Artificial Intelligence Approach to Predicting River Water Quality: A Review

Journal of Environmental Treatment Techniques

2020, Volume 8, Issue 3, Pages: 1093-1100

J. Environ. Treat. Tech.

ISSN: 2309-1185

Journal web link: http://www.jett.dormaj.com

Artificial Intelligence Approach to Predicting River

Water Quality: A Review

1,3

1,2*

1,2

Ariani Dwi Astuti , Azmi Aris , Mohd Razman Salim , Shamila Azman , Salmiati and

Mohd Ismid Md Said¹

Department of Water and Environmental Engineering, School of Civil Engineering, Faculty of Engineering, Universiti Teknologi Malaysia, Johor, Malaysia

Centre for Environmental Sustainability and Water Security (IPASA), Universiti Teknologi Malaysia, Johor Bahru, Malaysia

Department of Environmental Engineering, Faculty of Landscape Architecture and Environmental Technology, Universitas Trisakti, Jakarta, Indonesia

Received: 26/05/2020

Accepted: 07/07/2020

Published: 20/09/2020

Abstract

Precise prediction of the water quality time series may provide directions for early warning of water pollution and help policymakers to

manage water resources more effectively. This prediction may reveal the proclivity of the characteristic water quality according to the most

recent water quality, shifting, and transformation rule of the pollutant in the watershed. The predictive capability of traditional models is

constrained due to variability, complexity, uncertainty, inaccuracy, non-stationary, and the non-linear interactions of the water quality

parameters. Since the middle of the 20th century, Artificial Intelligence (AI) approaches have been found efficient in bridging gaps,

simulating, complementing deficiencies, and improving the precision of the predictive models in terms of multiple evaluation measures for

better planning, design, deployment, and handling of multiple engineering systems. This article discusses the state-of-the-art implementation

of AI in water quality prediction, the type of AI approaches, the techniques adopted include the knowledge-based system as well as literature

and their potential future implementation in water quality modelling and prediction. The study also discusses and presents several possibilities

for future research.

Keywords: Water quality simulation; Artificial Intelligence; Knowledge-based system; Review

Introduction¹

have parameter-wide nonlinear relationships with each other.

Conventional data processing cannot address this significant

limitation (10–12). Nonlinear Artificial Intelligence (AI) models,

on the other hand, play a significant role in simulating complex

and nonlinear processes (13). This situation creates a big gap

between model designers and professionals. Selecting a suitable

numerical model is a challenging task for novice application

users. The forecast precision of traditional models is restricted

due to the uncertainty, unpredictability, obscurity, and inaccuracy

of water quality information. Progress in AI technology has made

it possible over the past decade to apply the developments in

computational modelling systems to bridge the gap, as mentioned

above (8).

Chemical, physical, and biological properties found in water

are generally referred to as the quality of water (1). Accurate

evaluation of water quality using the Water Quality Management

program is important if decision-makers are to understand,

interpret, and use these data to support resource management

practices (2,3). Modelling of water quality parameters is an

essential part of every water systems analysis. In order to properly

manage the watershed, it is necessary to predict the quality of

surface water so that appropriate measurements can be taken to

avoid pollution from the permissible concentrations. Ideal

management of water resources is based on accurate and reliable

estimates of future changes (1,4–6).

QUAL2E, Water Quality Analysis Simulation, and the U.S.

Army Corps of Engineers Hydrological Engineering Center-5Q

are several models commonly applied to water quality

management (7). These models, however, are not only time

consuming and expensive, but also lack user-friendliness and

effective knowledge transfer in model interpretation. Therefore,

more models, which do not suffer from these problems, are

needed to be developed (8,9). Several scientists noted that the

prediction of water quality is impacted by various variables that

AI methods are currently capable of mimicking this behaviour

(14), complementing the defect, and improving the precision of

forecast models in terms of multiple assessment measures for

better planning, design, operation, and management of distinct

engineering systems (15, 16). The significant contributions of the

present review article are 1) to categorise AI methods

comprehensively and 2) to discuss their advanced application to

water quality modelling and prediction.

