Addis Ababa University

Addis Ababa Institute of Technology

School of Electrical and Computer Engineering

Machine Learning-Based Contamination Detection in Water

Distribution System

A Thesis Submitted to Addis Ababa University School of Graduate Studies in Partial

Fulfillment of the Requirement for the Degree of Master of Science in Computer
Engineering

By
Akalewold Fikre

Advisor
Getachew Alemu (Ph.D.)
June, 2020
Addis Ababa, Ethiopia
 Addis Ababa University
Addis Ababa Institute of Technology
School of Electrical and Computer Engineering

Machine Learning-Based Contamination Detection in Water

Distribution System

Approved by Board of Examiners

Name Signature Date

(Dean, School of Graduate committee)

Advisor

Internal Examiner

External Examiner
Declaration

I, the undersigned declare that this thesis is my original work and has not been presented for a
degree in this or any other university, and all sources of materials used for the thesis have been
fully acknowledged.

Name Signiture

Akalewold Fikre __________________

Place: Addis Ababa Institute of Technology, AAiT

Addis Ababa University, AAU
Addis Ababa,
Ethiopia.

Submitted in: June, 2022

This thesis has been submitted with my approval as a university advisor

Name Signiture

Getachew Alemu (Ph.D.) __________________

i
 Acknowledgment

“It is good that a man should both hope and quietly wait for the salvation of the LORD”.
Lamentations 3:26

Above all, Glory to Almighty God and the savior, Jesus Christ, for giving me all the courage and
strength in the course of the study and for the completion of this research. I know any good that
has or will come from the above.

I am grateful to all those who participated in this study, specifically to the managers of different
departments of AAWSA, and expertise as well as family members that are willing to participate
in the research. Obviously, without their openness to share their knowledge and experience, this
study would not have been possible. I also want to thank the institution, Addis Ababa water, and
sewerage Authorities, for their kind support in accessing the laboratory, manuals, and different
books and documents. I am also very pleased to thank the whole Staff, at Addis Ababa water and
sewerage authorities, for their devotion to providing the help I needed throughout the research
process.

My sincere thank also goes to Getachew Alemu (Ph.D.), my supervisor, for all his inspiring
comments, friendly advice, suggestions, and scholarly critics from the beginning to the end has
made it possible the completion of the thesis writing. I would also thank Menor Tekeba for his
time and comments that helped to improve the thesis.

I am also indebted to thank my family for their prayers, advice, and support. My special thanks
go to big Sister Amelework Fikre, Solomon, Leykun Gashawbeza, Firehiwot G/Silase, Mustafa
Mehedi, Tekel Tufa, Bekele Desta, Biruk Teshome, Sultan Hasen, and Hana Biruk for all kinds
of help and advice.

Last, but not least, it is my wish to thank all my Sunday BS team members, for their all-rounded
help and prayers

ii
 Abstract

Water is a necessary component of all human activities. According to the United Nations World
Water Assessment Program, every day, 2 million tons of sewage, manufacturing, and
agricultural waste are discharged into the world's water. Due to population demands and
dwindling clean water supplies as well as available water pollution management mechanisms,
there is an urgent need to use computational methods to intelligently manage available water. To
ensure the protection of drinking water, accurate detection of natural or deliberate pollution
events in water delivery pipes is essential. Companies that have water must ensure that it is safe
to drink. To resolve the global issue of rising water contamination, the design of water
contamination detector models has monitored the security of water in pipelines when
concentrations of water quality variables in the pipes surpass their maximum threshold is
presented in this paper. This paper proposes artificial neural networks, specifically
Convolutional Neural Networks, for automated water impurity, detection to refine the model
must a picture of turbid water in the pipe is used to detect events. The algorithm of deep learning
achieved 96.3 percent accuracy after extensive training with a dataset of 4220 images reflecting
various levels of contamination. Besides that, the machine learning algorithm uses an efficient
study of water turbidity and transparency levels to estimate the level of pollution in a specific
sample of water. As the established model is combined with the current framework, it will
provide a cost-effective way for the water company to obtain an estimate of water quality,
alerting local and national governments to take action, and potentially saving millions of people
throughout the world.

iii
 Table of Contents
Declaration .................................................................................................................................. i
Acknowledgment ....................................................................................................................... ii
Abstract ..................................................................................................................................... iii
Acronyms ................................................................................................................................ viii
Chapter One .................................................................................................................................. 1
1.1 Background of the Study .................................................................................................. 1
1.2 Statement of the Problem ................................................................................................. 4
1.3 Objective of the Research ................................................................................................ 5
1.3.1 Specific Objectives ................................................................................................... 5
1.4 Scope of the Research ...................................................................................................... 5
1.5 Limitations of the Study ................................................................................................... 5
Chapter Two .................................................................................................................................. 7
Literature Review............................................................................................................................ 7
Chapter Three ............................................................................................................................. 18
Research Methodology ............................................................................................................... 18
3.1 Introduction .................................................................................................................... 18
3.2 Research Site .................................................................................................................. 18
3.3 Data collection tools ....................................................................................................... 20
3.3.1 Secondary Data Sources ......................................................................................... 21
3.4 Method of Data Analysis................................................................................................ 21
3.4.1 Collection of water samples .................................................................................... 21
3.4.2 Preliminary-scale contaminant injection and monitoring system ........................... 23
3.4.3 Developing the dataset ............................................................................................ 24
3.5 Model Design ................................................................................................................. 27
3.5.1 Architectural design of the Proposed System ......................................................... 27
3.5.2 Experimental Setup ................................................................................................. 34
3.6 Ethical Consideration ..................................................................................................... 38
3.6.1 Informed Consent.................................................................................................... 38
3.6.2 Respecting for Person ............................................................................................. 39

iv
Chapter Four ............................................................................................................................... 40
Result and Discussion ................................................................................................................. 40
4.1. System Models Analysis ................................................................................................ 40
4.2. Optimization ................................................................................................................... 43
4.3. Discussion ...................................................................................................................... 51
Chapter Five ................................................................................................................................ 55
Conclusion and Recommendation ............................................................................................. 55
5.1. Conclusion...................................................................................................................... 55
5.2. Recommendation ............................................................................................................ 56
References .................................................................................................................................... 57
Appendix ...................................................................................................................................... 63
Appendix: A Sample code......................................................................................................... 63
Appendix: B Data Analysis Results of clean water .................................................................. 64
Appendix: C Data Analysis Results of contaminated water ..................................................... 66

v
 List of Figures
Figure 3. 1 The raw water at the time of sampling ..................................................................................... 22
Figure 3. 2 Water Reservoir at the time of sampling. ................................................................................. 22
Figure 3. 3 Customer Tap water at the time of sampling. .......................................................................... 23
Figure 3. 4 A process flow schematic of the Preliminary-scale system...................................................... 24
Figure 3. 5 (a) is clean water with normal light and (b) is clean water with blue light............................... 25
Figure 3. 6: (a) is contaminated water with 100mg of sodium nitrate and (b) is contaminated water with
400mg of sodium nitrate. ............................................................................................................................ 25
Figure 3. 7: (a) is contaminated water with clay at low concentration and (b) is contaminated water with
clay at high concentration. .......................................................................................................................... 26
Figure 3. 8: (a) is contaminated water with sand at low concentration and (b) is contaminated water with
sand at high concentration. ......................................................................................................................... 26
Figure 3. 9: (a) is contaminated water with silt at low concentration and (b) is contaminated water with
silt at high concentration. ............................................................................................................................ 26
Figure 3. 10: Contaminated water with 34mg/l of Iron. ............................................................................. 27
Figure 3. 11 Basic layout of a typical ML system has several stages. (Source: A Khan et al. (2020)). ..... 29
Figure 3. 12 Architecture of the Convolutional Neural Network. ............................................................. 30
Figure 3. 13 3x3 Output matrix.................................................................................................................. 31
Figure 3. 14 Max Pooling. .......................................................................................................................... 32
Figure 3. 15 Image processing and analysis. ............................................................................................. 34
Figure 3. 16 Segmentation Process ............................................................................................................. 37
Figure 4. 1 Accuracy, loss, validation accuracy, and validation loss for the proposed model.................... 40
Figure 4. 2 Training and validation accuracy for the proposed model ....................................................... 41
Figure 4. 3 Training and validation loss for the proposed model ............................................................... 41
Figure 4. 4 Training and validation accuracy for the modified proposed model ........................................ 42
Figure 4. 5 Training and validation loss for the modified proposed model ................................................ 43
Figure 4. 6 Validation accuracy for the top 3 models ................................................................................. 45
Figure 4. 7 Validation losses for the top 3 models...................................................................................... 46
Figure 4. 8 Training and validation accuracy for the best models .............................................................. 46
Figure 4. 9 Training and validation loss for the best models ...................................................................... 47
Figure 4. 10 Graphical user interfaces ........................................................................................................ 50

vi
 List of Tables
Table 4. 1 Model’s accuracy, loss, and validation accuracy and validation loss .......................... 44
Table 4. 2 Shows the top 9 models in terms of validation accuracy and validation loss. ............. 45
Table 4. 3 A multiclass Confusion Matrix of the classes ............................................................. 48
Table 4. 4 Precision, Recall, and F1 score of the classes.............................................................. 49

vii
 Acronyms

AAWSA – Addis Ababa Water and Sewerage Authorities

BCM - Billion Cubic Meters

CNN - Convolutional Neural Network

NTU – Nephelometric Turbidity Unit

SVM - Support Vector Machine

UNDP - United Nations Development Program

UNEP - United Nations Environmental Protection

UNESCO – United Nations Educational Science and Cultural Organization

U.S.EPA – US Environmental Protection Agency

US – United States

WDS – Water Distribution System

WHO – World Health Organization

WWF - World Wildlife Fund

viii
 Chapter One
1.1 Background of the Study

The water resource covering 70% of its surface is a source of supply for all kinds of life on the
planet (Jeremiah Castelo, 2021). When you consider how much more water is still in the ocean
and how much of the earth's atmosphere is filled with it, water scarcity will not be a big problem
for most people. However, the World Wildlife Fund for Nature estimates that over 1 billion
people lack healthy and clean water. On the other hand, Africa, especially Sub-Saharan Africa
and South Asia, is home to the majority of the world's people—more than 2.6 billion people—
who don't even have access to improved sanitation (WHO/UNESCO, 2010). According to other
reports, over 2.7 billion people are affected by water shortages at least once a year (Jeremiah
Castelo, 2021).

It is only 3% of the world's water is considered freshwater, even though it occupies 70% of the
planet's surface. Furthermore, humans are unable to enter about 2.6 percent of this freshwater.
They are, either trapped in polar glaciers or ice caps, frozen in the atmosphere or soil, heavily
polluted, or too far under the surface of the earth to be removed. It leaves about 0.4 percent of the
water in the world that is available and drinkable to share among the planet's 7 billion people
(World Atlas, 2018). Even so, most of this 0.4 percent is difficult to reach. The majority is
located within underground aquifers that can be reached by digging wells. The remainder is
found in rivers and streams, which we call surface water. With such a small proportion of clean
water on the planet's surface, most of the population of the world is suffering (Jeremiah Castelo,
2021).

Since surface water is easier to reach, it has become the most popular method for humans to
obtain healthy drinking water. Every day, nearly 321 billion gallons of surface water and 77
billion gallons of groundwater globally. Surface water includes lakes, rivers, streams, and
reservoirs located on the earth's surface. Surface water accounts for 80% of global daily water
use, and the bulk of that water is for irrigation and public supply. Oceans are the main source of
surface water, accounting for the world's 97 percent, but it is unusable for humans due to their
high salinity. (Jeremiah Castelo, 2021).

1
Drinking pure water is an issue for water supply companies around the world, and it is currently
a well-known problem due to numerous vulnerable threats. Changes in water chemistry can
occur before earthquakes, because of terrorist acts, or other man-made pollutants. The changes in
water chemistry caused by an earthquake can be an excellent indicator of earthquake predictions.
However, it can also be dangerous to human health because some changes are toxic. When a
system event occurs, such as large-scale water contamination because of an accident or a
malicious attack, it can affect society and the economy. One such instance is the 1993
contamination event in Milwaukee, which affected 403,000 people, resulting in thousands of
hospitalizations and a handful of fatalities, at $96.2 million in medical and productivity costs (P.
Corso, M. Kramer, K. Blair, D. Addiss, J. Davis, A. Haddix, 2003). Another incident occurred in
2014 in West Virginia (USA) when the drinking water distribution system was contaminated
accidentally with crude MCHM, an industrial chemical (W. J. Cooper, 2014). It affected 300,000
people.

Because of all the threats to public health, water system pollution detection is critical. As a
result, it is serious that water utilities assess the hydraulic and quality features of the water
distribution system to ensure that all customers have water that is safe to consume. Nowadays,
contamination-warning systems are used by water utilities to control the quality of drinking
water. They monitor water quality and environmental data at multiple measuring stations using
sensors. Despite these measurements, the system also requires a detection system that can
accurately alert users to changes in water quality based on the data collected. Machine learning
may produce good results for detecting contaminants in water quality, and this behavior
motivated the development of a model. The number of erroneous predictions can be reduced
dramatically using machine learning methods. The use of artificial neural networks for modeling
water quality to compute estimation errors was recently examined, and the contamination event
probability was calculated and compared to a given threshold to detect an event using a
sequential Bayesian rule (L. Perelman, J. Arad, M. Housh, A. Ostfeld, 2012). Neural networks
are a type of machine learning model that seeks to replicate characteristics of the human brain
and its biological neural networks. They are the foundation of most deep learning approaches, a
subset of machine learning that entails running data through numerous consecutive layers to
conduct classification or pattern analysis. Computer vision, bioinformatics, and natural language

2
processing are just a few domains where deep learning has aided growth. Convolutional neural
networks (CNNs) are a type of neural network, that is, a subset of deep learning. CNN's are large
networks of nodes called "neurons" that make connections as they learn from data. Because CNN
requires individuals to define specific features for a model to assess, they undergo supervised
learning. As a result, the model's creator constructs the model's structure, spawning multiple
layers that execute various functions and contribute to the model's performance in different ways.
Researchers and data scientists have discovered CNNs' aptitude for numerous jobs, most notably
image classification, in recent years, and they have increased in popularity as a result. The
suggested system is an image-based water quality evaluation system. There are in-depth water
quality measurements available, such as the index of water quality (WQI). The suggested model
aims to offer a rapid estimate of water quality that does not require comprehensive chemical
testing and can be performed anywhere and at any time.

