and y is B1, then f1 = p1 + q1 + r1 (3)
wi = w1 +w2
, i = 1, 2 (8)

Layer 4 The defuzzification method has been applied at this

R2. If x is A2 and y is B2, then f2 = p2 x + q2 y + r2 (4) layer, sometimes referred to as the normalization layer, to
where x and y are inputs for node i, A1, A2 and B1, B2 are produce the best possible input–output matching. The fol-
membership functions, and p1,p2, q1,q2, and r1,r2 are the lowing equation (Eq. 9) represents the process of this layer.
consequent parameters and f1 and f2 are the output function
Q4i = wi fi = wi (pi x + qi y + r) (9)
variables.
where wi is the previous output and pi, qi, and r are the sub-
Layer 1 In this layer, inputs have been fuzzified using an sequent parameters.
appropriate membership function. A node function can
represent each node (Eqs. 5 and 6). Membership functions Layer 5 The overall output from all the signals can be rep-
exist in various shapes, including general bell, triangular, resented in an output layer by Eq. 10.
trapezoidal, and Gaussian. The following equation generally
suggests a general bell shape feature in this layer.

Fig. 3  ANFIS architecture

13
Water Conservation Science and Engineering (2023) 8:16 Page 5 of 16 16

∑
� wi fi foundation in 1727 [48]. Jaipur is located at a latitude of
O5i = wi fi = ∑i (10) 26° 55′ north and a longitude of 75° 49′ east, with munici-
i wi
pal boundaries at latitudes 26° 46′ and 27° 01′ north and
75° 37′ and 76° 57′ east, respectively. The water supply in
Artificial Neural Network (ANN) Technique Jaipur has been in operation for over a century. The tap water
supply sourced initially from neighbouring open wells was
ANN is a reliable computational tool that effectively maps supplemented in 1918 by groundwater (16 enormous dia.
input and heterogeneous data in many scientific applications open wells on the Amanshah Nallah bed) and tap water sup-
[15]. Neurons are the smallest component of a neural network plies at various community areas. To counteract the grow-
and are typically organised in a network for prediction. The ing effects of water scarcity, the Bisalpur dam’s water sup-
ANN network is divided into three layers to estimate and fore- ply was extended in 2009. Currently, Jaipur City requires
cast a complex function: input, hidden, and output, as shown 462 MLD of water, with the Bisalpur dam providing 275
in Fig. 4. Trial and error was used to determine the number of MLD, the tube wells (1900 No.) providing 97 MLD, and
neurons in the hidden layer. The best weight vector is neces- the single-point tube wells providing 2 MLD (117 Nos.).
sary to generate the best approximation function during the The study area map and the data collection locations are
learning phase. shown in Fig. 5.
A dataset set is supplied into the input layer ­(I1, … ­In), and
the neuron output is estimated using Eq. 11. Collection of Water Samples
∑
n= wij xi + bi (11) During the initial part of the study, 349 water samples were
collected from several places throughout the city. Due to
where wij is the difference in weight between neurons in two accessibility issues, uniform sampling was impossible in
different layers, n is the neuron’s output, bi is the bias, and xi several urban locations. Water samples were obtained and
is the value of the neuron in the preceding layer. The sigmoid kept according to the guidelines given by USGS. For sam-
function was employed as the output transfer function, as ple collection, 1-Liter polypropylene bottles with a narrow
shown in Eq. 12. mouth were submerged in 5% HNO3 for 24 hrs. before being
1 regularly washed with deionised water. Before getting water
yi = ∑ , c1 > 0 (12) samples, the bottles were allowed to dry. Figure. 6 shows the
1 + e−c1 ( wij xi +bi )
water samples collected for the water quality index analysis.

Selection of Water Quality Parameters and Water Quality

Implementation Analysis

Study Area and Data Collection Water quality is influenced by physical parameters, inor-
ganic pollutant parameters, and biological parameters
Jaipur, popularly known as the Pink City, has been the state induced by various natural and anthropogenic activities
capital and the centre of Rajasthan’s economy since its [46]. To assess their health risk probability, it is necessary

Fig. 4  ANN architecture

13
 16 Page 6 of 16 Water Conservation Science and Engineering (2023) 8:16

Fig. 5  Study area map

to understand the water quality characteristics, standard safe water outlined by the WHO and the ICMR [25]. Water
permissible limits, and impacts on human health before quality indicators such as pH, dissolved oxygen (DO),
providing it to consumers to reduce the risk and ensure total dissolved solids (TDS), chloride ­(Cl−), hardness (as
CaCO3), fluoride ­(F−), nitrates ­(NO3−), iron (Fe), and cop-
per (Cu) were measured and analysed as per the standard
procedure. The tests used were titrimetric, colorimetric,
and atomic absorption spectrophotometer (AAS) as per
the standard guidelines proposed by the American Public
Health Association 2007 [37, 45].

