Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

Contents lists available at ScienceDirect

Journal of King Saud University –

Computer and Information Sciences
journal homepage: www.sciencedirect.com

Water quality prediction and classification based on principal

component regression and gradient boosting classifier approach
Md. Saikat Islam Khan a,d, Nazrul Islam b,d,⇑, Jia Uddin c, Sifatul Islam a,d, Mostofa Kamal Nasir a,d
a
Department of Computer Science and Engineering, Santosh, Tangail-1902, Bangladesh
b
Department of Information and Communication and Technology, Santosh, Tangail-1902, Bangladesh
c
Department of Technology Studies, Endicott College, Woosong University, Daejeon, South Korea
d
Mawlana Bhashani Science and Technology University, Santosh, Tangail 1902, Bangladesh

a r t i c l e i n f o a b s t r a c t

Article history: Estimating water quality has been one of the significant challenges faced by the world in recent decades.
Received 29 January 2021 This paper presents a water quality prediction model utilizing the principal component regression tech-
Revised 11 May 2021 nique. Firstly, the water quality index (WQI) is calculated using the weighted arithmetic index method.
Accepted 3 June 2021
Secondly, the principal component analysis (PCA) is applied to the dataset, and the most dominant WQI
Available online xxxx
parameters have been extracted. Thirdly, to predict the WQI, different regression algorithms are used to
the PCA output. Finally, the Gradient Boosting Classifier is utilized to classify the water quality status. The
Keywords:
proposed system is experimentally evaluated on a Gulshan Lake-related dataset. The results demonstrate
Water quality index
Principal component regression
95% prediction accuracy for the principal component regression method and 100% classification accuracy
Classification algorithm for the Gradient Boosting Classifier method, which show credible performance compared with the state-
Boxplot analysis of-art models.
Ó 2021 The Authors. Production and hosting by Elsevier B.V. on behalf of King Saud University. This is an
open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction 2020; Moon et al., 2004). In Bangladesh, many urban areas are
affected by water contamination due to unplanned urbanization
There are no living species on earth who can survive without and industrialization. In 2015, urban pollution caused Bangladesh
water. The flow of water from rivers and lakes contributes explic- to lose $6.5 billion, which was 3.4% of GDP, according to the world
itly or implicitly to both human well-being and the fisheries indus- bank report (World Bank Report, 2018), where Dhaka lost $1.44
try (Kar, 2019; Kar, 2013). However, water is frequently polluted billion, which was the 0.72% of the GDP.
because the industry has been growing every year on the back of Therefore, surveillance of water quality is mandatory. Though
spiralling demand, and hazardous waste is discharged into the riv- water quality can be tested using traditional techniques such as
ers and lakes by those industries. Millions of people died every collecting the water specimens manually and then analyzed it in
year, untold losses of income, agricultural land deteriorated due a laboratory (Wu and Liu, 2012). But it can be considered time-
to water pollution (Dzwairo et al., 2006). In recent years, several consuming and expensive. Sensors can also be regarded as another
studies have shown that the quality of groundwater has declined conventional approach. However, using sensors is considered
significantly in most countries (Adimalla, 2019; Gaikwad et al., costly to test all the water quality variables and often show low
precision (Oelen et al., 2018). Another solution for monitoring
⇑ Corresponding author at: Department of Information and Communication and water quality is predictive modelling using machine learning and
Technology, Mawlana Bhashani Science and Technology University, Santosh, deep learning approaches. Compared to other conventional meth-
Tangail 1902, Bangladesh. ods, it has several advantages: lower costs, efficient in terms of
E-mail addresses: [email protected] (M.S. Islam Khan), nazrul. time required for travel and collection, enables prediction under
[email protected] (N. Islam), [email protected] (J. Uddin), [email protected] various phases of a system, and predicts desirable values when
(S. Islam), [email protected] (M.K. Nasir).
accessing a site is inconvenient (Sinshaw et al., 2019).
Peer review under responsibility of King Saud University.
The researchers had extensively used prediction models for
their studies in the water quality management system over the last
few years including, the artificial neural network (Sinshaw et al.,
2019; Barzegar and Moghaddam, 2016; Barzegar et al., 2020;
Production and hosting by Elsevier

https://doi.org/10.1016/j.jksuci.2021.06.003
1319-1578/Ó 2021 The Authors. Production and hosting by Elsevier B.V. on behalf of King Saud University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Please cite this article as: Md. Saikat Islam Khan, N. Islam, J. Uddin et al., Water quality prediction and classification based on principal component regression
and gradient boosting classifier approach, Journal of King Saud University – Computer and Information Sciences, https://doi.org/10.1016/j.jksuci.2021.06.003
Md. Saikat Islam Khan, N. Islam, J. Uddin et al. Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