Corresponding author: Azmi Aris, (a) Department of Water and Environmental Engineering, School of Civil Engineering, Faculty of

Engineering, Universiti Teknologi Malaysia, Johor, Malaysia and (b) Centre for Environmental Sustainability and Water Security (IPASA),

Universiti Teknologi Malaysia, Johor Bahru, Malaysia. Email: azmi.aris@utm.my.

093

Journal of Environmental Treatment Techniques

2020, Volume 8, Issue 3, Pages: 1093-1100

extremely precise because it conveys all kinds of interactions

expressed in the information, including fundamental physics and

chemistry (9). Some studies (6, 11, 22-27) that explored river

water quality modelling issues using AI methods have revealed

encouraging outcomes in recent decades (Table 1).

Several researchers have attempted to predict water quality

parameters using AI-based models such as ANN, SVM, and k-

NN. In these studies, ANN has been frequently found a stronger

predictive model compared to conventional modelling

techniques. In the case of 47 sources (2007-2019) reviewed,

ANN, SVM, and k-NN have been used in 38, 10, and 1 source,

respectively. ANN has been widely used between 2007 and 2015,

but from 2015 to 2019, ANFIS and SVM have surpassed ANN,

as more recent approaches of AI. Some studies made a

comparison between the models.

The study found that different parameters are needed to be

used in water quality assessments using various techniques.

Different output parameters predictions have been studied, but the

ten most important parameters are DO, BOD, TSS, Total

Nitrogen, temperature, COD, turbidity, Total Phosphate, NH ,

and WQI. The monthly water quality data have been used most in

many of these studies to simulate water quality parameters [4, 5,

Artificial Intelligence-Based Model for River

Water Quality Simulation

The 1956 Dartmouth Conference was held at a time when AI

earned its name, purpose, and first accomplishments; it was

widely recognised as the birth of AI. Across various fields, the AI

field is currently playing an important role, focusing on machines

with a human-like mind (17). By incorporating descriptive

understanding, procedural knowledge, and reasoning, AI methods

enable researchers to simulate human knowledge in clearly

defined domains. In addition, advances in AI techniques have

enabled the creation of intelligent management systems through

the use of shells under established platforms such as MathLab,

Visual Basic, and C++. (8, 18).

Recently, AI has achieved significant progress in multiple

programs such as autonomous driving, big data, information

processing, smart search, image understanding, automatic

software development, robotics, and human-computer games,

which will have a significant effect on human society. Some of

the most important AI-based algorithms include artificial neural

networks (ANNs), support vector machine (SVM), random forest

(

RF), genetic algorithm (GA), enhanced regression tree (ERT),

simulated annealing (SA), imperialist competitive algorithm

ICA), and decision tree (DT). AI methods are also associated

0, 11, 16-28], which was followed by daily water quality data (3,

5, 35-40).

(

with experimental design (e.g., response surface methodology,

and standardised design) to improve the precision of the optimal

solution prediction (19). Advances in data science and data

mining techniques such as neural networks (NNs), supporting

vector machines (SVMs), and k-nearest neighbours (k-NN) have

helped to solve some complicated high-dimensional issues in

river water quality prediction (Figure 1).

Artificial Neural Network Modelling in River

Water Quality Monitoring

The theory of artificial neurons was first launched in 1943,

with the implementation of the back-propagation practice (BP)

algorithm for feedforward ANNs in 1986 (23). ANN is a recent

method with a versatile mathematical structure that can identify

complicated non-linear interactions between input and output

information compared to other traditional modelling approaches

(1, 25).