Computing the Novel Weight‑Integrated Health

Hazard Index (HI)

Calculation of Daily Dose (ADAD/ADRD/ADD)

The minimum required or maximum authorised daily doses

for total hardness, chloride, and TDS are not specified in
terms of human health in the field of toxicological databases.
The average daily intake ADAD represents the daily intake
of a component. In contrast, ADRD is equivalent to the vari-
Fig. 6  Water sample collection ous nutritional requirements of the RDA (recommended

13
Water Conservation Science and Engineering (2023) 8:16 Page 7 of 16 16

dietary allowances) [7]. If the facts allow for the computa- (days). The average body weight (kg) is ABW; the minimum
tion of such a dose, both dosages are reduced to exposure essential concentration (MRC) is the lowest concentration
from a single environmental compartment (in this case, of an element at which no harm to health is known. AT is an
drinking water). According to the US EPA’s Integrated Risk abbreviation for averaging time (days), which refers to the
Information System (IRIS) database, the reference dose for time over which the dose is averaged [45]. In this study, the
some standard water quality parameters is represented in exposure dose is averaged throughout the exposure, so AT
Table 1 ; [12 ,24]. equals ED in value. Table 3 displays the input exposure data
for estimating human health risk.
Steps Followed for Computing HQiW and HI:
xxii. To assess non-cancer risk, the hazard quotient (HQi)
is calculated by comparing the ADD and the reference
i. Selecting the key variables for the calculation of the
dose (RfD) (mg/kg/day) by using Eqs. 17 and 18.
hazard index. TDS, CL, F, hardness, nitrates, Fe, and
Cu were chosen due to their adverse effects on daily ADD
life and human health.
HQi =
RfD (17)
ii. Initial weights (wi) are based on the values assigned
to each parameter’s relative importance and contri- ADAD
bution to each site’s total drinking water quality (via HQi = (18)
ADRD
expert opinions).
iii. The following equation (Eq. 13) was used to calculate
the relative weights of the considered parameters, and vi. Individual parametric weighted hazard quotient
the results are displayed in Table 2. (HQiW) uses Eq. 19.

wi HQi W = HQi ∗ Wi (19)

Wi = ∑n (13)
i=1 wi vii. A hazard index (HI) is calculated as the sum of
iv. The average daily dose (ADD) as the exposure meas- HQiWs shown in Eq. 20.
ured for Cu, Fe, F−, and NO3− is calculated using Eq. HI = ΣHQi W (20)
14 [17]
C × IR × ED
ADD = (14)
ABW × AT

Table 2  The relative weight of chemical parameters

Qualitative parameter wi Wi = wi/Σwi
For chloride, TDS, and total hardness, two values of ADAD
and ADRD are calculated using Eqs. 15 and 16. TDS 3 0.115
Total Hardness 3 0.115
C × IR × ED
ADAD = (15) Fluoride 3 0.115
ABW × AT
Nitrate 5 0.192
Iron 4 0.154
MRC × IR × ED Chloride 3 0.115
ADRD = (16)
ABW × AT Copper 5 0.192
where C represents the number of contaminants in a litre
of water (mg/L), IR represents the amount of water drunk
daily (L), and ED represents the amount of time exposed Table 3  Input exposure data for the assessment of human health risks
Parameter Value Unit

Table 1  Reference Dosages ABW—average body weight 76.74 Kg

Qualitative Reference dos-
parameter age (mg/kg-day) C—concentration of chemical ele- Site-specific mg/l
ments in water
Fluoride 0.06 IR—daily water intake 2.59 L
Nitrate 1.6 ED—duration of exposure 1 Year
Iron 0.7 AT—averaging exposure time 365 Days
Copper 0.04 EF—exposure frequency 365 Days/year

13
 16 Page 8 of 16 Water Conservation Science and Engineering (2023) 8:16

Table 4  Hazard index classification of gravity (COG) method (Eq. 2). After defuzzification,
Risk level HI Risk classification the values from MATLAB tools were used to calculate the
water quality based on established categories, as shown in
1 ≤ 0.1 Without risk Table 5. The most conventional scoring scale for water qual-
2 > 0.1 Low risk ity is between 1 and 100.
3 > 1.0 Medium risk
4 > 4.0 High risk ANFIS Prediction Modelling