Previous studies have found a relationship between turbidity levels and gastrointestinal disease,
such as Schwartz and Levin in 1999 [Schwartz, J., Levin, R.]. As a result, the study believes that
assessing the water's turbidity images is valid to determine the risk of drinking that water.
Turbidity is an important indicator that may provide valuable information quickly, cheaply, and
regularly. Turbidity may not always imply a direct threat to public health. However, it can
indicate the presence of pathogenic microbes and serve as an early warning system for dangerous
events occurring throughout the water supply system, from watershed use. Turbidity can be
measured easily, accurately, and quickly and is often used for operational monitoring of control
measures contained in water safety plans (WSPs), the WHO's recommended strategy for
maintaining drinking-water quality (WHO, 2017). Instruments can only detect turbidities below
4 Nephelometric Turbidity Unit (NTU), but the murky, red-brown, or black suspension can be
detected at 4 (NTU) and above, lowering the acceptability of drinking water. Turbidity can be
used in operational settings. Water quality in distribution systems is monitored as an indicator of
network integrity and proper operation and maintenance. The rapid changes in turbidity can
indicate significant pollution incidents in surface water and groundwater catchments. However,
some particles with turbidity values ranging from 5 to 13 Nephelometric Turbidity Units (NTU)
are unseen by the naked human eye. This study aimed to develop a comprehensive machine
learning model that only requires an image of water to assess whether or not contaminants are

3
present, making water quality analysis and determining the safety of drinking water much
simple.

1.2 Statement of the Problem

Water distribution networks (WDN) are vital infrastructures that can be contaminated
chemically, biologically, or radioactively, either intentionally or accidentally. Water utilities
often take manual samples and perform chemical analyses regularly to monitor the water quality
in the system. On the other hand, manual sampling can take days to discover, usually after a
chemical analysis or consumer complaints. In recent years, aiming to monitor water quality,
multi-parameter sensors in WDNs have been developed. Water quality sensors that are
connected to the internet can be used to improve real-time monitoring of water quality. Because
specialized water quality sensors are often linked with high prices, only a few sensors are put at
strategic sites in the network in practice. This fact has prompted an extensive investigation into
water quality sensor installation in recent years. Low-cost sensors measuring water quality
parameters, such as oxidation-reduction potential, chlorine concentration, total organic carbon,
turbidity, conductivity, and pH, provide an alternative to special sensors for pollution detection.
In the past, various approaches were proposed to address pollution detection issues, including
single or multiple-type measurements that are analyzed separately or in combination from one or
more locations in the network, using model-based or model-free approaches. People who live in
areas where clean water is scarce might benefit from technology that can identify the level of
contamination in potential drinking water. This thesis, a deep learning approach to water quality
analysis that utilizes the power of modern hardware and machine learning techniques, was
created to provide this technology and quickly determine the level of contamination of water
based on an image. Turbidity is a calculation of a liquid's relative visibility. Turbidity measures
the concentration of light reflected by materials in the water; it increases when the materials
found in the sample water rise. Clay, silt, organic and very tiny inorganic matter, algae, plankton,
ingested colored organic compounds, and other microscopic species all contribute to the turbidity
of water. As a result, this study aimed to develop a model-based approach for pollution detection
throughout the water distribution system using the image of the turbid water that is closely
correlated with the physicochemical properties of other water quality parameters.

4
 1.3 Objective of the Research
This study’s general objective was to detect contamination of water in water distribution
systems.
1.3.1 Specific Objectives
 Describe the physicochemical properties of water in the water distribution system that is
used for drinking and other purposes.
 Explore changes based on the turbidity of water quality measurement in physicochemical
properties of water for contaminated and non-contaminated water
 Collect the data of water for both turbid and non-turbid water quality parameters and,
based on the data, assess the image data difference.
 Developing the detection model for selected water quality parameters
 Observing the special characteristics obtained from the model and analyzing the output
 Offering a contamination detection model for the water distribution system to reduce the
impact of water pollution

1.4 Scope of the Research

The research targets the detection of contaminated water in a water distribution system in Addis
Ababa, Ethiopia, giving special attention to Legedadi Dam and Entoto groundwater sources. This
site was chosen because most parts of the city get the water from the Legedadi Dam, and the
network it covers is larger compared to other water sources, which makes it an appropriate place
for the research. It also helped to collect more water samples that covered most areas of the city.
The focus of this research was to detect pollutants in water distribution systems.

1.5 Limitations of the Study

This study has provided essential features regarding contamination detection in water
distribution systems. However, there were some limitations recognized in this study.

First, the cameras used during the data collection period were Nikon d5200, Nikon d3200, and
mobile telephone cameras, which affected the quality of the turbid water image. If the image is
clear, the model identifies it correctly. It is because the image captured by the camera shows very
tiny particles found in the water. However, the camera used for this study has a quality limit. For
some images, similarities have been observed, especially in the case of low concentrations of
5
contaminated water images. Therefore, considering camera quality as a part of the problem, this
research suggests future studies consider the image quality at a low concentration of pollutants.

Secondly, the turbid water image from five NTU to 15 NTU is very challenging to identify its
type of contamination. In addition to that, the study did not include the turbid water image,
especially from five NTU to eight NTU. Therefore, to keep the balance and credibility of the
study, this research suggests future studies consider including the water image from five to eight
NTU.

The final limitation of this study was lacking observing all behavioral changes in the chemical
properties of the water during chemical pollution. It means that the study did not include all
changes in the chemical properties of the water. Some chemical pollutants may not affect the
turbidity of the water. Hence, future studies should consider all changes in the chemical
properties of the water and the turbidity of the water in the case of different chemical
contamination.

6
 Chapter Two
Literature Review

This chapter focuses on presenting the research's theoretical and experimental roots, as well as
incorporating them with the literature on water pollution in water delivery systems. Because of
limited water resources, stringent budget requirements, a growing population, aging
infrastructure, increasingly stringent regulations, and increased attention to safeguarding water
supplies from accidental or deliberate contamination, drinking water utilities are facing new
challenges in their real-time operations. Existing technologies, whether laboratory-based or
online monitoring systems, are too slow to develop operational responses and do not provide a
degree of public health safety in real time, necessitating the development of better water
monitoring systems. Because of the potentially serious implications for human health, rapid
identification (and response) to instances of contamination is important. Water is such an integral
part of life that developing a new approach for assessing water quality and forecasting potential
water quality patterns is critical.

The water delivery system has a wide body of work on simulation and model-based pollution
detection. The Environmental Technology Confirmation program was developed by the US
government in 1995 to investigate how changes in water quality parameters can be detected by
real-time sensors. Academics, government departments, and private companies are all doing
extensive research in close collaboration. On commercially produced hardware and software,
these projects range from theoretical to applied science. Klise and McKenna (2006) and
McKenna et al. (2008) planned three water quality detection methods based on a comparison of
calculated and observed values and receiver operating characteristic (ROC) curves to test each
technique. The US Environmental Protection Agency (EPA) has tested over 30 pollutants
(including pesticides, insecticides, metals, and bacteria) that may be used in deliberate acts of
water pollution (U.S.EPA, 2005a; U.S. EPA, 2005b). Hall et al. (2007) published the results of
experiments that looked at how commercial water quality sensors of various designs and
technologies responded to biological and chemical loads. The most responsive metrics to the
common pollutants were identified to have been free chlorine, total organic carbon (TOC),
oxidation-reduction potential (ORP), conductivity, and chloride, according to these studies.

7
Byer and Carlson (2005) were the first to develop and evaluate an online drinking water delivery
monitoring system. They mixed four credible hazardous drinking water pollutants (sodium
fluoroacetate, sodium arsenate, sodium cyanide, and aldicarb) with water from the tap and tested
the detection accuracy in a delivery system at various concentrations. Ph., chlorine residual, total
organic carbon, and turbidity values were measured before and after the introduction of these
pollutants using benchtop analysis and online monitoring equipment. All four pollutants may be
observed at low concentrations, according to the findings. Three of the four chemicals were
found in concentrations that were not harmful to human health.

Yang et al. (2009) used a pilot-scale pipe system to monitor 11 pollutants at various
concentrations. Background noise was minimized and contaminant signals were improved using
the proposed adaptive transformation of sensory measurements. Pollutant classification and
detection based on chlorine kinetics were made possible as a result of this.

Event detection systems were proposed by Arad et al. (2013), Perelman et al. (2012), and Murray
et al. (2010). These experiments also show a correlation between water quality indexes. Some
machine learning methods predict the expected calculated value of the next time phase by
learning the behavior of each parameter time sequence. The models can then classify
measurement outliers and define deviations from expected behavior. Most of the predictor
variables are combined to determine the likelihood of the risk occurring.

The system developed by Guepie et al. (2012) was based on residual chlorine decay. They
hypothesized that a contaminant in the WDS would absorb a large portion of the measured
chlorine and that this single parameter would provide enough information. The methods used in
the previous studies were supervised classification methods. According to a given training data
set, the classifier learns to differentiate between regular operation and event time measurements
using these methods. When there are no real-time measurements of contamination events, the
models must be trained and tested using simulated contamination events. The models add some
random disturbances to the measured data to reflect the contaminant effect to preserve generality
and then in the absence of sufficient information.

8
Depending upon pollutant reaction kinetics and uncertainty, Uber et al. (2007) presented
recommendations for event simulation. The foregoing studies applied simultaneous processes of
unvaried analysis to water quality time series to detect pollution events. Following the concurrent
methods, the results were combined to determine the likelihood of an occurrence occurring. A
multivariate analysis varies from previous research in that it requires the simultaneous
observation and analysis of several parameters.

Jonathan Arad, Mashor Housh, Lina Perelman, and Avi Ostfeld (2013) suggest a complex
threshold scheme for water delivery system contamination event detection. Off-line and web
phases are used to establish the technique. The algorithm of genetic (GA) is used to tune five
decision variables throughout the off-line stage: positively and negatively filters, positively and
negatively dynamic thresholds, and window size. Using the five design parameters, a recursive
Bayes' rule is used to detect actual online events during the online point. Both visual and
statistical signs of contamination incidents can be obtained using this technique. These
conclusions are focused on a single or several criteria that are regularly calculated.
Contamination incidents were simulated, but with different intensity severity effect
characteristics, since the exact behavior of pollutants in the actual water delivery system is
unknown.

The architecture and implementation of a coupled SVM-evolutionary optimization methodology

for water quality event detection in water delivery systems is proposed by (Nurit Oliker and Avi
Ostfeld, 2013). Since the suggested model is based on simultaneous conjunction analysis of the
entire water quality time series data set, it differs from parallel one-dimensional data analysis.
The multivariate analysis looks at the connections between water quality indexes and looks for
shifts in their patterns. All measurements were taken in a typical working setting. The model is
made up of time series data for six water quality indexes: total chlorine, electrical conductivity,
total organic carbon, temperature, pH, and turbidity, as well as simulated events. The model's
ability to identify events that were only partially reflected in the data is noteworthy (i.e.,
affecting only some of the measured parameters). In comparison to the previous model, the
proposed model has a higher accuracy and detection ability. A minimum volume ellipsoid is
developed and validated for the detection of outliers' measurements. The ellipsoid could take the
place of the current model's SVM point.

9
Nurit Oliker and Avi Ostfeld (2014) propose a two-step classification model for outlier
identification and event classification based on a minimal volume ellipsoid classifier and a
corresponding sequence analysis. The ellipsoid construction only uses normal vectors (measured
in normal operation time) in the known data set and does not include abnormal (events
representing vectors) examples. All measurements were taken in a typical working setting. Total
chlorine, electrical conductivity, pH, temperature, Total organic carbon, and turbidity are all data
that are similar to SVM-evolutionary data. The proposed model includes a full decision-making
method for detecting contamination events. The model also uses applied multivariate analysis of
the water quality metrics data, which differs from most previous studies' parallel analysis of the
parameters. The model produces strong results in terms of accuracy and identification, but it
should be noted that it is susceptible to any sudden changes.

Advanced strategies for improving source separation in a composite sensor signal and thus
improving sensor detection capability are discussed in a variety of publications (e.g., Choi et al.,
2005; Feintuch, 1990; Lin et al., 2007). Using advanced sensor data processing techniques or
evolving hardware, such as new generations of optical sensors for bacterial and chemical
contaminations, improves detection accuracy and contaminant selectivity in traditional water
quality sensors (e.g., U.S. EPA, 2005; NRC, 1995). Data analysis and other related statistical
techniques such as multi-variant cluster analysis (Klise and McKenna, 2006), adaptive data
treatment (Yang et al., 2007), and control chart-based robust statistics are used in the first
approach (Kroll and King, 2006). In pilot-scale pipe flow experiments, Szabo et al. (2007)
recorded observable sensor responses to the presence of multiple pollutants. The minimum
concentration threshold of pollutants above which their presence can be registered in water
quality sensors was stated by Haught et al. (2005) based on bench-scale research.

Yang et al. (2007) introduced an iterative detection system for actual occurrence adaptive
detection, recognition, and alert (READiw). In a field application, water quality measurements
are sent to a processing unit, which analyzes the sensor signals in real-time to distinguish
contaminants from natural history. A statistical comparison of new measurements to an
adaptively back-tracked backdrop in a moving time window governs real-time anomaly
detection.

10
The researchers have continued to work on developing pollutant detection systems that
incorporate multiple sensor responses. Kroll (2006) defined the Hach HST method, which uses
multiple sensors to detect events and identify contaminants. While replies from numerous
sensors are used in Kroll's system, their internal relationship is not explored. Arad et al. (2013)
provided an overall approach for detecting quality defects in water delivery systems using
multivariate water quality time series that combines a data-driven estimation model with
sequential probability updating.

Centered on the enhancement and development of a Dempster–Shafer (D-S) proof principle,

Dibo Hou, Huimei He, Pingjie Huang, Guangxin Zhang, and Hugo Loaiciga (2013) propose a
water-contamination event detector. The advanced D-S theory for the monitoring of water-
contamination events, according to their research, is based on time series of residuals of water-
quality variable predictions, including the use of weighted-averaging and time-dimension details
to address conflicts or ambiguities that occur when attempting to detect water-contamination
events. They tested potassium ferricyanide (K3 [Fe (CN) 6]) and ferric ammonium sulfate
(NH4Fe (SO4) 2), two toxic compound pollutants. Virtual and real-time controlled water-quality
data were used to evaluate the proposed process. The findings show that using actual water
quality monitoring; the suggested extended D-S fusion system can sense water-contamination
events with simulated strengths greater than 1.2. However, applying the future extended D-S
fusion approach to drinking water infrastructure in combination with actual early warning
systems would necessitate (1) the development of an accurate library of contaminant-sensitive
variables and (2) the integration of regular operational systems to minimize false alarms.