Hameed et al., 2017; Kadam et al., 2019; Zhang et al., 2019), mul- and WQI methods, respectively. They also found that the FLI
tiple linear regression (Choubin, 2016; Ewaid et al., 2018), least method is superior to the WQI method because of the relationship
square method (Salari et al., 2018), decision tree (Saghebian between measured values and WQI standard values. For the esti-
et al., 2014), random forest method (Yajima and Derot, 2018), mation of dissolved oxygen in aquaculture, Li et al. (2018), sug-
wavelet neural network approach (Xu and Liu, 2013), recurrent gested a synthetic model by combining Sparse-autoencoder and
neural network approach (Li et al., 2019), neuro-fuzzy techniques long short-term memory networks (LSTM). Although both CNN-
(Aghel et al., 2019; Khadr, 2017; Kisi et al., 2019; Zhu et al., LSTM and Sparse-autoencoder-LSTM models showed excellent
2019), and support vector machine method (Leong et al., 2019; performance since they predicted only DO and chlorophyll, it
Mohammadpour, 2015). In recent decades, those methods have may be challenging to deal with more water quality variables using
had a tremendous impact on the aquatic climate. But most of the such models. In another research, Asadollah et al. (2021), applied
models, including artificial neural network, wavelet neural net- an ensemble machine learning method called Extra Tree Regres-
work, recurrent neural network, and decision tree, required lots sion (ETR) which combines multiple week learners such as decision
of input parameters and computational power, which are consid- tree to predict WQI values in Tsuen River, Hong Kong. They applied
ered expensive to construct such models. the ETR method on ten water quality variables. Results indicated
With this motivation, this paper used water quality index that the ETR method achieved 98% prediction accuracy, which out-
(WQI), a combination of different water quality metrics that performed the other state-of-the-art models such as support vector
demonstrates the water quality condition of a particular region, regression and decision tree. Further, Hameed et al. (2017), devel-
and is applied both prediction and classification models to predict oped two neural artificial network techniques: a radial basis func-
WQI and classify the water quality status. Principal component tion neural network (RBFNN) and a backpropagation neural
regression (PCR) is used to predict WQI in this analysis, combining network (BNN) to predict the WQI in the tropical region of Malay-
both supervised and unsupervised techniques. PCR’s basic concept sia. The WQI was measured using sub-indices equations in this
is that the principal component analysis is applied to the dataset to study (Agamuthu and Victor, 2011). In both RBFNN and BNN
minimize the dimension. At the same time, a regression algorithm strategies, the training is faster, but the prediction takes a long
is used for the PCA output. Since it can solve dataset multi- time, making the model slow. Bui et al. (2020), proposed a hybrid
collinearity issues and allows fewer water quality specimens, PCR machine learning algorithm by combining the random tree and
can predict the WQI more efficiently than the other techniques. bagging (BA-RT) technique. The BA-RT method achieved 94% pre-
The Gradient Boosting Classifier method is used in this study for diction accuracy using a 10–fold cross-validation technique, outdo-
the classification task. It is an ensemble technique that can operate ing 15 standalone and hybrid algorithms. A more comprehensive
with a small amount of data. study into the application of machine learning methods for model-
Besides, the paper’s main contributions are as follows: ing river water quality was performed by Rajaee et al. (2020),
where they reviewed a total of 51 articles published from 2000

This study’s water quality index is determined using the form of to 2016. According to this study, artificial neural networks and
a ”weighted arithmetic index” method. wavelet-neural networks were the most widely used methods for

For the first time, the principal component regression method is predicting water quality. Furthermore, Samsudin et al. (2019),
applied to predict the water quality index. developed an artificial neural network. For this study, the most sig-

The water quality parameters are reduced using the PCR nificant water quality parameters were found through a spatially
approach, which allows using any water quality specimens. discriminant analysis (SDA). But these studies can barely show

For classifying the water quality status, a classification model is 71% accuracy. In another research, Yilma et al. (2018), applied an
presented. artificial neural network for predicting WQI in Ethiopia’s Akaki

A boxplot analysis is also carried out on the dataset to deter- River. In this analysis, an artificial neural network with eight hid-
mine the most dominant WQI parameter. den layers and 15 hidden neurons predict WQI with more than
90% accuracy. Also, Imani et al. (2021), applied an artificial neural
This paper has been formulated as follows: authors have dis- network with a single hidden layer for predicting water quality
cussed the literature review in Section 2. Then, Section 3 intro- resilience in São Paulo, Brazil. Applying neural networks to predict
duces the architecture of the proposed models with a proper WQI required lots of water quality data, which is expensive and
explanation of the subsection. Additionally, the authors have dis- time-consuming. Ho et al. (2019), applied a decision tree for clas-
cussed the experimental setup and the result analysis in Section 4. sifying water quality status in Klang River, Malaysia. They consid-
Finally, the authors have concluded the paper with limitations and ered three scenarios where they used six water quality variables in
future scope in Section 5. the first scenario. After that, in each procedure, they removed
water quality parameters such as NH3-N, pH, and SS to evaluate
the decision tree algorithm’s ability in different situations. They
2. Related works achieved 84.09%, 81.82%, and 77.27% classification accuracy in each
scenario, which is higher than the 75% classification accuracy
This section demonstrated the existing literature survey. The benchmark. Besides, to predict the WQI, Ahmed et al. (2019), used
author took the most common approaches to detect and classify several supervised machine learning methods. They conducted
the water quality, including deep neural network, recurrent neural their model on four water quality parameters. They found that
network, neuro-fuzzy inference, and support vector regression. by using gradient boosting and polynomial regression, the WQI is
For example, Barzegar et al. (2020), applied a CNN-LSTM amal- more successfully predicted where a multilayer perceptron classi-
gam model to predict two water quality variables, named Dis- fies the water quality category more effectively. However, this
solved Oxygen (DO) and chlorophyll-a. Results indicated that the study worked with fewer water quality parameters, but both pro-
CNN-LSTM amalgam model outperformed both the individual posed prediction and classification models did not show more than
CNN and LSTM model and the machine learning models such as 75% accuracy. On the other hand, Wang et al. (2017), applied sup-
SVR, Decision Tree. Oladipo et al. (2021), compared two statistical port vector regression to predict WQI. More than 90% of accuracy
methods, including Fuzzy Logic Inference (FLI) and WQI methods, was achieved in this analysis. In this study, 22 specimens of water
for evaluating the water quality in the Ikare community, Nigeria. quality were used, which makes the model computationally costly.
They found moderate and poor water quality conditions using FLI Li et al. (2019), proposed an amalgam model for the study of time-
2
Md. Saikat Islam Khan, N. Islam, J. Uddin et al. Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