ANFIS is a modelling approach used to resolve complex

The following scale, shown in Table 4, has been adopted non-linear systems. ANFIS provides learning via ANN and
to evaluate the degree of non-carcinogenic health risks con- high-level reasoning via a fuzzy model as a result. The opti-
sumers may experience due to routinely consuming toxic mum ANFIS model can be developed by selecting the right
water. amount of membership functions, fuzzification functions,
the type of output function, the optimization method, the
Computing the Fuzzy Index training algorithm, and the number of epochs. The ANFIS
model has been applied individually for both calculated
For conducting the water analysis, the measurement water fuzzy and HI indexes. To attain the best simulation results
quality variables were considered: pH, dissolved oxygen with a minor error, the ANFIS editor in MATLAB allows
(DO), total dissolved solids (TDS), chloride (­ Cl−), hardness modification of the consideration. When executing for the
(as CaCO3), fluoride ­(F−), nitrates ­(NO3−), iron (Fe), and fuzzy inputs, the number of membership functions was kept
copper (Cu). The study shows that pH and DO values for all similar to the categorization of water quality parameter. In
collected samples are within the BIS-established standard case of performing simulation for the HI index, the stand-
limits, absolving them from incorporating in the fuzzy index. ard values of the membership function as 3 were kept for
The experimental value range and Eq. 21 were initially used all the input water quality parameters, and modelling was
to define the triangle membership function for each parame- performed for the triangular, trapezoidal, general bell, and
ter. In the proposed study, the linguistic terms very low, low, Gaussian membership functions. The input data is split into
moderate, high, and very high were used to define the fuzzy three categories: training (70%), testing (20%), and valida-
set for total dissolved solids (TDS), chloride ­(Cl−), hardness tion (10%). The detailed characterization of the analysis for
(as CaCO3), and fluoride (­ F−), while low, medium, and high both indexes is shown in Table 6.
were used to define the fuzzy set for nitrates ­(NO3−) and iron
(Fe). For copper (Cu), low and high terms were applied, and ANN Prediction Modelling
each input parameter was sorted according to the member-
ship functions into one of the three fuzzy sets. An artificial neural network (ANN), a black-box technique,
was able to model and learn complicated non-linear relation-
⎧ 0, �x < aorc < x ⎫ ships between the input and output. Input and target data for
⎪ (a−x) ⎪ this study were made up of 70% dataset and remaining 30%
f (x;a, b, c) = ⎨ (a−b) , �a ≤ x ≤ b ⎬ (21)
⎪ (c−x) , b ≤ x ≤ c ⎪ dataset used for testing. For both fuzzy and HI indexes, the
⎩ (c−b) ⎭ classified values are given as objectives during implementa-
tion, and the predicted values are taken as the output from the
where the parameters {a, b, c} with (a < b < c) specify the
prediction model. The input consists of the values for the water
x-axis coordinates of the triangle membership function, and
quality parameter. A single hidden layer in ANN is gener-
x represents the site-specific experimental value.
ally sufficient to handle any function approximation problem.
The threshold values have been specified as the cut-
off values for choosing the ranges for each variable. Each
parameter’s below-average levels are categorised as low,
while its above-average values are categorised as high. Table 5  Water quality Range Description
Additionally, the rules were developed using professional classification
knowledge. The Mamdani inference system was used in a 0–44 Poor
study to incorporate the knowledge of the experts and apply 45–64 Marginal
it to create a fuzzy index [31, 47, 59]. The “fuzzy logic 65–79 Fair
toolbox” in MATLAB version 7.9.0. 2.7 was used for all 80–94 Good
calculations. The outputs were defuzzified using the centre 95–100 Excellent

13
Water Conservation Science and Engineering (2023) 8:16 Page 9 of 16 16

Table 6  ANFIS prediction modelling characteristics robust and fault tolerant. An extensive data set is a precursor
Indexes/memberships 2
R values for fuzzy 2
R values to developing predictive models for any location. Most of the
index for HI index literature highlights the increasing usage of classification
and predicting models in the river and groundwater quality
Triangular 0.8413 0.99647
assessments [21, 29, 49]. The outcome of the classification
Trapezoidal 0.7832 0.20787
and prediction models has been provided below.
General Bell 0.8284 0.96544
Gaussian 0.8099 0.97770
Water Quality Testing and Classification