Over the past two decades, river water pollutants have been

considered as one of the global issues that need the full attention

of environmental scientists. River water quality, however, is one

of the main characteristics to which environmental scholars need

to pay full attention. In all developing countries, water quality is

a growing concern. Water abstraction mechanisms of domestic

use, farming, mining, energy generation, and forestry practices

may lead to a decline of water quality and quantity, which affects

not only aquatic ecosystems but also the allocation of safe water

for human consumptives (20). Thus, the assessment of surface

water quality is important in the management of water resources

and is very important in monitoring the concentration of

pollutants in rivers. Monitoring water quality is costly because

pollution control and efficient water resource management

require large quantities of data (21). Therefore, AI can be

recommended as an alternative technique with high prediction

accuracy for predicting the river water quality. AI benefits from

traditional techniques since they take account of the non-linear

relationship between influential variables and reduce the

complexity required to obtain experimental equations (20).

The overall concept behind AI techniques is to explore hidden

interactions in large quantities of information and to create

models that represent physical procedures governing the system

being studied. A model derived from data reflects a correlation

between variables of input and output. Such a model can be

ANNs are common instruments applicable to modeling

extremely complex relations, processes, and phenomena. ANNs

have been also widely used to predict water quality variables to

address contaminant source uncertainty and nonlinearity of water

quality data. Nevertheless, the issue with the initial weight

parameter and the unbalanced training data set makes it hard to

determine the optimal outcomes and hinders ANN modeling

efficiency (25). ANN consists of very basic processors called

neurons that are strongly interconnected and act together to solve

a problem (41). A neuron is an information processing unit,

essential for the functioning of the NN; it comprises weight and

activation.

From 2007 to 2019, eight types of ANN were applied by

different researchers to the prediction of river water quality,

namely Back Propagation NN (BPNN), Wavelet NN, Generalized

Regression NN (GRNN), Radial Basic NN (RBNN), Feed

Forward NN (FFNN), Multi-layer Perceptron NN (MLPNN),

Multi-layer Feed Forward NN (MLFFNN) and Adaptive

Network-Based Fuzzy Inference System (ANFIS). Among them,

five most widely-used models MLPNN (10), RBNN (6), FNNN

(5), ANFIS (5), and MLFFNN (4).

094

Journal of Environmental Treatment Techniques

2020, Volume 8, Issue 3, Pages: 1093-1100

Artificial Neural Network (ANN)

Single Approach

Support Vector Machine

k-nearest neighbors

Artificial Intellegency (AI)

Technique

WAVELET-ANFIS

WAVELET-ANN

ANN-ARIMA

Hybrid Approach

Figure 1: Classification Tree of AI Techniques Applied in Literature to River Water Quality

Several indicators often used to evaluate the ANN model's

simplified architecture of two weight layers with basic function

parameters in the first layer, while the second layer contains linear

combinations of those basic functions for the processing of the

output and also contains parameters for water quality (1). Some

performance are as follow: Coefficient of correlation (R) between

the values observed and the expected values, the mean square

error (which can be used to calculate how well the network output

corresponds to the expected output), mean absolute error (MAE),

root mean square error (RMSE); Coefficient of efficiency (CE),

Mean absolute prediction error (MAPE) (which usually expresses

accuracy as a percentage), interquartile range (IQR) (which refers

researchers have predicted DO, COD, TDS, NH -N, Turbidity,

and WQI using RBFNN. Cobaner et al. in 2009 used GRNN,

MLP, and RBNN to forecast SS. Their results confirmed that

RBNN performed slightly better than the others (37). In addition,

Ahmed (2017) compared some models regarding the DO

prediction, which again showed the superiority of the RBNN's

performance over the rivals (33).

to difference between the 25 and 75 percentile and is used to

calculate the entire bias error between the means of the ensemble

and the values observed), Nash-Sutcliffe coefficient (NSC), and

Determination coefficient (DC) (Table. 2).

FFNN propagates the data linearly from input to output; in many

practical applications, they are the most popular and widely-used

models (43). FFNN is used to predict DO, BOD, TN, and

temperature. Four scientists implemented monthly data in FFNN

to predict DO using different input parameters (11, 13, 33, 43).