series water quality data by integrating a recurrent neural network

with the Dempster-Shafer Theory (DST), where the RNN is capable
of analyzing time-series data effectively to predict WQI and DST,
which is a probability method used to amalgamate the outcome
of RNNs. It can be challenging to predict WQI using RNN and DST
since specialized handling of the data is required when fitting
and testing the model. Besides, Ahmed et al. (2019), proposed a
neuro-fuzzy inference method based on a wavelet-de-noise tech-
nique to predict water quality parameters. Results indicated that
this model outperformed the other neural network model, such
as RBF and MLP. But the neuro-fuzzy inference method causes a
curse of dimensionality problem, which occurs when high dimen-
sional data is analyzed and classified.
In summary, from the above studies, most of the current
approaches were based on a predictive model but did not provide
any classification model. Most of the models showed less accuracy
and used many water quality specimens. The proposed method is
applied to address the limitations described in the current Fig. 2. Lake position in Dhaka city.
approaches above. Also, the proposed model gives a dynamic
approach to use any number of water quality specimens.
3.2. Computation of WQI
3. Proposed method
This section illustrated the way the WQI is measured. A total of
The authors have presented a brief discussion on the proposed nine parameters were used to estimate the WQI, including pH, DO,
architectures for predicting and classifying WQI in this section, as COD, TDS, EC, Turbidity, Chloride, SS, and Alkalinity. The water is
showed in Fig. 1. A novel algorithm for the principal component then classified into a separate class, based on the estimated WQI
regression method is also presented in this section. The sub- [as presented in Table 2].
sections below presents descriptions of blocks related to the pro- Here, WQI is computed using the ‘‘weighted arithmetic index
posed models. method” (Tyagi et al., 2013), which was first proposed by Horton
(1965). According to this technique, water quality rating (Q j ) is
an integral part of the WQI and is determined using the following
3.1. Region analysis and dataset expression:

Gulshan Lake is situated in the northernmost part of Dhaka in Q j ¼ ððM j  lj Þ=ðSj  lj ÞÞ  100 ð1Þ
Bangladesh. It is considered an essential part of Dhaka city since
where Q j is considered to be the quality rating of the jth water qual-
it is one of Dhaka city’s remaining water bodies. It is surrounded
ity measurement, M j is referred to as the measured value from the
by the Diplomatic Zones of Baridhara, Tejgaon Thana, and Shahjad-
Gulshan lake, Sj is regarded as the standard value of the water coef-
pur. In such regions, Gulshan lake is considered a significant source
ficient recommended by WHO (World Health Organization et al.,
of surface water recharge. Gulshan Lake is 3.8 km long, with a total
2004), and lj is referred to as the ideal value for water quality
surface area of about 100 hectares. It is positioned at latitudes 23°
parameters. The ideal value considered for pH and DO, are 7 and
48 North and 90° 25 East with a mean depth of 2.5 m and a density
14.6 mg/l, respectively, and for other water quality measurements,
of 12–105 m3 (Rahman and Hossain, 2019). Fig. 2 indicates the
it is equal to zero. After calculating Q j , the unit weight W j is esti-
lake’s location in Bangladesh. In this analysis, the water quality
mated using the following expression:
variables were measured every month in 2016. WQI has been
determined utilizing indicators such as pH, suspended solids (SS), W j ¼ 1=Sj ð2Þ
electrical conductivity (EC), total dissolved solids (TDS), turbidity,
where, W j is referred to as the relative unit weight, Sj is regarded as
dissolved oxygen (DO), alkalinity, chloride, and demand for chem-
the standard value of the jth parameters, and 1 is defined as the
ical oxygen (COD). In this analysis, the dataset consisted of 108
constant of proportion. The unit weight factor (W j ) and the quality
specimens. The collected samples were obtained from the Environ-
ment Department (DOE) and the Environment and Forest Ministry, rating (Q j ) are combined to form WQI, which is expressed with the
Bangladesh (Dataset, 2016). Table 1 provides a detailed statistical following expression:
overview of the dataset.

Fig. 1. Prediction and Classification of water quality index.

3
Md. Saikat Islam Khan, N. Islam, J. Uddin et al. Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

Table 1
Statistical description of water quality specimens.