The proposed study’s water quality parameters were

Table 7  Determination of the number of neurons selected based on their negative impacts on individuals. In
Number of neurons Number of neurons Overall ­R2 value
accordance with the study, every sample were examined to
meet the requirements for pH and DO in drinking water,
2*NI 14 0.9531 with reported values that ranged from 6.5 to 8.5 and 4 to 6,
NI + NO 8 0.73579 respectively. With a maximum observed value of 331 mg/l,
0.75 * NI 5 0.95099 nitrates causing methemoglobinemia in children were nota-
2* NI + 1 15 0.95105 bly high, and around 60% of the water samples surpassed
NI 7 0.95236 the allowable threshold of 45 mg/l. The two heavy metals
(NI + NO)/2 4 00.94487 analysed for the study area were Fe and Cu; copper levels
ranged from 0.1 to 2 mg/l, and an incredible 75% of samples
exceeded the allowed limits. Despite the iron levels being
However, some problems may be easier to solve with more between 0.1 and 1 mg/l, 51.7% of the samples were above
than one hidden layer, especially two hidden layers [39]. Typi- the allowed limit of 0.3 mg/l. The determined chloride lev-
cally, a trial-and-error process was used to control the number els in the samples were 23 to 3100 mg/l on average, and a
of neurons in a buried layer [57]. To determine the number large percentage of them (72.86%) confirmed the stand-
of neurons, the specific empirical criteria by considering the ard value of 250 mg/l. With a maximum recorded value
number of input (NI) and the output (NO) is represented in of 6840 mg/l, 11.4% of the collected samples had TDS
Table 7. The final model has a 7:14:1 architecture, mean- values over the 2000 mg/l allowable barrier, and 60.57%
ing there are seven inputs (WQ parameters), fourteen hidden of samples had TDS values between 500 and 200 mg/l.
nodes, and one output (WQI), as shown in Fig. 4 A feed-for- About 27.71% had TDS below the acceptable level of
ward neural backpropagation network using Levenberg–Mar- 500 mg/l. With recorded readings ranging between 30
quardt (TRAINLM) training algorithm and the activation and 1840 mg/l, averaging 303.67 mg/l, hardness is mostly
function tan sigmoid for the hidden layer and log sigmoid for caused by calcium, magnesium, and other chemicals. Most
the output layer has been employed. samples, or about 62.6%, were categorised as very hard.
R2 (coefficient of determination) is typically used to eval- Only 1.71% of the samples were classified as soft, whereas
uate the trained network’s accuracy. The degree to which 20.29% and 15.14% fell into the medium and hard catego-
the independent factors considered adequately explain the ries. According to standards, ­F− concentrations must not
measured dependent variable is indicated by the coefficient exceed 0.5 mg/l. However, most samples (57.14%) had
of determination. A better predictive relationship is indi- ­F− readings < 0.5 mg/l. A total of 21.71% and 15.43% of
cated by a higher ­R2 score [9]. the samples fell between 0.5 and 1 mg/l and 1 and 1.5 mg/l,
respectively (Fig. 7). The threshold value of 1.5 mg/l was
exceeded in about 5.43% of the observed samples.
Results and Discussion The classification performed using the HI shows that
165, 177, and 7 samples fall in the low, medium, and high-
In proposed study, developed fuzzy index aims to provide risk zones, revealing that 52% of the city area comes under
a concise method for expressing ambient water quality and the medium and high-risk zones, respectively, proving
the health hazard index aims to correlate water quality with detrimental to human health. The classification performed
human health. In the past, ANN and ANFIS have been using the fuzzy logic shows that the 67, 154, 41, 84, and 4
proven an efficient in modelling the complex and unknown samples fell into the poor, marginal, fair, good, and excel-
non-linear relationship between raw water quality data and lent categories, respectively. About 63.7% of the samples
indexes without needing a numerical equation. Furthermore, belonged to the marginal and poor category and demanded
the AI models are adaptable to changes in the dataset, uni- effective water treatment. Fig. 8 shows the classification of
versal approximation, and mapping capacity, making them water samples based on fuzzy and HI indexes.

13
 16 Page 10 of 16 Water Conservation Science and Engineering (2023) 8:16

Fig. 7  Water quality parameter range

ANFIS Prediction Modelling The model predictive quality of ANFIS for both indexes
was compared against their calculated values. The ANFIS
The ANFIS model consisted of 7 independent inputs and one models have good predictions; however, statistically, ANFIS
dependent output, outlining a site-specific prediction algo- for the fuzzy index (R2 = 0.84) and ANFIS for the HI index
rithm. The analysis was performed for the triangular, trap- (R2 = 0.996), as shown in Fig. 9.
ezoidal, general bell, and Gaussian membership functions.
The triangular membership function for both the fuzzy and ANN Prediction Modelling
HI indexes gives the lower RMSE values and is considered
the optimal topology and used for further analysis. A signifi- Several empirical formulae for the number of neurons in the
cant novelty of the study lies in maintaining congruency in hidden layer were employed to deduce the network architec-
the membership functions for both the fuzzy classification ture, wherein 14 neurons, i.e. (2* Number of Inputs), pro-
and ANFIS modelling, primarily for the fuzzy index, thereby vided the best overall ­R2 value as 0.9531. The final network,
achieving improved R ­ 2 values. In contrast, the HI index was which had a 7–14-2 architecture, was utilised to forecast the
executed through the established membership functions results for both indexes. The datasets of indexes developed
and rules developed by the ANFIS editor. Table 8 provides from the real-time data were used to feed the optimised net-
detailed characteristics of the ANFIS model. work to test and validate the trained model. Figure 10 shows

13
Water Conservation Science and Engineering (2023) 8:16 Page 11 of 16 16

Fig. 8  Fuzzy and HI water quality classification

Table 8  ANFIS model properties for fuzzy and HI indexes 7.27] to [− 14.379, 10.21], the residuals observed for the
Properties Fuzzy HI
fuzzy forecast range from [− 36.26, 58.24] to [− 28.40,
35.83] with both ANFIS and ANN models. As a result of
Number of nodes 22,566 4426 these findings, the ANFIS model offers a better prediction
Number of linear parameters 11,250 2187 model and better characterises the developed site-specific
Number of non-linear parameters 11,362 63 water quality indexes.
Number of training data pairs 245 245 Figure. 12 shows the MAE, RMSE, ­R2, and R values for
Number of checking data pairs 0 0 each predictive modelling applied for both indexes, which
Number of fuzzy rules 11,250 2187 shows that the ANFIS modelling provides the best results
for both the fuzzy and HI indexes.