Eliades et al. (2014) looked into the issue of water quality by using a model-based method for
detecting pollution events in water treatment systems. The suggested system takes into account
well-known chlorine input signals and produces bounds of the predicted chlorine concentration
at different sensing locations at each time stage by running several Monte-Carlo simulations in
parallel with the real system. The sensor measurements are then compared to the approximate
limits, and if these bounds are violated in compliance with a guideline, event logic raises an
event warning flag. The result demonstrates that the system will adjust the detection bounds as
the chlorine concentration input varies. Chlorination is the method of adding chlorine to
drinkable water to disinfect it and destroy germs. It fixes one of many issues here.

11
Contamination, on the other hand, occurs for a variety of causes and has varying effects on
various parameters. As a result, detecting pollution based on chlorine concentration does not
provide a complete control mechanism.

From Kuhnert et al. (2014), an online security monitoring toolkit for a water delivery method
based on sensor measurements of water quality should be created. They use an easy-to-
parameterize event detection algorithm based on principal component analysis to implement the
proposed process. To calculate the threshold, the algorithm only needs two parameters: the
number of principal components and the sigma environment. They put the algorithm through its
paces for both clean and contaminated water. They also measure the water's conductivity, or the
concentration of ions in the water, here. Dissolved salts and inorganic materials such as alkalis,
chlorides, sulfides, and carbonate compounds have these conductive ions. Consequently, treating
one parameter of water quality affects the other, necessitating careful consideration.

(Mohammed et al., n.d.) Introduce Adaptive neuro-fuzzy inference system (ANFIS) models for
determining the protection of water in pipe networks when concentrations of water quality
variables in the pipes surpass their maximum thresholds. During the testing stage, the proposed
model will correctly detect between 92 and 96 percent of the water in the pipe network's safety
state, with a 1 percent false alarm rate. This work prompted me to redesign the algorithm for
certain water quality parameters that are highly correlated with one another and have an impact
on the algorithm's efficiency, i.e. some algorithms perform poorly on highly correlated
parameters. So it's intriguing to find out which parameters are closely correlated and then
remodel the algorithm.

Pollutant Detection Using Multiple Conventional Water Quality Sensors in an Early Warning
System for detecting the safety status of source water (Che & Liu, 2014). This approach
proposes a method for detecting pollutant events in a short period using data series from
multivariate parameters. Conductivity, UV-254, Ph., temperature, oxidation-reduction potential,
turbidity, nitrate-nitrogen, and phosphate are the eight parameters used in their proposed method,
and the most widely used herbicide, glyphosate, is chosen as the test contaminant. Numerous
sensors of water quality are presumed to react to a contaminant simultaneously in the proposed
process. Their findings indicate that with the use of feedback from online sensors of water

12
quality, the proposed method could detect glyphosate contamination 1 minute after the
contaminant was introduced at a concentration of 2 mg/l. The intensity of sensor correlative
reactions was also linked to pollutant concentration, and the patterns of sensor correlative
responses were contaminant-based. Since the orders of magnitude of the sensors' responses were
linked to contaminant concentrations, all associated sensors do not react at the same stage.

Christos C. Anastasiou, Marios M. Polycarpou, Christos G. Panayiotou, and Theofanis P.

Lambrou (2014) suggest the development and operation of such a low-cost device that can be
used in consumers' homes to continuously track qualitative water parameters and fuse multi-
parametric sensor responses to determine the risk of water consumption. They put their method
to the test by injecting different concentrations of two critical pollutants (Escherichia coli
bacteria and arsenic) at distinct time intervals and evaluating the efficiency of the event detection
algorithms in real-time. When microbiological pollutants (E.coli) were injected into chlorinated
potable water, the TU and EC sensors performed well. Almost all sensors in the case of
chemically (arsenic) contaminated drinking water did not react at low levels of arsenic
contamination. However, pH and ORP sensors responded well at concentrations greater than 25
g/L, and all sensors responded well at higher concentrations (greater than 500 g/L).

Mashor Housh and Avi Ostfeld (2015) propose an integrated logit model for pollution event
detection in water delivery systems, to improve an early event detection model called the
dynamic threshold method (DTM) by optimally incorporating detection from all different water
quality parameters into the event detection structure. Total chlorine (mg/L), electrical
conductivity (EC) (mS/cm), pH (_), temperature (_C), total organic carbon (TOC) (ppb), and
turbidity are the six water quality parameters they use to evaluate their device (NTU). The
developed approach is linked to data collected from a single monitoring station site, and it does
not take into consideration data from multiple sensors. However, although the model detects
events with greater likelihood, the system's ability to incorporate multiple sensor data into event
detection is constrained.

Shuming Liu, Han Che, Kate Smith, and Tian Chang (2015) created a real-time contaminant
classification system based on quantitative analysis of data from a variety of traditional water
quality sensors. This suggested approach is tested with data from laboratory pollutant dosing

13
experiments. Turbidity, temperature, oxidation-reduction potential (ORP), pH, conductivity,
phosphate, UV-254, and nitrate can all be measured simultaneously and continuously. They used
heavy metals (nickel nitrate, cadmium nitrate), herbicides (glyphosate, atrazine), and inorganic
salts to assess the three classes with the most common six toxic substances: herbicides (atrazine,
glyphosate), heavy metals (cadmium nitrate, nickel nitrate), and inorganic salts (sodium fluoride,
sodium nitrate). Additionally, clustering is used to describe the type of contaminant. And they
discovered that sensor response varied from contaminant to contaminant in their experiment. In
terms of categorizing the kind of pollutant in real-time, the proposed approach has great
performance and robustness. Whenever a glyphosate solution of 2.8 mg/l (4 times the national
limit) is detected, for example, the suggested method would accurately classify a pollutant within
4 minutes after detection, with such a true positive rate of 0.88.

Mohammadpour et al. (2015) used three separate algorithms, SVM, and two artificial neural
network approaches to investigate the issue of water quality. R2, RMSE, and MAE are used to
compare results. The SVM algorithm is competitive with neural networks in terms of results.

Pingjie Huang, Yu Jin, Dibo Hou, Jie Yu, Dezhan Tu, Yitong Cao, and Guangxin Zhang (2017)
propose a multi-classification probability output-based pollution classification system (MCPO).
This method takes a binary SVM and transforms it into a multi-classification SVM with a
likelihood of production. Throughout the process of developing a pattern library, the proposed
approach decreases the effect of pollutant concentrations. The proposed method will determine if
a sample feature is readily visible. A proposed model will determine if a sample feature is readily
visible. The samples with readily evident classification characteristics can decide their category;
the samples with non-apparent classification characteristics can avoid making a single
classification decision. As a consequence, samples with a broad maximum likelihood and
standard deviation have a distinct classification function, resulting in a highly credible single-
classification result. The classification function is not as evident in the early stages of
contaminant injection in samples with limited maximum likelihood and standard deviation.
However, for certain pollutants, high-concentration samples will accurately represent the
properties of material while having little impact on the construction of a classification decision
boundary.

14
Yue, Wong, Zhu, and Zhang (2017) designed a lightweight anomaly-finding methodology for
water quality parameters that utilizes dual time-moving windows to distinguish anomalous data
from historical trends in real time. The algorithm uses statistical models, specifically the
autoregressive linear combination model. They put the algorithm to the test using 3-month PH
data from a real water quality monitoring station in a river system. Experiments show that their
algorithms have a lower rate of false positives and do better than the AD and ADAM algorithms
at detecting anomalies.

Xie, Gao, and Kang (n.d.) have created a model for predicting water quality. Artificial Neural
Network with Nonlinear Autoregressive showed the best results. In 2009, Xiang & Jiang used a
combination of least squares support vector machine (LS-SVM) and particle swarm optimization
methods to predict water quality and overcome the limitations of traditional back-propagation
algorithms, such as being moderately difficult to reach and easy to achieve the extreme minimum
value. They also realized that the model is very good at determining the quality of the water of
the Liuxi River Xiang and Jiang via simulation testing (2009). Recently, the Recurrent Neural
Network has been used in large-scale vision and speech problems as a dynamical device whose
next state and output are dependent on the current network state and input (Gregor, Danihelka,
Graves, Rezende, & Wierstra, 2015). RNNs and LSTMs excel at extracting patterns from input
feature space, even when the data spans long sequences. They can nearly seamlessly system
problems with numerous input variables, which is particularly useful in time series forecasting,
where traditional linear models can be difficult to adapt to multivariate or multiple input
problems (Che, Purushotham, Cho, Sontag, & Liu, 2018).

Revathi, S. Kavi, and G. Shenbagalakshmi (2018) suggest using a wireless sensor network to
develop and implement an actual system of water quality control for drinking water. The
pollution finding algorithm and fuzzy rules assist in defining and classifying water based on
pollution in the pipeline. The smart sensors' low-frequency communication connection is a pipe-
to-pipe data transfer that is also long distance. Signal attenuation is a significant factor in signal
propagation through a drenched medium at high frequencies. Therefore, the key challenge is
wireless communication between the sensors and signal transmission from the sensors. The
proposed system is low in cost, lightweight, and consumes little power.

15
Jeniffer Jepkoech and Consolata Gakii (2019) created a classification model based on a decision
tree to analyze water quality and determine whether it is safe or not. Something like a decision
tree as just a data mining technique to predict clean water based on water quality parameters will
help the laboratory technologist's job by predicting which water samples can go on to the next
stage of analysis. Conductivity, pH level, and water alkalinity analysis can all help to determine
water quality. The model was developed using five decision tree classifiers: Random forest,
LMT, J48, Hoeffding tree, and Decision Stump, and the precision was compared. Decision
Stump had the lowest accuracy of 83 percent, while the J48 decision tree had the highest
accuracy of 94 percent.

Florin Leon, Doina Logofătu, and Fitore Muharemi (2019) suggest finding the best performing
model for water quality data for event detection on time series data, to find the best model for
anomaly detection on water quality systems. To see if machine learning models are more
accurate than logistic regression, they use the F-score metric to evaluate the output of each
algorithm. For just a long period, time series have been studied using statistical algorithms.
Machine learning Algorithms are increasingly common for their ability to perform well on time
series data sets. The results indicate that the whole algorithms are vulnerable, with ANN, logistic
regressions, and SVM being the least vulnerable, while LSTM, RNN, and DNN being the most
vulnerable.

Automated Aqua Sight water pollution detection, proposed by Elliott Ruebush and Ankit Gupta
(2019), uses a picture to assess the level of contamination in water. Believe that determining the
turbidity of a water picture is a viable method for determining the risk of consuming that water.
They have used a convolutional neural network to do this, which involves an image of water to
decide whether or not pollutants are present. They checked the algorithm and found it to be
accurate 96.2 percent of the time. However, the dataset they used to arrive at this conclusion
contains only 105 water images, and the data is not balanced, i.e. they only focus on the physical
properties of the water and they are not considering other properties. They concentrate on the
physical properties of the water, which implies that the contaminants applied to the water to alter
the turbidity in order were: sand, sand and salt, sand, salt, and black pepper, sand, salt, black
pepper, and oil-paint. The chemicals, however, affect the turbidity of the water. As a

16
consequence, during the data collection process, the chemical properties of the water are ignored.
Finally, all of the differences have an impact on the model's preparation and outcome.

17
 Chapter Three
Research Methodology
3.1 Introduction
This chapter focuses on the methodology that the study followed in the research. It begins with a
brief description of the research area, an account of the research participants, and the techniques
used to recruit follows. The chapter also makes a brief description of the research methods used
to gather the necessary data and achieve the research objective. The chapter also briefly
describes the method employed for analyzing the collected data. In the end, the ethical issues are
presented.

3.2 Research Site

Ethiopia is one of Africa's largest countries, with more than a population of 110 million people
and a capital city of about 6 million people, Addis Ababa (UN, 2017). The capital city, Addis
Ababa, is situated on the banks of the Akaki River. Entoto Mountain Range forms the town's
northern border. Mt. Furi, to the southwest of a town, and Mt. Yerer, to the southeast, are high
massive volcanic centers rising to elevations of 2,839 and 3,100 meters, respectively. The east
and west drainages that divide the Akaki River are formed by such two peaks.

The location of town is in the highlands, at a height of 2,200 to 2,500 meters above sea level.
The city's river network can be divided into two catchments: the Great Akaki catchment (900
km2) and the Little Akaki catchment (540 km2), all of which drain into the lake Aba Samuel
(Aschale et al., 2017). The River of Kebena and the Great Akaki River are two main river
branches that originate from the Great Akaki catchment, each with its network of smaller
tributaries. The River of Kebena flows through the city's densely populated areas before joining
the Great Akaki River, creating a sub-basin of the Great Akaki catchment. Apart from the
Kebena River, the Kebena catchment has only one major river, the Bantyketu.

As the Addis Ababa Observatory, the long-term average annual rainfall is 1254mm. The
maximum temperature fluctuates between 20° C in the wet season and 25° C in the dry season,
while the minimum temperature fluctuates between 7° C and 12° C all year. In the county, the
speed of wind is usually moderate, with average values ranging from 0.5 to 0.9m/s. In November

18
and December, the average sunshine hours per day reach 9.5 hours, while in July and August,
this number drops to 3 hours or less. Monthly pan evaporation records collected from previous
studies show that the average monthly pan evaporation at Addis Ababa Observatory during the
dry season (November) is around 180 mm and this value decreases to around 75 mm during the
wet season (July).

The city is rapidly expanding because of urbanization, and it is undergoing massive

infrastructure growth. There are over 2,000 industries in Addis Ababa, including potable water,
cement, clothes, beer and alcohol, tobacco, leather, tannery, rubber, and food factories. The city
serves as the country's industrial, cultural, governmental, commercial, and modern nerve center
(Aschale, 2016). With its many foreign organizations and agencies, the city is also one of
Africa's hubs. Addis Abeba is the center for the United Nations Economic Commission for
Africa (ECA) and over a hundred embassies. It is said to be Africa's diplomatic capital and a
symbol of the continent's current humanitarian development.

Addis Ababa was chosen partly because of the natural springs on the site when it was created. As
the city expanded, new water sources were required. Therefore, in 1938, a plant was built to treat
water from the nearby rivers. The city began to develop, and the first dam (Gefersa I) was
installed in 1944 to meet the increased demand. Many springs degraded in quality during the
1950s and were withdrawn from service. The Gefersa treatment plant, which has a capacity of
30,000 m3/day, was constructed in the 1960s and another dam (Gefersa III) was built in 1966 to
increase capacity and serve as a sediment trap. Two transmission pipelines, each 400mm in
diameter, ran from the Gefersa treatment plant to Addis Ababa's distribution reservoirs.