Ahmed in 2017 predicted DO using BOD and COD parameters

as input, and obtained R = 0.936 and RMSE = 0.709. Ranković et

al. in 2010 used FFNN to check its capability to predict DO. They

added more variables of water quality. Their findings showed that

pH and the water temperature are the most powerful variables in

DO prediction. In (44), ANFIS was used to learn neural network

algorithms and fuzzy logic was used to construct a non-linear

mapping between inputs and outputs. Ahmed et al. (2017) and

Khaled et al. (2018) conducted a study into BOD prediction using

ANFIS, and they suggested that the ANFIS technique could be

successfully applied to building models for predicting the river

water quality (34, 55). Elkiran et al. (2018) attempted to predict

DO at Yamuna River using ANFIS, FFN, and MLR. They found

that even both FFNN and ANFIS were found capable of handling

nonlinear interactions, the ANFIS model performed better than

FFNN (13). The most predicted parameters for ANN are done,

DO, BOD, COD and WQI, respectively 13, 5, 5, and 4 studies.

And so far, for certain output parameters from certain inputs, each

NN type has achieved good results.

The fundamental MLPNN model has three layers: (i) input

layer, (ii) hidden layer, and (iii) output layer. The input layer

supplies the input data set, the hidden layer processes the features,

and finally, the output layer shows the expected results. As can be

seen in Table 2, MLPNN is widely used to predict the DO

parameters. (15, 28, 31, 32). Moreover, MLPNN is used to predict

BOD, COD, EC, TDS, turbidity, and WQI. In predicting the DO

parameters, different inputs are used. However, Olyaie et al. in

017 and Ay and Kişi in 2017 both used pH, EC, temperature,

and flow as input parameters (15,31). The difference is that Olyaie

et al. used daily data of seven years, while Ay and Kişi used

monthly data of 15 years. Regarding the performance, Olyaie et

al. obtained R = 0.955 and root mean square error (RMSE) =

.594, while Ay and Kişi obtained R = 0.98 and RMSE = 0.52.

The time scale of the data does not seem to have any impact on

outcomes.

Radial basis function neural network (RBFNN) is a type of

NN applicable to general purposes and to various problems. The

RBFNN model is more advantageous than other types of NN that

have a grouping phase during training, where the hidden node's

central location is calculated (36). RBFNN was developed as one

of the most common three-layer neural feedforward networks

(42) to determine parameters of water quality. This model has a

095

Journal of Environmental Treatment Techniques

2020, Volume 8, Issue 3, Pages: 1093-1100

Table 1: Artificial Intelligence-Based Models Applied to Water Quality Prediction

Type of

Approach

Artificial

Neural

Methods

Output Parameter

River

Authors

Back Propagation

neural networks

COD, DO, NH

Sediment

Dahan River, Taiwan; Jishan

Lake, China; River Suktel,

India; Yuqiao reservoir in

Tianjin

Zhao et al. (2007); Xu and

Liu (2013); Chang et al.

(2015); Ghose and

Network

(ANN)

Samantaray (2018)

Wavelet Neural

Network

Jishan Lake, China

Xu and Liu (2013)

Generalized

COD

Cark Creek, Turkey

Ay and Kisi (2014)

Regression NN

Radial Basic

Neural Networks

COD, DO, NH

Surma River, Bangladesh;

Yangtze River, China; Johor

River, Malaysia; Cark Creek,

Turkey; Kopili River, India

Ahmed (2017); Basis et al.

(2014); Najah et al. (2013);

Ay and Kisi (2014); Kumar et

al. (2016)

TDS, Turbidity,

Suspended

sediment

Feed Forward

Neural Network

BOD, DO, Total

Nitrogen,

Temperature,

WQI

Gomti river, India; Melen River, Singh et al. (2009); Dogan et

Turkey; Surma River,

al. (2009); Ahmed (2017); He

et al. (2011); Gazzaz et al.

(2012); Elkiran et al. (2018)

Bangladesh; 59 rivers in Japan;

Kinta River, Malaysia; Yamuna

River

Multi-layer

Perceptron Neural

Networks

WQI, DO, BOD,

COD, TDS,

Turbidity,

Johor River, Malaysia; Heihe

River, China; Cark Creek,