the comparison between calculated values for predicted val-

ues for the indexes. Figure. 10 also demonstrates two lines: Conclusion
the perfect fit y = X (predicted data = experimental data), and
the other is the best fit indicated by a solid line. The reported Approximately 22% of the water samples were collected
­R2 value for the HI index is 0.839, which indicates that the from industrial hand pump and tube wells, indicating poor
input variables included in prediction modelling are suffi- water quality and a high level of risk. Initially, water qual-
cient for predicting the HI index. In contrast, the R ­ 2 value ity classification was done through fuzzy and novel health
for the fuzzy index is 0.524, which indicates that the input hazard (HI) indexes. According to the analysis, only 1.2% of
variables included in prediction modelling are insufficient the collected samples were categorised as good, while 19%
for predicting the fuzzy index with ANN. were categorised as poor. Simultaneously, the health hazard
The residual plots are shown in Fig. 11 which help to index values designate 47% of the collected water samples as
comprehend deviation of predicted from each model from low risk, while 2% are classified as high risk. Most city areas
the experimental data. While the residuals for HI from the have poor water quality, which is closely associated with a
developed ANFIS and ANN models range from [− 7.23, “medium” level of health risk due to anthropogenic activity

13
 16 Page 12 of 16 Water Conservation Science and Engineering (2023) 8:16

Fig. 9  Predicted outcome from ANN for (a) fuzzy index and (b) HI index

Fig. 10  The predicted outcome from ANN for the (a) fuzzy index and (b) HI index

such as textile dying and printing industries, which employs optimum results with an R ­ 2 value of 0.8413 with triangular
a variety of chemicals and azo dyes. Furthermore, ANFIS membership. While performing the ANFIS modelling for
and ANN models were developed on the generated indexes the HI index, the standard values of the membership func-
values to enable easy prediction and decision-making. While tion with the triangular membership provide the optimum
performing the ANFIS modelling for the fuzzy index, the results with the ­R2 value of 0.996. The ­R2 values for the pre-
modification in a few membership function provides the dicted values show that the incorporated inputs are sufficient

13
Water Conservation Science and Engineering (2023) 8:16 Page 13 of 16 16

Fig. 11  Residuals observed for

the predicted outcome from
ANFIS and ANN model for (a)
HI index and (b) Fuzzy index