The Legadadi Dam, treatment plant, and transmission pipeline were completed in 1970. The
water distribution system for this lower dam required a pumping system. The Legadadi treatment
plant's capacity was increased from 50,000 to 150,000 m3/day in the 1980s, and a second
transmission pipeline was constructed. Due to Addis Ababa's population, the engineers knew that
these sources would only be adequate until 1992. They looked at various locations for the next
great dam, which would supply water until 2020. This project was delayed (and is still being
postponed) due to numerous political and economic instability and two emergency projects, the
building of the dire dam and the Akaki area, had to be implemented instead.

19
The Legadadi treatment plant was expanded in tandem with the building of the Dire dam,
reaching its current capacity of 165,000 m3/day (Sime et al., 1998). Its capacity has increased to
195,000 m3/day. The Akaki well field has a 43,000 m3/day range. The American Association of
Women's Sports Administrators (AAWSA) published a report in 2011.

The city's factories are situated along the River of Little Akaki and its tributaries, which
influence the river. While the Great Akaki's catchment area is sparsely populated, the upstream
areas are highly populated with residential and commercial areas. Besides, it passes through
agricultural fields before joining the Kebena River. Before leaving the city and entering another
agricultural area, the river flows through a denser city area with industries present. Aba Samuel
is a lake about 53 kilometers from Addis Ababa's city center. It was designed in the late 1930s as
part of a dam project to produce electricity. Now, it is highly contaminated by industrial and
urban waste wastewater Little Akaki or the Great Akaki (Yohannes & Elias, 2017).

Residential, leisure, urban agriculture, manufacturing, commercial, open-market, and green areas
are all part of the city's land use. Many riverbank areas and surrounding areas in the city have
settlers living there without permission. In Addis Ababa, illegal settlement is one of the problems
and the river is used as a garbage dump and toilet, and a washing machine. Domestic waste is
discharged into rivers without being treated, and pipelines from different sources have outputs
that discharge directly into rivers (Yohannes & Elias, 2017).

3.3 Data collection tools

This research used several techniques to gather the data needed to fulfill the study's target. The
goal of combining several research instruments to investigate the same phenomenon made it
crucial to deepen our understanding of how to regulate water quality in water delivery systems.
Furthermore, combining multiple research tools allowed researchers to triangulate their findings
and compensate for the shortcomings of one approach with the strength of another. As a result, in
this study, data collection instruments such as laboratory tests, secondary data sources and the
researcher's field notes were used to gather the information needed to achieve the goal of the
research.

20
 3.3.1 Secondary Data Sources
The study also collected secondary data from a variety of sources, including books, journals,
research papers, and other related unpublished documents, such as information from the Addis
Ababa Water and Sewage Authority (AAWSA). It focused on the scope of the water pollution
issue and assessed whether any intervention initiatives were implemented to address it. The
opportunity exists via evaluating historical information regarding water quality and pollution
events that occur during various periods and their cumulative effect on human health.

3.4 Method of Data Analysis

The first step in data processing is to obtain water samples from various locations. The most
important aspect of data processing was making a list of coding categories. This was achieved by
searching for trends and continuity in the collected data and recognizing the themes.
Furthermore, coding categories were generated using keywords and phrases that reflect the
topics and themes. Creating a list of themes and subthemes made processing the collected data
much easier, which was essential for the data analysis. The meanings of the code were listed and
then grouped into common categories or themes. The significant themes of the phenomena were
then developed using the clustered themes and meanings.

3.4.1 Collection of water samples

From various locations, a total of 216 water samples were collected (Figure 3.1-3.3). The
samples are collected throughout different phases of a treatment plant, including raw water,
during sedimentation, and finally from the reservoir, city reservoir, and household. Parts of the
city where water pipes were exposed to domestic waste were chosen as sampling sites. Another
criterion used in the selection process was proximity to contamination sources like industrial
wastewater and hospital effluent outlets. Since the water pipe could not be accessed anywhere,
locations had to be selected where it would be convenient to take a sample if it broke down due
to construction or some other reason.

The sampling took place over eleven months in 2020, from January 8 to December 12. The
majority of samples were taken in the morning. The collection protocol followed the Addis
Ababa Water and Sewage Authority's procedures (AAWSA). In summary, the 600 ml sampling
bottle was used to extract water from the sample points in both clean and contaminated

21
situations. Water samples were directly stored in a cooler after being sampled and kept
refrigerated in the laboratory to avoid any reactions or microbiological activity.

Physicochemical parameters were studied in the samples (Turbidity, pH, EC, TDS, Total
Alkalinity, Calcium Hardness, Total Hardness, Magnesium Hardness, Ammonia, Nitrate, Nitrite,
Phosphate, Fluoride, Iron, Manganese, Silica, Chloride, and Bicarbonate Alkalinity). Following
the sampling, all measurements were carried out in line with WHO requirements (WHO). All
sample processing is carried out at the Addis Ababa Water and Sewage Authority's regular
laboratory (AAWSA). Appendix C and D contain the results of all samples taken from a real
network of water delivery systems for both safe and polluted water.

Figure 3. 1 The raw water at the time of sampling

Figure 3. 2 Water Reservoir at the time of sampling.

22
 Figure 3. 3 Customer Tap water at the time of sampling.

However, laboratory findings indicate that most physical and bacteriological parameters have
changed in samples taken from various points along with the water distribution network.
Chemical parameters change just slightly. It is not to say there is no chemical pollution. This
analysis is focused on physicochemical parameters and excludes bacteriological findings. As a
result, it is risky to add chemicals to the water distribution line, to better understand chemical
contamination, so it is important to consider changing the physical and chemical parameters by
simulating the drinking water line and incorporating different chemicals that are harmful to
human health.

3.4.2 Preliminary-scale contaminant injection and monitoring system

The preliminary-scale device used in this study is a water delivery system simulator to learn how
physicochemical parameters change when harmful chemicals are injected into the system. Figure
3.4 illustrates the process flow of the Preliminary-scale method, which was used for baseline
establishment and multi-pass contaminant checks. The water tank is about 60 cm high with 40
cm in diameter, with a total capacity of 20 L. 20L of water supply flows via the multi-direction
to the end line in this mode. A screwed valve is located between the source and the end line and
used to inject a chemical into the device. Finally, a 600 ml sampling bottle was used to extract
the sample for all types of contamination, and it was weighed in the laboratory in the same way
as the previous sample and the adjustment of parameters for the proposed method was studied.

23
 Figure 3. 4 A process flow schematic of the Preliminary-scale system

The system can measure the following eight parameters: pH, turbidity, TDS, Total Alkalinity,
Calcium Hardness, iron, manganese, Chloride, nitrate-nitrogen, and Silica.

3.4.3 Developing the dataset

We looked at the overall method of collecting data in the previous portion. In addition, lab tests
were conducted on samples from different distribution lines, including clean and polluted water
samples, to determine their physicochemical properties in detail. The outcome, on the other
hand, is not the target in and of itself, but it does point in the right direction. Designers should
consider changes throughout the physicochemical properties of water, concentrating on the
turbidity of the water that is influenced by the chemical and biological particles of the
contaminant, based on the findings of the previous parts. Turbidity is a term that refers to the
cloudiness of water caused by suspended particles like clay and silts, chemical precipitates like
manganese and iron, and organic particles like plant debris and species. As a result, a dataset for
the proposed method is generated based on the properties of the turbidity of the water. As the
turbidity of the water rises, it decreases the purity of the water to transmit light by scattering and
absorbing the light. A total of 216 collected water samples image were used in the proposed
system's dataset. But it doesn’t mean the whole dataset is only 216 water images. It was divided
into three parts. The first 121 images contained both safe and contaminated water in various
forms. The second 69 images were all of the clean water, and they added a variety of light levels
to the dataset. A total of 26 images were added to the dataset from the third round of data
collection, all of which were of polluted water with a broader variety of light levels. The

24
proposed approach considers different effects of light on the collected water to create the image
dataset based on the turbidity of the water.

From those methods:

 The image is taken with 5 colored lights: light Blue, Blue, Green, Yellow, and Red.
 Image was taken with light-Off
 Image taken with varying levels of brightness Images
 taken at different positions

(a) (b)

Figure 3. 5 (a) is clean water with normal light and (b) is clean water with blue light.

(a) (b)

Figure 3. 6: (a) is contaminated water with 100mg of sodium nitrate and (b) is contaminated water with
400mg of sodium nitrate.

25
 (a) (b)
Figure 3. 7: (a) is contaminated water with clay at low concentration and (b) is contaminated water with
clay at high concentration.

(a) (b)
Figure 3. 8: (a) is contaminated water with sand at low concentration and (b) is contaminated water with
sand at high concentration.

(a) (b)
Figure 3. 9: (a) is contaminated water with silt at low concentration and (b) is contaminated water with
silt at high concentration.

26
 Figure 3. 10: Contaminated water with 34mg/l of Iron.

3.5 Model Design

Today's water supply companies and people in the city need a more advanced and easy-to-use
and access purified water management system that will benefit from technology that allows them
to assess how polluted potential drinking water is. This research proposes a deep learning
approach to water quality analysis that utilizes the power of modern hardware and machine
learning techniques to provide this technology and quickly determine the degree of pollution of
water based on a picture.

3.5.1 Architectural design of the Proposed System

Machine Learning (ML) algorithms are well-known for learning the underlying relationships in
data and making decisions without the need for explicit instructions. Convolutional neural
networks (CNNs) are a form of neural network that is used in deep learning. CNNs are large
networks of nodes called “neurons” that create connections as they learn from data. CNNs are
one of the most powerful learning algorithms for comprehending image information, with
excellent results in image segmentation, classification, identification, and retrieval tasks (Ciresan
et al. 2012; Liu et al. 2019).

The ability of CNN to manipulate spatial or temporal similarity in data is one of its most
appealing features. CNN is divided into several learning levels, each of which consists of a
mixture of convolutional layers, nonlinear processing units, and subsampling layers (Jarrett et al.
2009). CNN is a feed-forward multilayered hierarchical network in which each layer performs
multiple transformations using a bank of convolutional kernels (LeCun et al. 2010). The
convolution operation aids in the extraction of useful features from data points that are globally

27
correlated. The non-linear processing unit (activation function) receives the output of the
convolutional kernels, which not only aids in learning abstractions but also embeds non-linearity
in the feature space. This non-linearity results in various activation patterns of different
reactions, making it easier to understand semantic differences in images. Subsampling is
commonly applied to the output of the non-linear activation function, which aids in the
summarization of the results while also making the input invariant to geometrical distortions
(Scherer et al. 2010; LeCun et al. 2010). The ability of CNN to automatically extract features
eliminates the need for a separate feature extractor (Najafabadi et al. 2015). As a result, CNN can
learn a decent internal representation from raw pixels without undergoing extensive processing.
Learning with Hierarchical, multi-tasking, automatic feature extraction, and weight sharing are
among CNN's interesting features (Guo et al. 2016; Liu et al. 2017; Abbas et al. 2019).

During preparation, CNN learns by controlling the shift in weights according to the goal using a
back-propagation algorithm. The optimization of an objective function with a back-propagation
algorithm is close to the human brain's response-based learning. Deep CNN's multilayered,
hierarchical structure allows it to retrieve high-, mid-, and low-level features. Mid and low-
level features are joined to make high-level features (more abstract features). CNN's hierarchical
feature extraction ability mimics the Neocortex throughout the human brain's deep and layered
processes of learning, which dynamically learns features from raw data (Bengio 2009). CNN's
fame stems largely from its ability to collect hierarchical features.

3.5.1.1 Basic CNN components

CNN can learn representations from grid-like data, and it has recently demonstrated significant
performance improvements in a variety of machine learning applications. Figure 2 depicts a
typical block diagram of an ML system. A typical CNN architecture consists of pooling and
convolution layers alternated with one or more completely connected layers at the end. A fully
connected layer may be replaced with a global average pooling layer in some cases. Various
regulatory units like dropout and batch normalization are also implemented to improve CNN
performance, in addition to various mapping functions (Bouvrie 2006). The arrangement of CNN
components is critical during the implementation of the new design architecture to achieve
improved performance.

28
 Figure 3. 11 Basic layout of a typical ML system has several stages. (Source: A Khan et al. (2020)).

While they could be used with three-dimensional data and one-dimensional, CNN is a special
form of neural network model designed for interacting with two-dimensional image data.

3.5.1.2 Convolutional layer

The convolutional layer gives its network its name and serves as the heart of CNN. This layer
executes a process called "convolution." Every neuron performs as a kernel throughout the
convolution layers, which would be made up of a collection of convolution kernels. The
operation of convolution becomes a correlated operation if the kernel is symmetric. The
convolutional kernel divides the image into small slices, known as receptive fields, as shown in
fig. Extracting feature motifs is easier when an image is divided into small blocks. Like an older
neural network, a convolution involves a linear operation multiplying a set of inputs with the
weights. When multiplication is performed with an array of input data and the weights of a two-

29
dimensional array, called a kernel or a filter, the method was developed for two-dimensional
input (Khan et al., 2020).

Figure 3. 12 Architecture of the Convolutional Neural Network.

Convolution operations can be expressed as follows:

Flk (p, q) = ∑𝑐 ∑𝑥,𝑦 𝑖 c(x, y).elk (u, v) (4.1)

Where, ic(x, y): an element of the input image tensor

Ic,: is element-wise multiplied by elk(u, v)

elk(u, v): index of the kth convolutional kernel kl of the lth layer.

Whereas the output feature-map of the kth convolutional operation can be expressed as Flk = [flk
(1, 1)… flk (p, q)… flk (P, Q)].

The input data is greater than the filter, and the dot product is used to multiply a filter-sized patch
of the input with the filter. A dot product is the element-wise multiplication of the input and
filter's filter-sized patch, which is then averaged, always yielding a single value. The operation is
often referred to as the "scalar product" because it produces a single value.

Consider a 5 x 5 image with pixel values of 0, 1 and a 3 x 3 filter matrix, as shown below.

30
 Figure 3. 13 3x3 Output matrix

Various kinds of features inside of an image can be extracted by sliding kernels with the same set
of weights on the image, making CNN parameters more powerful than fully connected networks.
Convolution operations can be classified based on the number and size of filters, the type of
padding, and the direction of convolution.

 Strides

The number of pixels shifted over the input matrix is referred to as the stride. When the stride is
set to 1, the filters are moved one pixel at a time. We switch the filters two pixels at a time when
the stride is two, and so on. Convolution will function with a stride of 1, as seen in the diagram.

 Padding

Filters do not always perfectly fit the input image. There are two possibilities:

 To make the image fit, pad it with zero (zero padding).

 Remove the section of the image where the filter didn't fit. This is referred to as valid
padding, and it preserves only the valid portions of the image.

3.5.1.3 Pooling layer

Feature motifs, which appear in the image as a result of the convolution process, may appear in a
variety of places (see Fig. 1). Once features have also been retrieved, their actual position is less
important as long as their relative position to others is preserved. Down-sampling, also known as
pooling, is an intriguing local process. It compiles similar data in the general neighborhood of
the receptive field and outputs the dominant response for that area.