Parameter Count Min Max Mean SD Weight Relative weight

pH 108 6.85 8.27 7.35 .242 2.54 .1204
DO 108 0 15.3 5.20 3.76 4.09 .1938
COD 108 24 303 89.87 44.55 2 .0947
TDS 108 122.5 432 274.72 56.11 2.75 .1303
Turbidity 108 16 82 51.23 19.89 2.4 .1137
Chloride 108 15.8 76.9 33.43 11.03 1 .0473
SS 108 24 142 56.81 24.48 1.5 .0711
Alkalinity 108 104 210 150.68 19.82 1.6 .0758
EC 108 250 869 557.97 108.93 3.22 .1526
WQI 108 63 144 109.52 20.86

Table 2 The proposed PCR method for predicting the WQI is given as
WQI categories and applications (Tyagi et al., 2013).
Algorithm. 1.
WQI Range Water Quality Status Applications
0–25 Excellent Drinking, industrial and irrigation Algorithm 1. PCR algorithm
25–50 Good Drinking, industrial and irrigation
50–75 Medium Industrial and irrigation 1: Input: Af is the prediction matrix of the models in f.
75–100 Bad Irrigation 2: Output: aj , predicting WQI
Above 100 Unsuitable for drinking Treatment require before use
3: B ¼ cov ðAf Þ
. Where, cov() is a covariance matrix
X X 4: n = select_component(PCs)
WQI ¼ W jQ j= Wj ð3Þ b
5: f ¼ b1 PCs1 +b2 PCs2 +. . ...+bn PCsn
. Where, b=(PCsTn PCs1 T
n )PCsn z
PN
3.3. WQI prediction model 6: aj ¼ n¼1 bn cn ;j
. Where, cn ;j is the jth coefficient of the nth principal
This section demonstrated the development of the principal
component.
component regression model to predict the water quality index.
7: return a
PCR comes with the idea that performed PCA on the dataset and
then performed the regression model on the new PCs.
where the select_component() is used to select the number of com-

Data-Preprocessing: The dataset included some null values. For
ponents, however, it allows to use any number of components, but
handling such null values, the median method is used in this
it must be less than or equal to the number of features used initially.
analysis. Furthermore, Min–Max scalar is used to scale the data,
The best way to grasp the PCR model is to look at it mathemat-
which makes the computation easier.
ically. We can derive our regression equation using the following

Dimensionality-Reduction: For dimensionality reduction,
formula:
Principal Component Analysis (PCA) is applied to the dataset,
which extracts the most dominant water quality parameters. y ¼ bX þ  ð4Þ
PCA is a statistical analysis that reduced a dataset’s dimension-
From simple linear regression, multiple linear regression can be
ality that is influenced by the multi-correlated variables. Since
derived using the following expression:
PCA takes all the inter-correlated variables, it transformed them
into a small number of non-correlated variables that described y ¼ b0 X 0 þ b1 X 1 þ b2 X 2 þ    þ bk X k þ  ð5Þ
all the variances. The uncorrelated variables obtained from the
Here y is denoted as the expected value of the predictor vari-
PCA are known as principal components (PCs).
ables. b0 ; b1 ; b2 . . .bk is the regression coefficient function associ-

Data-Split: The collected data are then divided into two sets
ated with the independent X 0 ; X 1 ; X 2 , . . .X k variables, respectively.
after conducting PCA on the dataset: training and testing set
Then b is used to determine the interaction between the predictor
with a proportion of 80 and 20 percent.
variables and the independent variable. And finally,  is denoted as

Learning -Algorithm: For predicting the WQI, machine learning
vector of random errors. In PCR our first step is to normalize the
algorithms can be used. In this analysis, several machine learn-
data before executing PCA, as PCA is reactive to the data. Therefore,
ing algorithms are used, including Linear Regression, Gradient
we used one of the pre-processing steps to normalize the data in
Boosting regression, Random Forest Regression, and Support
such a way that r ¼ 1 and l ¼ 0. We conduct PCA on matrix X in
Vector Regression. Such algorithms are independently paired
the next step which is our independent variables. We can obtain
with the PCA method and implemented on the dataset. Finally,
X ¼ PDV 0 by performing PCA on the independent variable X by sin-
the best prediction model is selected by comparing the perfor-
gular value decomposition. Here, D is a diagonal matrix consisting
mances of such models. Table 3 presents the experimental
of q non-negative singular values and q is denoted as the explana-
parameters used for those models.
tory variables.
Table 3 2 3
D1 0  0
Experimental parameters for regression model.
6 0 D1  0 7
6 7
Model Parameters Values D ¼ diagonal½D1 ; D2 ; . . . Dq  ¼ 6
6 .. .. .. .. 7
7 ð6Þ
Muliple Linear Regression fit_intercept True 4 . . . . 5
Support Vector Regression kernel, C, degree poly, 200, 3 0 0    Dq
Gradient Boosting Regression learning_rate, n_estimators .1, 60
Random Forest Regression n_estimators 10