Fig. 12  Statistical error comparison for predictive models

13
 16 Page 14 of 16 Water Conservation Science and Engineering (2023) 8:16

to create the prediction modelling for both fuzzy and HI Declarations

indexes through ANFIS. Additionally, multiple neuron
combinations were created while implementing the ANN, Ethics Approval and Consent to Participate Not applicable.
and the 14 neurons in the hidden layer provide the optimum Consent for Publication Not applicable.
results. The observed R ­ 2 values for the fuzzy and HI index
were 0.524 and 0.998, respectively. From the calculated R ­ 2 Competing Interests The authors declare no competing interests.
values, the ANN prediction modelling did not perform opti-
mally for the fuzzy index.
The study concluded that ANFIS proves trustworthy and
accurate approach for predicting WQI for potable water. References
Using these methods automates the calculation of WQI and
1. Aani Al, Saif TB, Hasan SW, Hilal N (2019) Can machine lan-
considerably decreases computation time; thus, which can guage and artificial intelligence revolutionize process automation
be utilised in any aquatic environment across the world as an for water treatment and desalination? Desalination 458:84–96
effective technique for forecasting water quality, particularly 2. Abtahi M et al (2015) A modified drinking water quality index
(DWQI) for assessing drinking source water quality in rural com-
for specific water resource problems that demand high preci-
munities of Khuzestan Province, Iran. Ecol Ind 53:283–291
sion forecasting and conventional methods cannot provide 3. Adelodun B et al (2021) Assessment of socioeconomic inequality
high accuracy. based on virus-contaminated water usage in developing countries:
a review. Environ Res 192:110309
4. Asgari G, Komijani E, Seid A, Mohammadi, and Khazaei M
Future Studies (2021) “Assessment the quality of bottled drinking water through
Mamdani fuzzy water quality index”. Water Resources Manag
35(15):5431–52.https://​doi.​org/​10.​1007/​s11269-​021-​03013-z
Despite advantages, there are still several obstacles to 5. Babbar R and Babbar S (2017) “Predicting river water quality
overcome before applying machine learning techniques to index using data mining techniques”. Environ Earth Sci 76(14):1–
evaluate water quality, which primarily requires a large 15. https://​link.​sprin​ger.​com/​artic​le/​10.​1007/​s12665-​017-​6845-9.
quantity of data and complicated phenomena governing Accessed 22 Aug 2022
6. Buragohain M, Mahanta C (2008) A novel approach for ANFIS
real-world water treatment and management systems. modelling based on full factorial design. Appl Soft Comput
To address the obstacles, future research and engineer- 8(1):609–625
ing processes should consider the following factors: (1) 7. Tillotson V, Tillotson GHR, Sachdev G, Sachdev V (2002) Build-
More modern sensors, particularly soft sensors, should be ing Jaipur: The making of an Indian City. United Kingdom: Uni-
versity of Chicago Press. Cabral Pinto MMS, Ordens CM, Con-
developed and implemented in water quality monitoring to desso de Melo MT, Inácio M, Almeida A, Pinto E, & Ferreira da
collect data with sufficient precision to permit the applica- Silva EA (2020) An inter-disciplinary approach to evaluate human
tion of machine learning techniques. (2) The practicality health risks due to long-term exposure to contaminated groundwa-
and dependability of the algorithms must be enhanced, and ter near a chemical complex. Expo Health 12(2):199–214. https://​
doi.​org/​10.​1007/​S12403-​019-​00305-Z/​FIGUR​ES/6. Accessed 20
more universal algorithms and models must be developed Jul 2022
to meet the needs of water treatment and management. (3) 8. Chang NB, Chen HW, Ning SK (2001) Identification of river
Multidisciplinary talent should be trained to build more water quality using the fuzzy synthetic evaluation approach. J
advanced machine learning algorithms and implement Environ Manage 63(3):293–305
9. Chang NB, Yeh SC, Wu GC (2010) Stability analysis of grey com-
them in engineering. promise programming and its application to watershed land-use
planning. Int J Syst Sci 30(6):571–589. https://​doi.​org/​10.​1080/​
00207​72992​92092
Author Contribution Deshbhushan Patil: Conceptualization, Meth- 10. Chau K (2006) A review on integration of artificial intelligence
odology, Formal analysis, Visualization, Writing original draft, and into water quality modelling. Mar Pollut Bull 52(7):726–733
Paper administration; Soumya Kar: Conceptualization, Article writing, 11. Chau KW (2006) A review on integration of artificial intel-
Formal analysis, Review and Editing; Rajiv Gupta: Conceptualization, ligence into water quality modelling. Marine pollution bulletin
Guidance, Supervision, and Motivation. 52(7):726-733
12. Environmental Protection Agency (n.d.) 40 CFR Parts 141 and
Funding The authors are grateful to the Department of Science & 142 National Primary Drinking Water Regulations for Lead and
Technology, New Delhi, for providing financial assistance under the Copper: Short-Term Regulatory Revisions and Clarifications.
project vide grant no. DST/TMD/EWO/WTI/2K19/UWS-04(C1) www.​regul​ations.​gov. Accessed 20 Jul 2022
titled Structured Dialogues for Sustainable Urban Water Management 13. Gallo G, Perfilieva I, Spagnuolo M, & Spinello S (1999) Geo-
(SDSUWM). graphical data analysis via mountain function. Int J Intell Syst
14(4):359–373
Data Availability Some or all data, models or code generated or used 14. Gharibi H et al (2012) Development of a dairy cattle drinking
during the study are proprietary or confidential and may only be pro- water quality index (DCWQI) based on fuzzy inference systems.
vided with restrictions. Ecol Ind 20:228–237