Zik = gp (Flk) (4.2)

31
Equation (4.2) depicts the pooling process, where Zik denotes the pooled feature-map of the lth
layer for the kth input feature-map Flk, and gp (.) denotes the pooling operation form. The pooling
operation aids in the extraction of a set of features that are insensitive to translational shifts and
minor distortions. Eliminating the feature-map size to an invariant feature set not only governs
the network's complexity, but it also aids generalization by reducing over-fitting.

Spatial Pooling can be of different types:

 Max pooling
 Average Pooling
 Sum Pooling
Max pooling takes the largest element from the rectified feature map. Taking the largest element
could also take the average pool. The sum of all elements in the feature map is called sum
pooling.

Figure 3. 14 Max Pooling.

3.5.1.4 Activation function

The decision-making function is an activation function that aids in the learning of complex
patterns. The use of the right activation function would speed up the learning process. Equation
(4.3) defines the activation function for a convolved feature map.

Tik = ga (Flk) (4.3)

Where Flk: is a convolution output that is allocated to the activation function

ga (.): introduces non-linearity and returns a transformed output
Tik: for the lth layer.

32
To inculcate non-linear combinations of features, various activation functions such as sigmoid,
tanh, maxout, SWISH, ReLU, and variants of ReLU such as leaky ReLU, ELU, and PReLU are
used. ReLU and its variants, on the other hand, are preferred because they help overcome the
vanishing gradient problem. For a non-linear operation, ReLU stands for Rectified Linear Unit.
The output is ƒ(x) = max(0, x). Why ReLU is important: ReLU’s purpose is to introduce non-
linearity in our ConvNet. Since the real-world data would want the ConvNet to learn would be
non-negative linear values.

3.5.1.5 Batch normalization

The issue of internal covariance change within feature maps is addressed using batch
normalization. The internal covariance shift is a change in the value distribution of hidden units
that delays convergence (by pushing the learning rate to a small value) and necessitates cautious
parameter initialization. In equation (4.4), batch normalization for a transformed feature-map Flk
is shown.

(4.4)

Where, Nlk: represents a normalized feature-map

Flk is the input feature-map,
µB and 𝜎B2 depict the mean and variance of a feature map for a mini batch
respectively.

To avoid division by zero, 𝜀 is added for numerical stability. By setting the distribution of
feature-map values to zero mean and unit variance, batch normalization unifies the distribution.
Furthermore, it smoothest the gradient flow and serves as a controlling factor, thus assisting in
the network's generalization.

3.5.1.6 Dropout

Dropout adds regularization to the network, which increases generalization by skipping certain
divisions or connections at a certain probability at random. Multiple connections learning a non-
linear relation in NNs are often co-adapted, resulting in over-fitting. This random dropping of
certain connections or units results in several thinned network architectures, from which one

33
representative network is chosen with small weights. After that, the chosen architecture is used to
estimate all of the suggested networks.

3.5.1.7 Fully connected layer

The fully connected layer is typically used for classification at the network's end. It is a global
operation, unlike pooling and convolution. It takes data from the feature extraction stages and
analyzes the output of all the layers before it. As a result, it creates a non-linear combination of
selected features that are used to classify data (Khan et al., 2020).

3.5.2 Experimental Setup

A collection of tools and libraries is required to implement the framework.

 Python
 TensorFlow: Machine learning models can be implemented, trained, and deployed using
this open-source platform.
 Keras: An open-source library that can be used to implement neural network architectures
on both CPUs and GPUs.
 Pandas: Data analysis and modification library.
 Matplotlib: In Python, this tool is used to generate visualization plots such as charts,
graphs, and more.
 Numpy: Array data structures can be used to perform a variety of mathematical
computations and operations.

After installing all of the necessary software and libraries, the data collected in the previous
section must be classified into chemical, clay, clean, sand, and silt and vegetable categories.
Image processing techniques are used to detect contamination in water distribution systems in
four phases: image acquisition, pre-processing (noise removal and image enhancement), image
segmentation, and image analysis. Each of the foregoing processes has been improved to
produce a high-quality image. Figure 3. 15 depict a typical image processing block diagram.


IMAGE IMAGE IMAGE IMAGE
ACQUISITION PREPROCESSING SEGMENTATION ANALYSIS

Figure 3. 16 Image processing and analysis.

34
Image Acquisition: This step is crucial and is regarded as the foundation of a proper image
processing system. Three processes are involved in the acquisition: the region of interest object
reflection, an optical system for focusing the energy, and sensors for detecting the amount of
energy. After that, the image is saved on a storage medium. The dataset, which includes 2,000
images of water, was generated by the researcher. Of the 2,000 water images, 71% of the images
are collected from sampled water, and another 29% of the images are collected from different
sources. It comprises 1,600 training image datasets and 400 testing image datasets, all of which
are colored images of various scale measurements grouped into five groups. However, the entire
set of categorized images is not numbered, making it difficult to proceed to the next level.
Labeling usually takes a collection of unlabeled data and adds meaningful, descriptive tags to
each piece of that unlabeled data. As a consequence, any image in the category must be labeled.

Image Preprocessing- After sample collection, preprocessing ensures that the sample is clean
and devoid of unnecessary deformities, as well as improves the sample's critical parameters for
subsequent processing. Enhancement of the image is the process of increasing the visual appeal
of an image by altering the pixel brightness. It consists of approaches for transforming images
into a more understandable format for both humans and machines. Image enhancement focuses
on and improves the image's certain features or attributes rather than increasing the image's
intrinsic details. Based on the fact that once all images are labeled, the next part that is important
for the implementation is selecting a standardized image size. The images will have to be
reshaped before modeling so that all images have the same shape. This is often a small square
image. If the source data has a variety of aspect ratios, some portraits, some landscapes, and the
target object is typically in the center, a square crop from the middle may be a good compromise.
There are many ways to accomplish this, but the most popular is to use a simple resize operation
to stretch and distort an image's aspect ratio and push it into a new shape. When loading all of
the images, I could examine the distribution of picture widths and heights, after which I created a
new image size that better represents what it is most likely to see in person. Smaller inputs
indicate a faster-to-train model, and this is generally the driving force in image size selection. In
this case, the method is used and set to a fixed size of 256x256 pixels, which can be found in
Appendix A code 1. The next step in the preprocessing phase is data cleaning and data
augmentation.

35
Data cleaning is the process of creating a homogeneous input data set to prevent the model from
being biased against some misinterpretations of images during training and therefore always
predicting certain categories.

Image Data Augmentation: Image Data Augmentation is a technique for artificially increasing
the size of a training dataset by creating altered versions of the images in the dataset. Deep
learning neural network models that are trained on more data become more skilled, and
augmentation techniques may generate variations of the images that enhance the fit models'
ability to generalize what they've learned to new images. Data augmentation can also be used as
a regularization strategy by introducing noise to the training data and allowing the model to learn
the same features regardless of where they appear in the input. Small shifts and horizontal flips
to the input images of clean and contaminated water may be useful for this issue. The
ImageDataGenerator used for the training dataset will take these augmentations as arguments.
Since we want to assess the model's output on the unaltered images, the augmentations should
not be used for the test dataset. In this case, images in the training dataset will be augmented with
small (10%) random horizontal and vertical shifts, rotated with (30%), zoomed with (20%), and
random horizontal flips that create a mirror image of a photo, which can be found in Appendix A
code 2. Images in both the training and test steps will have their pixel values scaled in the same
way.

Image Segmentation Process: - The process of breaking an image into smaller portions, or
segments, is known as image segmentation (Figure 3. 17). The approach is used to isolate an item
or region of interest (ROI) from the visual background, allowing for easy assessment of the most
essential and meaningful region. In addition to that, defining image properties is important.

36
 Figure 3. 18 Segmentation Process

Since data augmentation increases the training data, the image size has been redefined, which
can be found in Appendix A code 3. This has an impact on the model's training time and requires
more hardware resources. There are also different phases in the segmentation process.

Data preparation is an iterator that must load images for a single dataset in a sequential manner.
It also helps to customize more information on how images are loaded. The "target size" claim is
noteworthy because it allows you to resize all images to a particular size, which is often needed
when modeling. Furthermore, the "class mode" argument allows you to determine the type of
classification task, which can be either "binary" or "categorical," which is a multi-class
classification. This is a multi-class classification analysis. When practicing, the default "batch
size" is 32, which means 32 randomly selected images from around the dataset's classes will be
returned in each batch, which can be found in Appendix A code 4. It is also possible to save
augmented images to a file at the same time. Keras performs the data processing and
augmentation in real-time. This is good for recall, but it might need the same images used during
training. For example, you might want to use them later in a different software package, or you
might just want to generate them once and use them on several deep learning models or
configurations.

Create the model: - The model is made up of four convolution blocks, each with a max pool
layer. A relu activation feature is used to enable a completely connected layer with 512 units on
top. This model hasn't been fine-tuned for high precision.

37
Image Analysis: - The method of gathering useful information from a digital image is known as image
analysis. Scene analysis, picture description, image interpretation, pattern recognition, and
computer/machine vision are examples of automatic or semi-automatic methods for extracting
information.

3.6 Ethical Consideration

Ethical questions arise often in all research activities involving a wide range of individuals.
Nonetheless, because of the sensitivity and ethical complexities of researching certain research
topics, researchers are more careful. To ensure voluntary informed consent and protect the
confidentiality and privacy of all companies and persons concerned, sampling and data collection
must be conducted with caution. This was primarily done to safeguard the researcher's
emotional, physical, and mental well-being. As a result, ethical values such as obtaining consent,
maintaining privacy, and confidentiality were heavily considered during the research process.

3.6.1 Informed Consent

Securing the consent of participants and organizations with a stake in and responsibility for the
potential research participants with whom the researcher aims to perform the study is one of the
fundamentals of research activities, especially fieldwork. Informed consent means that a study
participant has decided to participate in research after learning about and knowing the researcher,
the research's goals, methodology and procedures, subjects, data use, and the ability to withdraw
from the study at any time. However, obtaining consent entails not only providing facts but also

38
doing so in terms that participants understand. Potential research participants were told about the
overall aspects of the study, including the study's purpose, methodology, and procedures, as well
as their right to withdraw from the study at any time if they felt uncomfortable.

3.6.2 Respecting for Person

In this research, different participants are involved, including Addis Ababa Water and Sewage
Authority, (AAWSA). This means, that individuals (the primary research participants) were
contacted and then asked to provide personal experiences and views on water quality as well as
the Authority data used for the research. Accordingly, privacy, avoiding undue interference in
personal affairs, was taken as a vital ethical issue that should be considered in this research
activity. This included recognizing the rights of research participants to control all personal
information they provide. The standards focused on protecting participants from any harm during
the research, ensuring that consent was obtained, and then the risks posed were modest. The
dignity of all research participants and the privacy of the institution were respected. Thus,
adherence to this principle ensures that people are not used simply as a means to achieve
research objectives. Therefore, in this study, the research participants were informed not to talk
and expose matters that they have made private to avoid negative feelings for themselves and
others as well. Furthermore, as part of the research responsibility, the research confines itself to
social and ethical principles so that it will not infringe on the private matters of the research
participants and the Authority.

39
 Chapter Four
Result and Discussion

The outcomes of a proposed technique are discussed in this chapter. First, we look at the overall
model accuracy of the proposed architecture's training and validation results. Second, we show
how different architectures were optimized to analyze the model that was used to find the best
accuracy during the validation data identification phase.

4.1. System Models Analysis

The error rates are significant indicators in assessing the model's success in the proposed
architecture. The accuracy indicates the likelihood that the picture will fit the target mark
correctly. The accuracy level of the model varies throughout the training cycle due to the dataset
used for training and validating the model. In 100 epochs, the model was conditioned. The batch
size is 32 files. The overall accuracy of the training was 95%.

Figure 4. 1 Accuracy, loss, validation accuracy, and validation loss for the proposed model.

The model which is created in the previous chapter is made up of 4 convolutional layers, 4 max-
pooling layers, and 1 dense layer is used. The accuracy of validation is 80.49 percent. This
means we can correctly classify 80.49 percent of the images in the validation collection that the
model missed.

40
 Figure 4. 2 Training and validation accuracy for the proposed model

Figure 4. 3 Training and validation loss for the proposed model

From Figures 5.4 and 5.5, we can see that the training accuracy is above the validation accuracy
and training loss is way below the validation loss, especially after the 20th epoch. When there are
few training datasets, the model can learn from sounds or unwanted information in the training
dataset, which can harm the model's output on the new dataset. This phenomenon is known as
over-fitting. It means that the model will have a difficult time generalizing on the new dataset

There are many methods for combating over-fitting during the training period. Using data
augmentation and adding Dropout to the model are the two main methods.

Data augmentation is a technique to produce further training data from an existing dataset by
augmenting it with random transformations that result in believable-looking photos. This allows
the model to be exposed to more facets of the data and generalize more effectively. The updated

41
model was trained with 4220 images, and its testing dataset is 1055 images, based on the
principle of data augmentation.

Dropout, the kind of regularization, is another technique for reducing overfitting in the network.
When Dropout is applied to a layer, it randomly removes several output units from the layer
during the training phase (by setting the activation to zero). Dropout accepts fractional numbers
in the manner of 0.1, 0.2, 0.4, and so on as data. This means that 10%, 20%, or 40% of its output
nodes from the applied layer would be randomly removed. After applying the dense layer, the
updated model randomly drops out twice 20 percent of the output units.

A little modification on the model which is created in the previous section is used to train the
model with the new dataset, i.e., 4 convolutional layers, 4 max pooling, and 2 dense layers are
used to achieve the accuracy level, followed by 2 dropouts to prevent over-fitting. The validation
accuracy is 90.6 percent, and with a good balance of training and validation metrics, the accuracy
increases.

Figure 4. 4 Training and validation accuracy for the modified proposed model

42
 Figure 4. 5 Training and validation loss for the modified proposed model

The model improvement phase will go on until the training and validation precision, as well as
training and validation loss, is almost perfectly aligned.

4.2. Optimization

We discussed the model accuracy and how to improve it using augmenting the data and dropout
in the previous part. In this section, we will keep optimizing the model until it achieves the best
accuracy on training and validation, as well as relatively low loss model architecture. Layers and
nodes per layer, as well as 0, 1, or 2 dense layers, are the simplest things to change in the model.
As a result, we can create 27 different models based on layers, nodes per layer, and dense layers.

The same dataset is used to train and test all models. Table 4.1 shows the accuracy, loss,
validation accuracy, and validation loss of the produced models. Models contain some
randomness. There will be no two rounds of optimizations that are the same. They can resemble
each other but not be identical. Random weights are also used to start models. This can have a
major impact on models, particularly when there are fewer epochs or a limited training set.