4
Md. Saikat Islam Khan, N. Islam, J. Uddin et al. Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

That’s how q  q matrix can be obtained with V as eigenvectors. selected using the statistical parameters like R2 , RMSE, MAE, Recall,
The matrix K ¼ XV can be obtained in the next step using eigenvec- Precision, and F1-score.
tor V, where each column is represented as a major part of q. Such
columns are orthogonal in the matrix and do not guarantee 4.1. Evaluation metrics
collinearity between the variables. We can cut the number of prin-
cipal components to reduce the collinearity between the variables. Three statistical parameters, including R squared error (RSE),
In matrix V, we must cut the components to r. After cutting the root mean squared error (RMSE) and mean absolute error (MAE),
components, the matrix looks like the K zr ¼ X zq V zr ; where z is are used to test the PCR model efficiency. These criteria are
denoted as the number of observations. We then perform regres- expressed as follows:
sion on the K matrix consisting of principal components q or r in
the next step. The coefficient can be determined using the follow- RSE ¼ 1  ððexplained v ariationÞ=ðtotal v ariationÞÞ ð10Þ
ing formula after conducting regression on the principal qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
component. RMSE ¼ ðyobs  ypred Þ2 =n ð11Þ
0 1 0
bk ¼ ðK KÞ ÞK y ð7Þ X
MAE ¼ ðjjyobs  ypred jjÞ=n ð12Þ
From the Eq. (7) the formula for the calculation is:
where, yobs = actual value, ypred = predicted value and n = total num-
y ¼ Kbk ¼ XðVbkÞ ¼ Xbx ð8Þ
ber of samples.
Finally, we can derive the following expression: Four classification metrics, including Accuracy, Recall, Precision,
and F1-score, are used to test the classification model performance.
bx ¼ Vbk ð9Þ These metrics are expressed as follows:
TP þ TN
Accuracy ¼ ð13Þ
3.4. WQS classification TP þ TN þ FP þ FN

This section demonstrated the development of a classification TP

Recall ¼ ð14Þ
model to classify water quality status. The water quality status is TP þ FN
divided into five groups, as suggested in Table 2.
TP
Precision ¼ ð15Þ

Resampling: Since most of the data belongs to group 4 and TP þ FP
group 5, resampling is performed on the dataset to remove
the data imbalance problem. After resampling the dataset, Precision  Recall
F1  score ¼ 2  ð16Þ
group 1, 2, 3, 4, and 5 contained a total of 10, 18, 19, 29, and Precision þ Recall
32 water quality specimens.
where, TP = True Positive, TN = True Negative, FP = False Positive,

Data-Preprocessing: In this technique, the median method is
FN = False Negative.
used to handle the null values and a min–max scalar to scale
the data.
4.2. Data and boxplot analysis

Data-Split: The dataset is divided into two sets: training and
testing with a proportion of 80 and 20 percent. Also to validate
Table 1 provides a detailed statistical overview of the dataset
the proposed model, we have changed the random state value
and the factors that affect water quality. Standard deviation was
(like 0, 1, or 41) multiple times during splitting the dataset so
maximum for TDS, EC, and COD, while the lowest standard devia-
that for every run, the train and test datasets would have differ-
tion was found for DO and pH. The minimum and maximum DO
ent values. Finally we take the average accuracy from the over-
values are 0 and 15.3 mg/ l, with a 5.20 mg/ l mean value. The tur-
all run.
bidity ranged from 16 to 82 NTU with a mean value of 51.23 NTU,

Classification-Algorithm: For classifying the water quality sta-
which means that these values are not distributed adequately
tus (WQS), several machine learning algorithms are used,
(World Health Organization et al., 2004). The relative weight has
including the Ada-Boost classifier, Random Forest classifier,
Support Vector classifier, and Gradient Boosting. From such
algorithms, the best algorithm is chosen by comparing the per-
formance. Table 4 presents the experimental parameters used
for those algorithms.

4. Experimental setup and result analysis

The objective of this section is to evaluate the performance of

the PCR and classification model. In this section, different PCR
and classification models are implemented, and the best model is

Table 4
Experimental parameters for classification model.

Model Parameters Values

Random Forest Classifier n_estimators, Criterion 15, gini
Support Vector Classifier C, Kernel, degree 2, linear, 1
Gradient Boosting Classifier learning_rate, n_estimators .1, 100
AdaBoost Classifier learning_rate,n_estimators 1, 50
Fig. 3. Water quality variables outlier detection.

5
Md. Saikat Islam Khan, N. Islam, J. Uddin et al. Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

Table 5 These parameters have a positive effect on estimating WQI as they

Comparison of different PCR models for predicting WQI. meet the standard values recommended by the WHO (World
Model R2 RMSE MAE Health Organization et al., 2004). Other parameters like COD,
TDS, SS, Alkalinity, and EC were extremely distorted and had high
PCA+ Multiple Linear Regressor .932 5.72 5.42
PCA+ Random Forest Regressor .839 8.87 7.82 skewed values compared to the standard value. So the average
PCA+ Support Vector Regressor .950 4.93 4.37 value of WQI at Gulshan Lake is high due to COD, TDS, SS, Alkalin-
PCA+ Gradient Boosting Regressor .722 11.6 9.15 ity, and EC. Therefore, it required adequate care.