13
Water Conservation Science and Engineering (2023) 8:16 Page 15 of 16 16

15. Gupta R, Singh AN, Singhal A (2019) Application of ANN for 34. Mohammad Khan F, Gupta R, Sekhri S (2001) A novel PCA-FA-
water quality index. Int J Mach Learn Comput 9(5):688–693 ANN based hybrid model for prediction of fluoride. Stoch Env Res
16. Horton RK (1965) An index-number system for rating water Risk Assess 35:2022.https://d​ oi.o​ rg/1​ 0.1​ 007/s​ 00477-0​ 21-0​ 2001-4
quality.&nbsp;J Water Pollut Control Fed 37(3):300–306 35. Nasiri F, Maqsood I, Huang G, Fuller N (2007) Water qual-
17. Hu G et al (2022) Integrated probabilistic-fuzzy synthetic evalu- ity index: A fuzzy river-pollution decision support expert
ation of drinking water quality in rural and remote communities. system.&nbsp;J Water Resour 133(2):95–105
J Environ Manag 301(2021):113937. https://​doi.​org/​10.​1016/j.​ 36. Nayak PC, Sudheer KP, Rangan DM, Ramasastri KS (2004) A
jenvm​an.​2021.​113937 neuro-fuzzy computing technique for modeling hydrological time
18. Icaga Y (2007) Fuzzy evaluation of water quality classification. series. J Hydrol 291(1–2):52–66
Ecol Ind 7(3):710–718 37. Nelson C, Lurie N, Wasserman J, and Zakowski S (2007) Con-
19. Ighalo JO, Adeniyi AG, and Marques G (2021a) “Internet of ceptualizing and defining public health emergency preparedness.
things for water quality monitoring and assessment: a compre- Am J Public Health 97 (Suppl 1). https://​ajph.​aphap​ublic​ations.​
hensive review”. Studies Comput Intell 912:245–59. https://​link.​ org/​doi/​10.​2105/​AJPH.​2007.​114496. Accessed 20 Jul 2022
sprin​ger.​com/​chapt​er/​10.​1007/​978-3-​030-​51920-9_​13. Accessed 38. Ocampo-Duque W et al (2013) Water quality analysis in rivers with
22 Aug 2022 non-parametric probability distributions and fuzzy inference systems:
20. Ighalo JO, Adeniyi AG, and Marques G (2021b) “Artificial intelli- application to the Cauca River, Colombia. Environ Int 52:17–28
gence for surface water quality monitoring and assessment: a sys- 39. Partovi FY, Anandarajan M (2002) Classifying Inventory using an
tematic literature analysis”. Model Earth Syst Environ 7(2):669– artificial neural network approach. Comput Ind Eng 41(4):389–404
81. https://l​ ink.s​ pring​ er.c​ om/a​ rticl​ e/1​ 0.1​ 007/s​ 40808-0​ 20-0​ 1041-z. 40. Patil D, Kar S, Shastri V, Gupta R (2023a) Qualitative and health
Accessed 18 Aug 2022 risk assessment of water using a novel weight-integrated health
21. Jagaba AH et al (2020) Water quality hazard assessment for hand hazard and fuzzy-derived indices. Sustainable Water Resources
dug wells in Rafin Zurfi, Bauchi State, Nigeria. Ain Shams Eng J Management 9(2):55
11(4):983–999 41. Patil D, Kumar G, Kumar A, Gupta R (2023b) A systematic basin-
22. Jang JSR (1993) ANFIS: adaptive-network-based fuzzy inference wide approach for locating and assessing volumetric potential of
system. IEEE Trans Syst Man Cybern 23(3):665–685 rainwater harvesting sites in the urban area.&nbsp;Environ Sci
23. Jang JSR, and Sun C-T (1995) “Neuro-fuzzy modeling and con- Pollut Res&nbsp;30(6):14707–14721
trol”. Proc IEEE 83(3):378–406 42. Patki VK, Shrihari S, Manu B, Deka PC (2015) Fuzzy system
24. Javed M, and Usmani N (2016) “Accumulation of heavy metals modeling for forecasting water quality index in municipal distribu-
and human health risk assessment via the consumption of fresh- tion system. Urban Water J 12(2):89–110. https://d​ oi.o​ rg/1​ 0.1​ 080/​
water fish Mastacembelus armatus inhabiting, thermal power plant 15730​62X.​2013.​820333
effluent loaded canal”. SpringerPlus 5(1):1–8. https://s​ pring​ erplu​ s.​ 43. Pham QB et al. (2021) Application of soft computing to predict
spring​ erope​ n.c​ om/a​ rticl​ es/1​ 0.1​ 186/s​ 40064-0​ 16-2​ 471-3. Accessed water quality in wetland. Environ Sci Pollut Res 28(1):185–200.
20 Jul 2022 https://​link.​sprin​ger.​com/​artic​le/​10.​1007/​s11356-​020-​10344-8.
25. Jinturkar AM, Deshmukh SS, Agarkar SV, Chavhan GR (2010) Accessed 18 Aug 2022
Determination of water quality index by fuzzy logic approach: 44. RadFard M et al (2019) Protocol for the estimation of drinking
a case of ground water in an Indian town. Water Sci Technol water quality index (DWQI) in water resources: artificial neural
61(8):1987–1994 network (ANFIS) and Arc-Gis. MethodsX 6:1021–1029
26. Karr CL, Gentry EJ (1993) Fuzzy control of Ph using genetic 45. Rapant S et al. (2020) Proposal of new health risk assessment
algorithms. IEEE Trans Fuzzy Syst 1(1):46–53 method for deficient essential elements in drinking water—case
27. Kar S, Gupta R (2023) Fluoride toxicity in Rajasthan, India: study of the Slovak Republic. Int J Environ Res Public Health
Human health risk assessment, low-cost water filter prepara- 17(16):5915. https://w ​ ww.m ​ dpi.c​ om/1​ 660-4​ 601/1​ 7/1​ 6/5​ 915/h​ tm.
tion, and contaminant remediation.&nbsp;Water Conserv Sci Accessed 20 Jul 2022
Eng&nbsp;8(1):3 46. Robinson DT et al. (2018) Assessing the impact of a risk-based
28. Khan Y, Chai SS (2017) Ensemble of ANN and ANFIS for water intervention on piped water quality in rural communities: the
quality prediction and analysis - a data driven approach. J Tel- case of mid-western Nepal. Int J Environ Res Public Health
ecommun, Electron Comput Eng 9(2–9):117–122 15(8):1616. https://​www.​mdpi.​com/​1660-​4601/​15/8/​1616/​htm.
29. Khandare AL et al. (2020) “Health risk assessment of heavy met- Accessed 21 Jul 2022
als and strontium in groundwater used for drinking and cooking 47. Ross TJ (2005) Fuzzy logic with engineering applications. Ger-
in 58 villages of Prakasam District, Andhra Pradesh, India”. Envi- many: Wiley.Gharibi, H., Mahvi AH, Nabizadeh R, Arabalibeik
ron Geochem Health 42(11):3675–3701. https://l​ ink.s​ pring​ er.c​ om/​ H, Yunesian M, & Sowlat MH (2012) A novel approach in water
artic​le/​10.​1007/​s10653-​020-​00596-1. Accessed 20 Jul 2022 quality assessment based on fuzzy logic.&nbsp; J Environ Manage
30. Kuncheva LI (2001) Fuzzy logic with engineering applications, 112:87–95. https://​doi.​org/​10.​1016/J.​JENVM​AN.​2012.​07.​007
Timothy J. Ross, (Ed.); McGraw Hill, New York, 1995, pp. 592. 48. Sachdev V, Tillotson GHR, Sachdev G, Tillotson V (2002) Build-
Neurocomputing 41(1–4):187–187 ing Jaipur: the making of an Indian city. Reaktion Books
31. Le T, Hassan F, Le C, and Jeong HD (2019) “Understanding 49. Saha P, Paul B (2018) Assessment of heavy metal toxicity related
dynamic data interaction between civil integrated management with human health risk in the surface water of an industrialized
technologies: a review of use cases and enabling techniques”. area by a novel technique. Human Ecol Risk 25(4):966–87. https://​
https://​doi.​org/​10.​1080/​15623​599.​2019.​16788​63 doi.​org/​10.​1080/​10807​039.​2018.​14585​95
32. Liu Li et al (2010) Using Fuzzy theory and information entropy 50. Shwetank S, Chaudhary JK (2022) Hybridization of ANFIS and
for water quality assessment in Three Gorges region, China. fuzzy logic for groundwater quality assessment. Groundwater Sus-
Expert Syst Appl 37(3):2517–2521 tain Dev 18(1):100777. https://d​ oi.o​ rg/1​ 0.1​ 016/j.g​ sd.2​ 022.1​ 00777
33. Loganathan C, Girija KV (2013) Cancer classification using adap- 51. Tillotson V, Sachdev G, Tillotson GHR, Sachdev V (2002) Build-
tive neuro fuzzy inference system with Runge Kutta learning. Int ing Jaipur: the making of an Indian city. University of Chicago
J Comput Appl 79(4):46–50 Press, United Kingdom