43
 Table 4. 1 Model’s accuracy, loss, and validation accuracy and validation loss

Metrics
Validation Validation
Accuracy Loss
Models Accuracy Loss
1 1-Conv-32-nodes-0-dense 0.2057 nan 0.2054 nan
2 1-Conv-32-nodes-1-dense 0.2057 nan 0.2054 nan
3 1-Conv-32-nodes-2-dense 0.9246 0.2206 0.9098 0.3099
4 1-Conv-64-nodes-0-dense 0.2057 nan 0.2054 nan
5 1-Conv-64-nodes-1-dense 0.9552 0.1282 0.9175 0.3176
6 1-Conv-64-nodes-2-dense 0.9392 0.1587 0.9136 0.2594
7 1-Conv-128-nodes-0-dense 0.2057 nan 0.2054 nan
8 1-Conv-128-nodes-1-dense 0.9392 0.1718 0.9213 0.2492
9 1-Conv-128-nodes-2-dense 0.9229 0.2133 0.8983 0.2651
10 2-Conv-32-nodes-0-dense 0.2057 nan 0.2054 nan
11 2-Conv-32-nodes-1-dense 0.9409 0.1588 0.9175 0.2249
12 2-Conv-32-nodes-2-dense 0.9399 0.1774 0.8829 0.3168
13 2-Conv-64-nodes-0-dense 0.2057 nan 0.2054 nan
14 2-Conv-64-nodes-1-dense 0.9656 0.1158 0.9487 0.1719
15 2-Conv-64-nodes-2-dense 0.9337 0.1986 0.8925 0.314
16 2-Conv-128-nodes-0-dense 0.2057 nan 0.2054 nan
17 2-Conv-128-nodes-1-dense 0.9614 0.1122 0.9367 0.185
18 2-Conv-128-nodes-2-dense 0.942 0.1736 0.9098 0.2338
19 3-Conv-32-nodes-0-dense 0.2057 nan 0.2054 nan
20 3-Conv-32-nodes-1-dense 0.9243 0.1962 0.906 0.2487
21 3-Conv-32-nodes-2-dense 0.9205 0.2257 0.8983 0.2875
22 3-Conv-64-nodes-0-dense 0.2057 nan 0.2054 nan
23 3-Conv-64-nodes-1-dense 0.9473 0.1501 0.9175 0.231
24 3-Conv-64-nodes-2-dense 0.9291 0.1919 0.9155 0.3459
25 3-Conv-128-nodes-0-dense 0.2057 nan 0.2054 nan
26 3-Conv-128-nodes-1-dense 0.9464 0.1504 0.9328 0.2184
27 3-Conv-128-nodes-2-dense 0.9119 0.2605 0.9002 0.3076

It may be tempting to choose the model with the highest validation accuracy, but it is preferable
to choose the best (lowest) validation loss model. When it comes to models, there is some
randomness, as previously said, but you should note patterns. For example, I noticed that models
with 1 dense layer performed better overall, while models with 0 dense layers performed the
worst. Thus, depending on such a validation accuracy of greater than 91 percent, this was chosen
because the previous component had already achieved 90 percent validation accuracy. Now let

44
us start by looking at some of the better ones by zooming in on the validation accuracy graph.
The following are the top 9:

Table 4. 2 Shows the top 9 models in terms of validation accuracy and validation loss.

Models Validation Accuracy Validation Loss

2-Conv-64-nodes-1-dense 0.9487 0.1719
2-Conv-128-nodes-1-dense 0.9367 0.185
3-Conv-128-nodes-1-dense 0.9328 0.2184
1-Conv-128-nodes-1-dense 0.9213 0.2492
1-Conv-64-nodes-1-dense 0.9175 0.3176
2-Conv-32-nodes-1-dense 0.9175 0.2249
3-Conv-64-nodes-1-dense 0.9175 0.231
3-Conv-64-nodes-2-dense 0.9155 0.3459
1-Conv-64-nodes-2-dense 0.9136 0.2594

When only training accuracy and training loss are considered, the model's output is 96.14 percent
and 0.112 percent, respectively. However, when the validation accuracy and validation loss of
this model are considered, it ranks second. From this one, I assume the model has been
comfortable with 1 dense and 2 convolutional layers, as each version of those 2 options has
proven to be better than the others. The top three models in terms of validation accuracy and
validation loss are as follows:

Figure 4. 6 Validation accuracy for the top 3 models

45
 Figure 4. 7 Validation losses for the top 3 models

Finally, the model achieved an almost match with well-balanced training and validation metrics,
with a validation accuracy of 94.82 percent and the highest (lowest) validation loss of 0.1719. The
result was achieved using two convolutional layers, each with 64 nodes, 2 max-pooling layers,
and one dense layer followed by a single dropout.

Figure 4. 8 Training and validation accuracy for the best models

46
 Figure 4. 9 Training and validation loss for the best models

If there are an unequal number of observations in each class, or if the dataset has more than two
classes, classification accuracy alone can be misleading. The fundamental issue with
classification accuracy is that it masks the information needed to properly evaluate your
classification model's performance. There are two circumstances in which this problem is most
likely to occur:

 If there are even more than two classifications in a dataset, with three or more classes, the
classification accuracy may reach 80%, although it's unclear whether this is because all
classes are predicted equally well or because one or two classes are ignored by the model.

 When the number of classes in the dataset is not even, the model may get a 90% or higher
accuracy rate, but this isn't a good score if 90 of every 100 records belong to one class,
and the model does this by always predicting the most prevalent class value.
In general, classification accuracy can mask the information essential to diagnose your model's
performance. As a result, generating a confusion matrix can help you figure out what the
classification model is getting right and where it is going wrong. Below is a multiclass confusion
matrix showing TP, TN, FP, and FN.

47
  The total number of test images of any class would be the sum of the corresponding rows
(i.e. the TP+FN for that class)
 The total number of FN’s for a class is the sum of values in the corresponding Row
(excluding the TP)
 The total number of FP’s for a class is the sum of values in the corresponding Column
(excluding the TP)
 The total number of TN’s for a certain class will be the sum of all columns and rows
excluding that class column and row

Table 4. 3 A multiclass Confusion Matrix of the classes

Predicted
Class Chemical Clay Clean Sand Silt

Chemical 208 0 0 0 2
Actual Clay 0 206 0 5 1
Clean 0 3 203 3 1
Sand 1 1 5 201 4
Silt 2 7 0 4 198

There are four machine learning-specific statistical measures. These tests were Accuracy,
Precision, recall, and F1 score. The equations for calculating each of the statistical measures are
shown below.

= 2∗ TPR(Ci)∗ PPV(Ci)
F1-Score of Class Ci F1 (C1) = (5.1)
(TPR(Ci) + PPV(Ci)

TP(Ci)
Recall of Class Ci TPR (Ci) = (TP(Ci) + FN(Ci)) (5.2)

TP(Ci)
Precision of Class Ci PPV (Ci) = (TP(Ci) + FP(Ci)) (5.3)

∑ N i=1 TP(Ci)
Accuracy (Areduced) = ∑ N i=1 ∑ N j=1 Ci,j (5.4)

208+206+203+201+198
Accuracy = 1055
= 0.963

48
 Table 4. 4 Precision, Recall and F1 scores of the classes

Precision Recall F1-score

Chemical 0.991 0.986 0.988

Clay 0.972 0.949 0.960

Clean 0.967 0.976 0.971
Sand 0.948 0.944 0.946
Silt 0.929 0.961 0.945

The F1-scores of each class can be combined to create a single measure for the entire model.
There are a few methods

1. Micro F1:- It is calculated by taking into account the model's total FP, total TP, and total
FN. It does not take each class into account separately; instead, the metrics are calculated
as a whole.
Total TP = (208 + 206 + 203 + 201 + 198) = 1016
Total FP = (1+2) + (3+1+7) + (5) + (5+3+4) + (2+1+1+4) = 39
Total FN = (2) + (5+1) + (3+3+1) + (1+1+5+4) + (2+7+4) = 39

Hence,

Precision = 1016/1016+39 = 0.963

Recall = 1016/1016+39 = 0.963
We can now use the regular F1-score algorithm to calculate the Micro F1-score using the
precision and recall values listed above.

2∗ Precision ∗ Recall 2∗ 0.963 ∗ 0.963

Micro F1 = =
Precision+ Recall 0.963+ 0.963

Micro F1 = 0.963

As you can see, when we calculate the metrics globally, all of the variables are equal. That
means Precision, Recall, Micro F1, and Accuracy all are equal. The accuracy calculated above is
also the same.

Precision = Recall = Micro F1 = Accuracy

49
 2. Macro F1:- It calculates metrics for each class separately and takes an unweighted

average of the results. As we have seen from table 4.4 Precision, Recall, Micro F1 for

each class,

Class chemical F1-score = 0.988 Class clay F1-score = 0.960

Class clean F1-score = 0.971 Class sand F1-score = 0.946

Class silt F1-score = 0.945

Hence,

Macro F1 = (0.988+0.960+0.971+0.946+0.945)/5 = 0.963

3. Weighted F1:- It uses a weighted mean of the metrics, unlike Macro F1. The total

number of samples in each class determines the weights. Since we had 211 chemicals,

217 clay, 208 clean, 213 sand, and 206 silt samples.

(0.988∗211)+ (0960∗217) + (0.971∗208) + (0.946∗213) + (0.945∗206)

Weighted F1 = = 0.963
1055

Figure 4. 10 Graphical user interfaces

50
After evaluating the model on the testing image dataset, it had an accuracy of 96.3 percent
(1055/39 images) and a loss of 0.1719. The 39 pictures that were missing were categorized as
follows. While all 210 images of chemically polluted water were marked as contaminated, two of
them were incorrectly labeled as silt contamination. Six of the 212 clay-contaminated water
images were misidentified. Of those six images, one image is classified as silt and five as sand
pollution. Seven of the images that were missing were clean images that were misclassified as
contaminated. From these seven missing images, three images were classified as clay, three
images as sand, and one image was classified as silt pollution. Unlike the other contaminated
classes, sand contaminated images are misclassified as clean water. Off all 212 sand
contaminated water images, eleven images were classified wrongly, i.e. five images identified as
clean, one image identified as clay, one image identified as chemical, and 4 images classified as
silt contamination. Finally, all silt and vegetation testing images were categorized as polluted,
but thirteen images were incorrectly classified, including four images identified as sand, two
images identified as chemical, and seven images identified as clay, similar to other contaminated
water images.

4.3. Discussion

The issue that this project looked into was finding easy-to-use methods of classifying water as
safe or polluted in a water delivery system that was open to both the water company influences.
This project aims to introduce simple and fast additional monitoring mechanisms to an existing
system, assisting a water company and people who do not have a basic way to determine whether
or not the distributed drinking water is polluted. Contaminated water is a major health risk, and
many people who are affected by it are unaware of the full extent of the risk. Furthermore, water
quality analysis can be difficult to obtain, especially in developing countries where costly
chemical testing kits or other advanced methods of pollution testing are unavailable. This
project, on the other hand, proposes a solution based on machine learning that enables the water
company to monitor the water delivery system while also providing people in such
circumstances with a simple way to evaluate water in their communities, resulting in bilateral
control mechanisms. As discussed in the results section, the model had relatively high
confidence in its predictions. The high accuracy value of 96.3 percent reflects this belief. The
model predicted more false positives than false negatives, with seven clean photos listed as

51
contaminated, and five of the contaminated images were incorrectly classified as clean. One
possible explanation for this phenomenon is that the model predicts three of the images as clay,
three of the images as sand, and the last as silt, based on these seven frames. Because of this,
when we saw the first image, the container containing the water was so close to the clay color,
and because of this, its prediction was incorrect. Because of water transference, when we saw the
other two images, the background was reflected through the water, and the background color
looked like clay. As a result, the model was predicted inaccurately. The image, which the model
assumed was sand, was taken from a distance. When we modify the image to some extent, for
example, by zooming the image, and the model predicts that image, it accurately classifies the
image. During the augmentation of the image, it changes some characteristics of the original
image. In addition to that, when a photo is taken far from the object, the other things in the area
are also included. These factors affect the classification process. One of the two photos, which
were misclassified as contaminated with sand, was taken on the front side, and the other was
taken from the top side. When we saw both images, they were close to being contaminated with
sand at a low concentration. It accurately estimates the same photos taken in the other direction.
When we see the last image, the two glasses overlap each other. For this reason, the model's
prediction is incorrect. But when we take the main image at different angles, like from a top view
or side view, and when the model predicts with those images, it perfectly classifies the image.

When we look at the other four groups, we can see that they have all incorrectly labeled the
picture in various sizes. However, only one class i.e. sand class can classify the contaminated
water as pure water, even though they classify the picture incorrectly. Let’s look at each one in
detail.

1. Chemical class:-Two of the 210 test images were incorrectly identified as silt
contamination. When we look at these two photographs, we can see that one of them is a
combination of chemicals, silt, and vegetation, but the chemical concentration is higher
than the silt and vegetation, so it is classified as a chemical. The classification is
misleading, however, since these two groups are very identical. The other is a top-down
image that has been contaminated with a small chemical. When we take this image from
a side angle and create a model to classify it, it is accurately classified, not only for these
images but also for other images taken from different angles.

52
2. Clay class: - Six pictures out of 212 were incorrectly identified as sand and silt
contamination. When we look at all these images, we can see that they are polluted with a
very low concentration of clay, which the light reflected into the water is a yellow and
light blue color, respectively, and that they are near other classes that are also
contaminated with a small amount. As a result, it incorrectly classifies the picture.
However, the model correctly classifies the image when we adjust the color of the light
that is reflected on the water.
3. Sand class: - Out of 212 test images, eleven were incorrectly classified: five as clean, one
as clay, one as chemical, and four as silt. All five photos, classified as pure water, have
the same container containing the water, and they are polluted with a very low
concentration of sand. The images were taken from a distance. For all images, the light
reflected into the water is a light blue color, which is very close to the container of the
water. If we take an image with other lights, the model will classify perfectly. On the
other hand, when we look at the image as sand pollution, we can see that it is polluted
with a very low concentration of sand, that the light reflected into the water is blue, and
that it is close to other groups that are contaminated with a small amount. The other
images that were taken as chemical pollution are also contaminated with a small amount,
similar to the previous one. However, when we take this picture from a different point of
view and use a model to identify it, it is correctly classified. When we look at the other
four pictures, we can see that two of them are contaminated with a low concentration and
the other two are highly polluted and fall into the silt class. As a consequence, the model
makes wrong predictions.
4. Silt class ፡ As compared to other classes, this one incorrectly classifies the image used for
testing purposes the most. Thirteen out of 211 test images were incorrectly categorized,
with seven classified as clay, four as sand, and two as chemical. When we examine the
seven photos taken as clay contamination, we can see that four images out of seven are
contaminated with a very low concentration of silt and vegetation, with turbidity values
of 8.45NTU and the rest three images are highly contaminated but when we look the
images there is a similarity with clay. The other two images that are taken as chemical
pollution is a top view image, close to the chemical contamination as seen above.
However, when we take this picture from the side and use a model to identify it, it is

53
perfectly classified. The remaining four images are thought to be contaminated with sand.
Some images are heavily polluted, while others are lightly polluted, and the remaining
two images are taken from above.