4.3. PCR model result assessment

Table 6
PCR model performance assessment using various components.
The proposed PCR method implemented using python. The
Model Components R2 RMSE MAE results of different PCR models showed in Table 5. From this table,
PCR1 n = 1 .042 21.64 19.56
PCA with Support Vector Regression has achieved the highest accu-
PCR2 n = 2 .411 16.97 14.57 racy compared to the other PCR techniques. Although other PCR
PCR3 n = 3 .569 14.52 12.51 models also performed well, PCA with Gradient Boosting Regres-
PCR4 n = 4 .564 14.59 12.54 sion proved to be a less useful model.
PCR5 n = 5 .565 14.58 11.67
Since the PCR model provides to work with fewer parameters,
PCR6 n = 6 .709 11.92 9.44
PCR7 n = 7 .927 5.97 5.33 so we reduced the number of components instead of taking all
PCR8 n = 8 .932 5.72 5.42 the features. The results of taking different features showed in
PCR9 n = 9 .932 5.72 5.42 Table 6. For this technique, PCA with Multiple linear regression is
selected since PCA is mostly related to multiple linear regression
to create new principal components. Table 6 illustrated that, with
nine and eight components, the PCR9 and PCR8 models showed the
been mentioned for manually measuring the WQI. WQI was deter-
best performance, where PCR9 clarified all the variance. The PCR8
mined using Eqs. (1)–(3). From Table 1, the minimum WQI value is
model gives the same result as the PCR9 model, and the number of
found 63, where the maximum value is 144, with a standard devi-
ation of 20.86. The mean value of WQI is found 109.5, which shows parameters is also reduced. The R2 value for the PCR8 model in
that the quality of water is inappropriate for drinking and irriga- testing steps is .932. If we reduce one more component from the
tion in most areas. PCR8 model, that model produced almost the same result as oper-
For establishing the intricate relationship between the water ating with all the components. The R2 value in the PCR7 model is
quality parameters, outlier detection carried out on the dataset. .927. After reducing one more component, the R2 value reduced
Fig. 3 demonstrates the boxplot analysis to find the outlier of the in the testing phases is .709. That shows less accuracy compared
water quality parameters. From this analysis, parameters such as with the PCR7 and PCR8 models. Yet in the water samples, PCR6
pH, DO, chloride, and turbidity are found normally distributed. still performed well. If we reduce more components from the

Fig. 4. Prediction and Classification of water quality indexPlot between observed and predicted WQI a. PCA+ Multiple Linear Regression model b. PCA+ Random Forest
Regression model c. PCA+ Support Vector Regression model d. PCA+ Gradient Boosting Regression model.

6
Md. Saikat Islam Khan, N. Islam, J. Uddin et al. Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

showed poor performance. Since PCR7 and PCR8 showed the same
results as working with the PCR9, we could infer that the PCR
method allows operating with fewer parameters instead of taking
all the features.
Fig. 4 illustrated the plot between the observed and predicted
WQI values for a better understanding of those models. Among
them, the value appeared closer to the regression fit line in the
PCA+ Support Vector Regression model because of the high train-
ing and testing accuracy.

4.4. Classification model result assessment

Different classification algorithms are implemented using

python. The results of varying classification models are presented
in Table 7. Among them, the Gradient Boosting Classifier has
achieved the highest accuracy and proved to be an efficient model
Fig. 5. Comparison of accuracy of each part in PCR model. to predict water quality status. The second-best model is Random
Forest Classifier, but to calculate recall, the Support Vector Classi-
fier performs better than the Random Forest Classifier. Ada-Boost
Table 7 Classifier is found less effective model compared to the other tech-
Comparison of different classification model. niques. The confusion matrix for those models is presented in
Fig. 6. From Fig. 6, we can observe that the Gradient Boosting Clas-
Model Accuracy Precision Recall F1-score
sifier classify all the testing data according to the water quality
Random Forest Classifier .91 .96 .85 .89
level where other models misclassified some of the testing data.
Support Vector Classifier .86 .82 .91 .84
Gradient Boosting Classifier 1.0 1.0 1.0 1.0
AdaBoost Classifier .77 .53 .60 .56 4.5. Performance comparison with other studies

From Table 8, it is clear that both PCR and GBC methods have
outperformed the previously developed models such as standalone
PCR model, the R2 value is barely 50 per cent, which shows low PCR machine learning (SVR, GB, DT), deep neural network (NN, MLP),
model efficiency. and hybrid (SDA-ANN, BA-RT) in terms of predicting and classify-
The accuracy comparison of the PCR model in each principal ing WQI. The best accuracy was found by Bui et al. (2020) (accu-
component showed in Fig. 5. It is evident from Fig. 5 that the model racy = 94%). However, they had not provided any classification
performed well with six, seven and eight components. After then, it model. Yilma et al. (2018) and Wang et al. (2017) achieved 93%

Fig. 6. Confusion matrix for classification algorithm. A. Random Forest Classifier b. Support Vector Classifier C. Gradient Boosting Classifier d. AdaBoost Classifier.

7
Md. Saikat Islam Khan, N. Islam, J. Uddin et al. Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

Table 8
Performance comparison between the proposed method and previous literatures work.

Name Location Total Specimens Prediction Method Classification Method Prediction Accuracy Classification Accuracy
Wang et al. (2017) China 22 SVR – 92% –
Yilma et al. (2018) Ethiopia 12 NN – 93% –
Samsudin et al. (2019) Malaysia 13 SDA-ANN – 71% –
Ahmed et al. (2019) Pakistan 4 GB MLP 74% 85%
Ho et al. (2019) Malaysia 6 – DT – 81%
Bui et al. (2020) Iran 10 BA-RT – 94% –
Proposed Methods Bangladesh 9 PCR GBC 95% 100%