13
 16 Page 16 of 16 Water Conservation Science and Engineering (2023) 8:16

52. Tiwari S, Babbar R, Kaur G (2018) Performance evaluation of two 57. Yassine D (2001) Neural network model for preformed-foam cel-
ANFIS models for predicting water quality index of River Satluj lular concrete. https://​www.​resea​rchga​te.​net/​publi​cation/​27957​
(India). Advances in Civil Engineering, 2018 3475. Accessed 22 Aug 2022
53. Turksen IB (1991) Measurement of membership functions and 58. Zadeh LA (1996) Fuzzy sets. 394–432. https://​doi.​org/​10.​1142/​
their acquisition. Fuzzy Sets Syst 40(1):5–38 97898​14261​302_​0021
54. Vyas V, Singh AP, and Srivastava A (2019) A decision making 59. Zadeh LA, Klir GJ, and Yuan B (1996) Fuzzy sets, fuzzy logic,
framework for condition evaluation of airfield pavements using and fuzzy systems. 6. https://w
​ ww.w​ orlds​ cient​ ific.c​ om/w
​ orlds​ cibo​
non-destructive testing”. Airfield Highway Pavements 2019: Innov oks/​10.​1142/​2895. Accessed 20 Jul 2022
Sustain Highway Airfield Pavement Technol - Selected Papers
from the Int Airfield and Highway Pavements Conf 2019:343–53 Publisher's Note Springer Nature remains neutral with regard to
55. Jacobs HL, Gabrielson IN, Horton RK, Lyon WA, Hubbar EC, jurisdictional claims in published maps and institutional affiliations.
& McCallumGE (1965) Water quality criteria-stream vs. efflu-
ent standards on JSTOR. https://​www.​jstor.​org/​stable/​25035​251#​ Springer Nature or its licensor (e.g. a society or other partner) holds
metad​ata_​info_​tab_​conte​nts. Accessed 22 Aug 2022 exclusive rights to this article under a publishing agreement with the
56. Yan H, Zou Z, Wang H (2010) Adaptive neuro fuzzy inference author(s) or other rightsholder(s); author self-archiving of the accepted
system for classification of water quality status. J Environ Sci manuscript version of this article is solely governed by the terms of
22(12):1891–1896 such publishing agreement and applicable law.

13