54
 Chapter Five
Conclusion and Recommendation
5.1. Conclusion

The objective of this research is to develop a model for detecting pollution in the water delivery
system using an image of turbidity, a water quality parameter that is closely related to the
Physico-chemical properties of water quality parameters. The suggested model, in particular, is a
CNN that introduced an outlier detector in a drinking water delivery system and compared its
output to sensor-based water quality models. A total of 2000 images were obtained from both
clean and contaminated water to implement the model. In this study, polluted water was divided
into four groups: chemicals, clay, sand and silt, and vegetation. The images obtained are
insufficient to train a model to achieve high detection accuracy. Since the model accuracy
improves as the amount of training data grows. Using data augmentation, the number of images
obtained in this study was artificially increased to 5275. Of 5275 images, 4220 images are used
for training, and 1055 images are used to test the model. The developed model not only detects
whether it is clean or contaminated water, but it also identifies the source of contamination, i.e.
chemicals, clay, sand and silt, and vegetation. This helps for treatment purposes. As a result, the
established CNN model was effective in a variety of ways:

 The model achieved an accuracy of 96.2 percent and a loss of 0.1719, analyzing each
sample in only 30s, much more efficient than other models like sensor-based
contamination detector.
 The model detects contamination at a very low concentration of pollutants, i.e. it detects
water pollution at the turbidity level of 6.45NTU and above.
 The model identifies the type of contamination that is used for treatment.
 The model provided easily understandable results, meaning that it uses a graphical user
interface to enable users to interpret the results, and the prediction values allowed us to
evaluate the model's confidence in its decision, while also providing a variety of
contamination types.

55
In conclusion, this study aids in the development of an effective system for water quality
management that is both efficient and simple for those water supply organizations that would
assess the protection of their water supplies.

5.2. Recommendation
Expand the research in the future to provide analysis of microscopic water images to monitor for
the existence of harmful bacteria and other types of a pollutant not covered in this study. I will
keep refining the model and developing the mobile application, which is incorporated with the
system and can be used by individuals to monitor water quality from anywhere, eventually
leading to a mature system that allows for successful water quality inspection without the use of
complicated chemical testing procedures.

56
 References
AAWSA. (2011). A brief summary, accessible as hardcopy in AAWSA Library. Unpublished
document.
Abbas Q, Ibrahim MEA, Jaffar MA. (2019). A comprehensive review of recent advances on
deep vision systems. Artif Intell Rev 52:39–76. doi: 10.1007/s10462-018-9633-3
Ademe, A. S. (2014). Source and Determinants of Water Pollution in Ethiopia: Distributed Lag
Modeling Approach. Intellectual Property Rights: Open Access, 2(2).
https://doi.ortests.4172/2375-4516.1000110
APHA. Standard methods for the examination of water and wastewater, 20th ed., American
Public Health Association/American Water Works Association/Water Environment
Federation, Washington DC, USA. 1998-99
Ara S, MA Khan, MY Zargar, et al. Physico-chemical characteristics of Dal Lake water. Aqua
Environ Toxicol. 2003; 12:129-134.
Arad, J., Housh, M., Perelman, L., Ostfeld, A. (2013). A dynamic thresholds scheme for
contaminant event detection in water distribution systems. Water Res. 47 (5), 1899e1908.
http://dx.doi.org/10.1016/j.watres.2013.01.017.Ben-Hur, A., Weston, J., 2010. A user’s
guide to support vector machines. In: Data Mining Techniques for the Life Sciences,
Methods in Molecular Biology, vol. 609. Humana Press, pp. 223e239 a part of Springer
Science þ Business Media, LLC 2010, http://dx.doi.org/10.1007/978-1-60327-241-4_13
Aschale, M., Sileshi, Y., Kelly-Quinn, M. & Hailu, D. (2016). Evaluation of potentially toxic
element pollution in the benthic sediments of the water bodies of the city of Addis Ababa,
Ethiopia. Journal of Environmental Chemical Engineering, 4(4, Part A), pp 4173–4183.
Available: http://www.sciencedirect.com/science/article/pii/S2213343716303207.
Aschale, M., Sileshi, Y., Kelly-Quinn, M. & Hailu, D. (2017). Pollution Assessment of Toxic
and Potentially Toxic Elements in Agricultural Soils of the City Addis Ababa, Ethiopia.
Bulletin of Environmental Contamination and Toxicology, 98(2), pp 234–243. Available:
https://doi.org/10.1007/s00128-016-1975-4.
Bengio Y. (2009 Learning Deep Architectures for AI. Found Trends® Mach Learn 2:1–127.
doi:10.1561/2200000006.
Bouvrie J. (2006). Introduction Notes on Convolutional Neural Networks. doi:
http://dx.doi.org/10.1016/ j.protcy.2014.09.007
Byer, D., & Carlson, K. H. (2005). Real-time detection of intentional chemical contamination in
the distribution system. Journal-American Water Works Association, 97(7).
Che, Z., Purushotham, S., Cho, K., Sontag, D., & Liu, Y. (2018). Recurrent neural networks for
multivariate time series with missing values. Scientific Reports, 8(1), 6085.
Che, H., & Liu, S. (2014). Contaminant detection using multiple conventional water quality
sensors in an early warning system. Procedia Engineering, 89, 479–487.
https://doi.org/10.1016/j.proeng.2014.11.239

57
Choi, S., Cichocki, A., Park, H.-M., Lee, S.-Y. (2005). Blind source separation and independent
component analysis: a review. Neural Info Proc – Let and Rev. 6(1), 1-57.

Ciresan D, Giusti A, Gambardella LM, Schmidhuber J. (2012). Deep neural networks segment
neuronal membranes in electron microscopy images. In: Advances in neural information
processing systems. pp 2843–2851
Eliades, D. G., Lambrou, T. P., Panayiotou, C. G., & Polycarpou, M. M. (2014). Contamination
event detection in water distribution systems using a model-based approach. Procedia
Engineering, 89, 1089–1096. https://doi.org/10.1016/j.proeng.2014.11.229
Feintuch, P.L. (1990). Active Adaptive Noise Canceller without Training Mode. U.S. Patent, No.
5117401. U.S. Patent Office
Gakii, C., & Jepkoech, J. (2019). A Classification Model for Water Quality Analysis Using
Decision Tree. Journal of Chemical Information and Modeling, 7(3), 1–8.
Gangwar RK, Khare P Singh, J Singh, et al. (2012). Assessment of physicochemical properties
of water. River Ramganga at Bareilly, UP, Journal of Chemical and Pharmaceutical
Research, 2012; 4(9):4231-34. https://pubs.usgs.gov/wri/wri004139/pdf/wrir00-4139.pdf
Gregor, K., Danihelka, I., Graves, A., Rezende, D. J., & Wierstra, D. (2015). Draw: A recurrent
neural network for image generation. arXiv preprint arXiv:1502.04623
Guepie, B.K., Fillatre, L., Nikiforov, I. (2012). Sequential monitoring of water distribution
network. In: Paper Presented at the IFAC Proceedings Volumes (IFAC-papers Online), vol.
16, pp. 392e397. PART 1, http://dx.doi.org/10.3182/20120711-3-BE-2027.00114.
Guo Y, Liu Y, Oerlemans A, et al. (2016). Deep learning for visual understanding: A review.
Neurocomputing 187:27–48. doi: 10.1016/ j.neucom.2015.09.116
Gupta, A., & Ruebush, E. (2019). AquaSight: Automatic Water Impurity Detection Utilizing
Convolutional Neural Networks. http://arxiv.org/abs/1907.07573
Hall, J., Zaffiro, A.D., Marx, R.B., Kefauver, P.C., Krishman, E.R.,Haught, R.C., Herrmann, J.G.
(2007). On-line water quality parameters as indicators of distribution system contamination.
Journal of the American Water Works Association 99 (1), 66e77.
Haught, R., Hall, J., Rahman, M., Richardson-Coy, R., Piao, H. (2005). The Bench-Scale
Minimum Threshold Experiment. WQTC, Toronto, Canada.
Hou, D., He, H., Huang, P., Zhang, G., & Loaiciga, H. (2013). Detection of water-quality
contamination events based on multi-sensor fusion using an extented Dempster-Shafer
method. Measurement Science and Technology, 24(5). https://doi.org/10.1088/0957-
0233/24/5/055801
Housh, M., & Ostfeld, A. (2015). An integrated logit model for contamination event detection in
water distribution systems. Water Research, 75, 210–223.
https://doi.org/10.1016/j.watres.2015.02.016
Huang, P., Jin, Y., Hou, D., Yu, J., Tu, D., Cao, Y., & Zhang, G. (2017). Online classification of
contaminants based on multi-classification support vector machine using conventional
58
 water quality sensors. Sensors (Switzerland), 17(3). https://doi.org/10.3390/s17030581
Kang, G. K., Gao, J. Z., & Xie, G. (n.d). Data-driven water quality analysis and prediction: A
survey.
Kavi Priya, S., Shenbagalakshmi, G., & Revathi, T. (2017). Design of smart sensors for real time
drinking water quality monitoring and contamination detection in water distributed mains.
International Journal of Engineering & Technology, 7(1.1), 47.
https://doi.org/10.14419/ijet.v7i1.1.8921
Khan MA. Recent bio-limnological pollution trends in Kashmir Himalayan Dal Lake ecosystem.
Development of red bloom. Ecol Energy. 1996; 22:41-77.
Khan, A., Sohail, A., Zahoora, U., & Qureshi, A. S. (2020). A survey of the recent architectures
of deep convolutional neural networks. Artificial Intelligence Review, 53(8), 5455–5516.
https://doi.org/10.1007/s10462-020-09825-6
Klise, K.A., McKenna, S.A. (2006). Water quality change detection: multivariate algorithms. In:
Saito, T.T., Lehrfeld, D. (Eds.), Proc. SPIE, 6203, pp. J1–J9.
Klise, K.A., McKenna, S.A. (2006). Multivariate application for detecting anomalous water
quality. In: Proceedings of the 8th Annual Water Distribution Systems Analysis
Symposium. WDSA, Cincinnati, Ohio, USA, pp. 1e11. https://doi.org/10.1061/40941 (247)
130
Kroll, D., King, K. (2006). Real world operational testing and development of an on-line water
security monitoring system. In: Eighth Ann WDSA Symp. Cincinnati, USA.
Kroll, D. (2006). Securing our water supply: protecting a vulnerable resource. Tulsa: Pennwell.
Kühnert, C., Bernard, T., Montalvo Arango, I., & Nitsche, R. (2014). Water quality supervision
of distribution networks based on machine learning algorithms and operator feedback.
Procedia Engineering, 89, 189–196. https://doi.org/10.1016/j.proeng.2014.11.176
L. Perelman, J. Arad, M. Housh, A. Ostfeld, Event detection in water distribution systems from
multivariate water quality time series, Environmental Science & Technology 46 (2012)
8212–8219
LeCun Y, Kavukcuoglu K, Farabet CC, others. (2010). Convolutional networks and applications
in vision. In: ISCAS. IEEE, pp 253–256
Lin, Q.-H., Zheng, Y.-R., Yin, F.-L., Liang, H., Calhoun, V.D. (2007). A fast algorithm for one-
unit ICA-R. Inf. Sci. 177, 1265–1275.
Liu, S., Che, H., Smith, K., & Chang, T. (2015). A real time method of contaminant
classification using conventional water quality sensors. Journal of Environmental
Management, 154, 13–21. https://doi.org/10.1016/j.jenvman.2015.02.023
Liu W, Wang Z, Liu X, et al. (2016). A survey of deep neural network architectures and their
applications. Neurocomputing 234:11–26. doi: 10.1016/j.neucom.2016.12.038
Liu X, Deng Z, Yang Y. (2019). Recent progress in semantic image segmentation. Artif Intell

59
 Rev 52:1089–1106. doi: 10.1007/s10462-018-9641-3
McKenna, S.A., Wilson, M., Klise, K.A. (2008). Detecting changes in water quality data. Journal
of the American Water Works Association 100 (1), 74e85.
Mohammed, H., Hameed, I. A., & Seidu, R. (n.d.). Machine Learning – Based Detection of
Water Contamination in Water Distribution Systems. 1664–1671.