and 92% prediction accuracy using artificial neural networks and Ahmed, Ali Najah et al., 2019. Machine learning methods for better water quality
prediction. Journal of Hydrology 578, 124084.
support vector regression techniques. But those techniques used
Ahmed, Umair et al., 2019. Efficient water quality prediction using supervised
12 and 22 specimens of water that can be considered expensive Machine Learning. Water 11 (11), 2210.
since most of the specimens were examined in the laboratory. Fur- Asadollah, Seyed Babak, Seyed, Haji, et al., 2021. River water quality index
thermore, Samsudin et al. (2019), Ahmed et al. (2019) and Ho et al. prediction and uncertainty analysis: A comparative study of machine learning
models. Journal of Environmental Chemical Engineering 9, (1) 104599.
(2019) used fewer water specimens to predict and classify WQI. Barzegar, Rahim, Moghaddam, Asghar Asghari, 2016. Combining the advantages of
Yet, their model accuracy is less than 85%. Contrary to the previous neural networks using the concept of committee machine in the groundwater
methods, the proposed models used only nine water specimens in salinity prediction. Modeling Earth Systems and Environment 2 (1), 26.
Barzegar, Rahim, Mohammad Taghi, Aalami, Jan, Adamowski, 2020. Short-term
this study and demonstrated 95% prediction accuracy and 100% water quality variable prediction using a hybrid CNN–LSTM deep learning
classification accuracy. model. Stochastic Environmental Research and Risk Assessment, pp. 1–19.
Bui, Duie Tien et al., 2020. Improving prediction of water quality indices using novel
hybrid machine-learning algorithms. Science of The Total Environment 721,
5. Conclusion 137612.
Choubin, Bahram et al., 2016. Multiple linear regression, multi-layer perceptron
network and adaptive neuro-fuzzy inference system for forecasting
This paper demonstrated a method for predicting and classify- precipitation based on largescale climate signals. Hydrological Sciences
ing the water quality using machine learning algorithms. The Journal 61 (6), 1001–1009.
water metrics, including PH, DO, SS, EC, Turbidity, Chloride, COD, Dataset, Gulshan Lake, 2016. Published on May 20, 2018. URL: http://www.doe.gov.
bd/site/publications/5132a8d7- 68e9 - 469d - a9af - 8981306b3b9f/ Surface -
TDS, and Alkalinity, were used in this study. For data preprocess-
and - Ground - Water-Quality-Report-2016.
ing, the median technique used to handle the null values and Dzwairo, Bloodless et al., 2006. Assessment of the impacts of pit latrines on
min–max scalar to scale the data. For the prediction purpose, we groundwater quality in rural areas: a case study from Marondera district,
Zimbabwe. Physics and Chemistry of the Earth, Parts A/B/C 31 (15-16), 779–788.
applied the principal component regression (PCR) method. After
Ewaid, Salam Hussein, Abed, Salwan Ali, Kadhum, Safaa A., 2018. Predicting the
analyzing the performance of multiple PCR models, PCA with Sup- Tigris River water quality within Baghdad, Iraq by using water quality index and
port Vector Regression seems to be more effective with an accu- regression analysis. Environmental Technology & Innovation 11, 390–398.
racy of 95%. However, if the number of components reduced, Gaikwad, Satyajit et al., 2020. Geochemical mobility of ions in groundwater from
the tropical western coast of Maharashtra, India: implication to groundwater
then PCA with the Multiple Linear Regression model proved to quality. Environment, Development and Sustainability 22 (3), 2591–2624.
be more effective. For the classification purpose, the Gradient Hameed, Mohammed et al., 2017. Application of artificial intelligence (AI)
Boosting classifier used to classify the water quality status. Besides, techniques in water quality index prediction: a case study in tropical region,
Malaysia. Neural Computing and Applications 28 (1), 893–905.
to check the performance of the model, the proposed model is Ho, Jun Yung et al., 2019. Towards a time and cost effective approach to water
compared with several state-of-art classifiers, including Ada- quality index class prediction. Journal of Hydrology 575, 148–165.
Boost Classifier, Support Vector Classifier, and Random Forest Clas- Horton, Robert K., 1965. An index number system for rating water quality. Journal
of Water Pollution Control Federation 37 (3), 300–306.
sifier. Experimental results showed that the Gradient Boosting Imani, Maryam, et al., 2021. A novel machine learning application: Water quality
Classifier classified water quality status more efficiently. Despite resilience prediction Model. Science of the Total Environment 768, 144459.
the achievements outlined in this paper, some improvements are Kadam, A.K. et al., 2019. Prediction of water quality index using artificial neural
network and multiple linear regression modelling approach in Shivganga River
still possible, including we can collect more training samples to
basin, India. Modeling Earth Systems and Environment 5 (3), 951–962.
make the model more stable and more progress is possible on Kar, Devashish, 2013. Wetlands and Lakes of the World. Springer New Delhi, India.
the prediction model. Those issues will be overcome in future Kar, 2019. Wetlands and their Fish Diversity in Assam (India). Transylvanian Review
of Systematical and Ecological Research 21 (3), 47–94.
research, perhaps by proper tuning of the PCR model and using
Khadr, Mosaad, 2017. Modeling of water quality parameters in Manzala lake using
deep neural network. adaptive neuro-fuzzy inference system and stochastic models. In: Egyptian
Coastal Lakes and Wetlands: Part II. Springer, pp. 47–69.
Kisi, Ozgur et al., 2019. Modeling groundwater quality parameters using hybrid
Declaration of Competing Interest
neuro-fuzzy methods. Water Resources Management 33 (2), 847–861.
Leong, Wei Cong et al., 2019. Prediction of water quality index (WQI) using support
The authors declare that they have no known competing finan- vector machine (SVM) and least square-support vector machine (LS-SVM).
cial interests or personal relationships that could have appeared International Journal of River Basin Management, 1–8.
Li, Zhenbo et al., 2018. Water quality prediction model combining sparse auto-
to influence the work reported in this paper. encoder and LSTM network. IFAC-PapersOnLine 51 (17), 831–836.
Li, Lei et al., 2019. Water quality prediction based on recurrent neural network and
improved evidence theory: a case study of Qiantang River, China.
References
Environmental Science and Pollution Research 26 (19), 19879–19896.
Mohammadpour, Reza et al., 2015. Prediction of water quality index in constructed
Adimalla, Narsimha, 2019. Groundwater quality for drinking and irrigation wetlands using support vector machine. Environmental Science and Pollution
purposes and potential health risks assessment: a case study from semi-arid Research 22 (8), 6208–6219.
region of South India. Exposure and Health 11 (2), 109–123. Moon, Sang-Ki, Woo, Nam C., Lee, Kwang S., 2004. Statistical analysis of
Agamuthu, Pariatamby, Victor, Dennis, 2011. Policy trends of extended producer hydrographs and water-table fluctuation to estimate groundwater recharge.
responsibility in Malaysia. Waste Management & Research 29 (9), 945–953. Journal of Hydrology 292 (1-4), 198–209.
Aghel, B., Rezaei, A., Mohadesi, M., 2019. Modeling and prediction of water quality Oelen, Allard, van Aart, Chris J., De Boer, Victor, 2018. Measuring surface water
parameters using a hybrid particle swarm optimization–neural fuzzy approach. quality using a low-cost sensor kit within the context of Rural Africa. In: P-
International Journal of Environmental Science and Technology 16 (8), 4823– ICT4D@ WebSci.
4832.

8
Md. Saikat Islam Khan, N. Islam, J. Uddin et al. Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

Oladipo, Johnson O. et al., 2021. Comparison between fuzzy logic and water quality World Bank Report, 2018. Published on Sept 16, 2018. URL: https: //www.
index methods: A case of water quality assessment in Ikare community, dhakatribune.com/ bangladesh/ environment/ 2018/09/16/world-bank-urban-
Southwestern Nigeria. Environmental Challenges 3, 100038. pollution-costs-bangladesh- 6–5bn-a-year.
Rahman, Shafkat Shamim, Hossain, Md Mahboob, 2019. Gulshan Lake, Dhaka City, World Health Organization, et al., 2004. Manganese in drinking-water: Background
Bangladesh, an onset of continuous pollution and its environmental impact: a document for development of WHO Guidelines for Drinking-Water Quality.
literature review. Sustainable Water Resources Management 5 (2), 767–777. Tech. rep. World Health Organization.
Rajaee, Taher, Khani, Salar, Ravansalar, Masoud, 2020. Artificial intelligence-based Wu, Yiping, Liu, Shuguang, 2012. Modeling of land use and reservoir effects on
single and hybrid models for prediction of water quality in rivers: A review. nonpoint source pollution in a highly agricultural basin. Journal of
Chemometrics and Intelligent Laboratory Systems 200, 103978. Environmental Monitoring 14 (9), 2350–2361.
Saghebian, S. Mehdi et al., 2014. Ground water quality classification by decision tree Xu, Longqin, Liu, Shuangyin, 2013. Study of short-term water quality prediction
method in Ardebil region, Iran. Arabian Journal of Geosciences 7 (11), 4767– model based on wavelet neural network. Mathematical and Computer
4777. Modelling 58 (3–4), 807–813.
Salari, Marjan et al., 2018. Quality assessment and artificial neural networks Yajima, Hiroshi, Derot, Jonathan, 2018. Application of the Random Forest model for
modeling for characterization of chemical and physical parameters of potable chlorophyll-a forecasts in fresh and brackish water bodies in Japan, using
water. Food and Chemical Toxicology 118, 212–219. multivariate long-term databases. Journal of Hydroinformatics 20 (1), 206–220.
Samsudin, Mohd Saiful et al., 2019. Comparison of prediction model using spatial Yilma, Mulugeta et al., 2018. Application of artificial neural network in water
discriminant analysis for marine water quality index in mangrove estuarine quality index prediction: a case study in Little Akaki River, Addis Ababa,
zones. Marine Pollution Bulletin 141, 472–481. Ethiopia. Modeling Earth Systems and Environment 4 (1), 175–187.
Sinshaw, Tadesse A. et al., 2019. Artificial neural network for prediction of total Zhang, Yanyang et al., 2019. Integrating water quality and operation into prediction
nitrogen and phosphorus in US Lakes. Journal of Environmental Engineering of water production in drinking water treatment plants by genetic algorithm
145 (6), 04019032. enhanced artificial neural network. Water Research 164, 114888.
Tyagi, Shweta et al., 2013. Water quality assessment in terms of water quality Zhu, Senlin et al., 2019. Two hybrid data-driven models for modeling water-air
index. American Journal of Water Resources 1 (3), 34–38. temperature relationship in rivers. Environmental Science and Pollution
Wang, Xiaoping, Zhang, Fei, Ding, Jianli, 2017. Evaluation of water quality based on Research 26 (12), 12622–12630.
a machine learning algorithm and water quality index for the Ebinur Lake
Watershed, China. Scientific Reports 7 (1), 1–18.