Mohammadpour, R., Shaharuddin, S., Chang, C. K., Zakaria, N. A., Ab Ghani, A., & Chan, N.
W. (2015). Prediction of water quality index in constructed wetlands using support vector
machine. Environmental Science and Pollution Research, 22(8), 6208–6219.
Muharemi, F., Logofătu, D., & Leon, F. (2019). Machine learning approaches for anomaly
detection of water quality on a real-world data set. Journal of Information and
Telecommunication, 3(3), 294–307. https://doi.org/10.1080/24751839.2019.1565653
Murray, R., Haxton, T., McKenna, S.A., Hart, D.B., Klise, K.A.,Koch, M., Vugrin, E.D., Martin,
S., Wilson, M., Cruze, V.A.,Cutler, L. (2010). Water quality event detection system for
drinking water contamination waring system: Development Testing and Application Of
CANARY. EPA/600/R-10/036. U.S. Environmental Protections Agency, Office of
Research and Development, National Homeland Security Research Center, Cincinnati,
Ohio,USA
http://cfpub.epa.gov/si/si_public_record_report.cfm?address¼nhsrc/&dirEntryId¼221394.
Najafabadi MM, Villanustre F, Khoshgoftaar TM, et al. (2015). Deep learning applications and
challenges in big data analytics. J Big Data 2:1–21. doi: 10.1186/s40537-014-0007-7.
National Research Council. (1995). Expanding the Vision of Sensor Materials. Academies Press,
Washington, D.C., 134 p.
Oliker, N., & Ostfeld, A. (2014a). A coupled classification - Evolutionary optimization model
for contamination event detection in water distribution systems. Water Research, 51, 234–
245. https://doi.org/10.1016/j.watres.2013.10.060
Oliker, N., & Ostfeld, A. (2014b). Minimum volume ellipsoid classification model for
contamination event detection in water distribution systems. Environmental Modelling and
Software, 57, 1–12. https://doi.org/10.1016/j.envsoft.2014.03.011
Olson E. (2004). Grading Drinking water in U.S cities what’s on Tap? National resource defense
Counsel, New York City, and Washington, D.C., Los Angeles, and San Francisco.
P. Corso, M. Kramer, K. Blair, D. Addiss, J. Davis, A. Haddix, Cost of illness in the 1993
waterborne cryptosporidium outbreak, Milwaukee, Wisconsin, Emerging Infectious
Diseases 9 (2003) 426–431.
Perelman, L., Arad, J., Housh, M., Ostfeld, A. (2012). Event detection in water distribution
systems from multivariate water quality time series. . Environ. Sci. Technol. 46 (15),
8212e8219. http:/dx.doi.org/10.1021/es3014024.
Scherer D, Müller A, Behnke S. (2010). Evaluation of pooling operations in convolutional
architectures for object recognition. In: Artificial Neural Networks--ICANN 2010. Springer,

60
 pp 92–101
Schwartz, J., Levin, R. (1999). Drinking water turbidity and health. Epidemiology, 86-90.
Sime, I. (1998). Addis Ababa water supply stage III-A project, Ethiopian Association of Civil
Engineers Bulletin, vol. 1, no. 1, Ethiopian Association of Civil Engineers, Addis Ababa.
Skadsen, J., Janke, R., Grayman, W., Samuels, W., Tenbroek, M., Steglitz, B., & Bahl, S. (2008).
Distribution system on-line monitoring for detecting contamination and water quality
changes. Journal / American Water Works Association, 100(7), 81–94.
https://doi.org/10.1002/j.1551-8833.2008.tb09678.x
Szabo, J.G., Hall, J.S., Meiners, G. (2007). Water quality sensor responses to contamination in a
single pass water distribution system simulator. EPA/600/R-07/001, p. 35.
The Water Project (Water in Crisis - Spotlight Ethiopia). (n.d.). Retrieved, 2015, from
http://thewaterproject.org/water-in-crisis-ethiopia
Uber, J.G., Murray, R., Magnuson, M., Umberg, K. (2007). Evaluating real-time event detection
algorithms using synthetic data. In: Paper Presented at the Restoring Our Natural Habitat-
Proceedings of the 2007 World Environmental and Water Resources Congress.
http://ascelibrary.org/doi/abs/10.1061/40927%28243%29499.
UN (United Nations). (2010). United Nations Millennium Development Goals. Accessed at
https:// www.un.org/millenniumgoals/environ.shtml .
UN. (2017). World Population Prospects - Population Division - United Nations. Available
https://population.un.org/wpp/Download/Standard/Population/
UN. (2018). World Urbanization Prospects - Population Division - United Nations. Available
https://population.un.org/wup/
UNDP (United Nations Development Program). (2006). Human development report 2006. New
York: UNDP.
U.S. Environmental Protection Agency. (2005). Technologies and Techniques for Early Warning
Systems to Monitor and Evaluate Drinking Water Quality: State-of the- Art Review. ORD,
NHSRC, Washington, DC.
U.S. EPA. (2005a). Water Sentinel: Online Water Quality Monitoring as an Indicator of
Drinking Water Contamination. Available at:
http://www.epa.gov/watersecurity/pubs/watersentinel_wq_monitoring.pdf
U.S. EPA. (2005b). Water Sentinel: System Architecture. Available at:
http://www.epa.gov/watersecurity/pubs/watersentinel_system_architecture.pdf
W.G. Wright Jr, and A.J. Thomas III. (2000). the Federal/Arkansas Water Pollution Control
Programs: Past, Present, and Future. UALR L. Rev., 23: 541 (2000).
W. J. Cooper, (2014). Responding to crisis: The West Virginia chemical spill, Environmental
Science & Technology 48 (2014) 3095–3095.
Worldwildlife. (2021) Water Scarcity | Threats | WWF. (n.d.). Retrieved May 29, 2021, from

61
 https://www.worldwildlife.org/threats/water-scarcity

Jeremiah Castelo. (2021) What is the Percentage of Drinkable Water on Earth? | World Water
Reserve. (n.d.). Retrieved May 29, 2021, from https://worldwaterreserve.com/water-
crisis/percentage-of-drinkable-water-on-earth/

WHO/UNESCO. (2010). Progress on Sanitation and Drinking water: 2010 Update. Geneva,
WHO press JMP, Update, 2010.
WHO, W. H. O. (2006). Guidelines for the safe use of wastewater, excreta and greywater -
Volume 1. (WHO), http://www.who.int/water_sanitation_health/publications/gsuweg1/en/
World Health Organization. (2011). Nitrate and Nitrite in Drinking-Water.
https://www.who.int/water_sanitation_health/dwq/chemicals/nitratenitrite2ndadd.pdf.
Xiang, Y., & Jiang, L. (2009). Water quality prediction using ls-svm and particle swarm
optimization. Knowledge Discovery and Data Mining, 2009. WKDD 2009. Second
International Workshop on (pp. 900–904). IEEE.
Yang, Y.J., Haught, R.C., Hall, J., Szabo, J., Clark, R.M., Meiners, G. (2007). Adaptive water
sensor signal processing: experimental results and implications for online contaminant
warning systems. In: Wat Env and Wat Res Congress, Tampa, Florida
Yang, J.Y., Haught, C.R., Goodrich, A.J. (2009). Real-time contaminant detection and
classification in a drinking water pipe using conventional water quality sensors: techniques
and experimental results. Journal of Environmental Management 90 (8), 2494e2506.
Yohannes, H. & Elias, E. (2017). Contamination of Rivers and Water Reservoirs in and Around
Addis Ababa City and Actions to Combat It. Environment Pollution and Climate Change,
01.
Zhang, J., Zhu, X., Yue, Y., & Wong, P. W. (2017). A real-time anomaly detection algorithm/or
water quality data using dual time-moving windows. Seventh international conference on
innovative computing technology (INTECH) (pp. 36–41). IEEE.

62
 Appendix
Appendix: A Sample code

Code 1

Code 2

Code 3

Code 4

63
Appendix: B Data Analysis Results of clean water

Table 1 Results from conducted analyses of collected clean water samples.

Total Hardness

N (Ammonia)

SO4 (Sulfate)

Cl (Chloride)
SiO2 (Silica)
(Manganese)
F- (Fluoride)

Bicarbonate
(Phosphate)
Parameters

N (Nitrate)
N (Nitrite)
Alkalinity

Alkalinity
Turbidity

Fe (Iron)
Total
TDS

PO4

Mn
PH
EC
Units
Date NTU mg/l µS mg/l mg/l mg/l mg/l mg/l mg/l mg/l mg/l mg/l mg/l mg/l mg/l mg/l
Jan 8, 2020 0.56 7.75 74 134.1 44 52 0.162 0.0055 0.48 0.1 0.086 0.27 0.122 0.018 11.6 6.5 44
Jan 12, 2020 1.2 7.61 72 127 47.4 49.6 0.133 0.0022 0.32 1.8 0.053 0.13 0.064 0.058 7.4 5.8 47.4
Jan 15, 2020 0.96 7.73 70 132.4 45 49.9 0.855 ND 0.2 3.4 ND 0.051 0.077 0.017 ND 5 45
Jan 19, 2020 0.98 7.58 76 137.6 51 48 0.025 0.0014 0.15 ND 0.294 ND 2.543 0.42 1.4 7 51
Jan 21, 2020 1.1 7.71 78 141.4 50.8 92.8 0.755 0.0043 0.42 1.5 1.363 ND 2.077 0.012 ND 3.6 50.8
Jan 24, 2020 0.78 7.15 74 135.5 63.2 58.6 0.336 ND 0.65 2.3 0.889 0.086 2.088 0.028 2.7 5.1 63.2
Jan 26, 2020 1.23 7.54 72 131.4 48 45.2 ND ND 0.8 ND 0.129 ND 1.993 0.017 5.6 5.4 48
Jan 28, 2020 0.75 6.76 71 129.4 52.2 54.4 0.325 0.0062 0.45 3.3 0.136 0.14 0.137 0.026 15.6 6.3 52.2
Jan 30, 2020 0.88 7.17 68 134.9 50.8 92.8 0.544 0.0043 0.42 1.5 1.363 ND 2.077 0.012 8.3 3.6 50.8
Jan 31, 2020 0.89 6.98 65 128.6 50.4 58.2 0.753 0.0026 1.1 0.9 0.103 0.18 1.662 0.349 12.3 4.1 50.4
Feb 2, 2020 0.91 7.05 64 110.6 45.2 56.4 0.887 0.0043 1.2 0.7 0.133 0.14 1.582 0.488 14.7 5.1 45.2
Feb 4, 2020 0.95 7.37 63 117.7 48.8 62.6 0.447 0.0037 0.89 0.9 0.095 0.19 0.464 0.228 9.3 6.4 48.8
Feb 7, 2020 0.72 7.57 66 118.3 44.6 54.8 0.262 0.063 0.75 0.6 0.077 0.14 0.263 0.118 10.6 6.2 44.6
Feb 10, 2020 0.45 7.65 73 138.4 45.4 51.2 0.146 0.0123 0.52 1.2 0.064 0.18 0.164 0.085 8.6 5.9 45.4
Feb 12, 2020 1.1 7.73 71 135.4 46 48 0.035 0.0146 0.28 2.5 0.188 0.074 1.273 0.116 2.3 6.1 46
Feb 15, 2020 0.99 7.76 78 144.5 62.2 108 0.034 0.0223 0.13 0.8 0.394 0.033 2.245 0.34 0.9 6.7 62.2
Feb 18, 2020 0.89 7.68 64 118.6 54.8 88.6 0.63 0.0034 0.24 1.8 1.133 0.041 2.144 0.025 1.9 3.5 54.8
Feb 21, 2020 0.85 7.21 65 138.5 61.2 61.6 0.146 0.0017 0.57 1.9 0.995 0.063 2.011 0.032 3.7 4.2 61.2
Feb 24, 2020 0.79 7.54 62 131.4 49 45.2 ND ND 0.8 ND 0.129 ND 1.993 0.017 5.6 5.4 49

64
Mar 1, 2020 0.77 7.51 75 142 47.2 76.4 0.074 ND 3.7 0.2 0.159 ND 1.484 0.434 13 2.3 47.2
Mar 4, 2020 0.88 7.26 71 149.5 63 65 0.876 0.009 0.11 0.3 0.071 0.32 1.771 0.644 13.2 6.3 63
Mar 10, 2020 0.91 7.08 73 136.4 53.2 60.2 0.785 0.001 1.4 0.7 0.098 0.29 1.887 0.457 14.1 5.3 53.2
Mar 13, 2020 1.1 6.95 71 135 43.4 54.8 0.336 0.006 1.9 0.5 0.121 0.01 1.982 0.557 15.3 4.5 43.4
Mar 16, 2020 1.2 7.57 70 133.2 48.4 64.8 0.385 0.0035 0.36 1.8 0.198 0.15 0.228 0.023 8.3 5.7 48.4
Mar 23, 2020 0.99 7.75 74 134.4 42 54 0.162 0.0055 0.48 0.1 0.086 0.27 0.122 0.018 11.6 6.5 42
Mar 26, 2020 0.95 7.61 72 137 47.4 50.6 0.133 0.0022 0.32 1.8 0.053 0.13 0.064 0.058 7.4 5.4 47.4
Mar 29, 2020 0.93 7.73 70 134.3 47 49.9 0.23 ND 0.2 3.4 ND 0.051 0.077 0.017 ND 4.8 47
Apr 2, 2020 0.92 7.58 76 137.6 51 48 0.025 0.0114 0.15 ND 0.294 ND 2.543 0.42 1.4 7 51
Apr 5, 2020 0.56 7.71 78 131.4 52.8 92.8 0.662 0.0043 0.42 1.5 1.363 ND 2.077 0.012 ND 4.2 52.8
Apr 11, 2020 0.75 7.15 74 135.5 64 58.6 0.26 ND 0.65 2.3 0.889 0.086 2.088 0.028 2.7 4.1 64
Apr 14, 2020 0.85 7.54 72 131.4 53 45.2 ND ND 0.8 ND 0.129 ND 1.993 0.017 5.6 5.3 53
Apr 17, 2020 0.86 6.76 71 129.4 57 54.4 0.325 0.0145 0.45 3.3 0.136 0.14 0.137 0.026 15.6 7.1 57
Apr 23, 2020 0.89 7.17 68 134.9 49 92.8 0.84 0.0054 0.42 1.5 1.363 ND 2.077 0.012 8.3 3.6 49
Apr 26, 2020 0.91 6.98 71 138.6 52 58.2 0.532 0.0233 1.1 0.9 0.103 0.18 1.662 0.349 12.3 4.1 52
Apr 29, 2020 0.99 7.05 74 130.6 45 56.4 0.322 0.0078 1.2 0.7 0.133 0.14 1.582 0.488 14.7 5.3 45
May 3, 2020 1.1 7.37 73 137.7 47 62.6 0.447 0.0056 0.89 0.9 0.095 0.19 0.464 0.228 9.3 6.4 47
May 6, 2020 1.02 7.57 76 138.3 45.4 54.8 0.346 0.0025 0.75 0.6 0.077 0.14 0.263 0.118 10.6 5.2 45.4
May 12, 2020 0.95 7.65 73 135.4 46.2 51.2 0.453 0.0089 0.52 1.2 0.064 0.18 0.164 0.085 8.6 6.3 46.2
May 15, 2020 0.88 7.73 71 133.8 48 48 0.335 0.0045 0.28 2.5 0.188 0.074 1.273 0.116 2.3 5.8 48
May 21, 2020 0.82 7.76 68 124.5 59 108 0.078 0.0156 0.13 0.8 0.394 0.033 2.245 0.34 0.9 5.6 59
May 24, 2020 0.78 7.68 74 138.5 56.8 88.6 0.583 0.0085 0.24 1.8 1.133 0.041 2.144 0.025 1.9 4.3 56.8
May 27, 2020 0.92 7.21 65 127.3 60.4 61.6 0.143 0.0074 0.57 1.9 0.995 0.063 2.011 0.032 3.7 4.4 60.4
May 30, 2020 0.65 7.54 72 131.4 49 45.2 0.089 0.0045 0.8 ND 0.129 ND 1.993 0.017 5.6 5.2 49
Jun 3, 2020 0.67 7.51 78 141.2 47.2 76.4 0.074 0.0053 3.7 0.2 0.159 ND 1.484 0.434 13 3.1 47.2
Jun 8, 2020 0.89 7.26 71 143.5 63 65 0.765 ND 0.11 0.3 0.071 0.32 1.771 0.644 13.2 5.9 63
Jun 9, 2020 0.92 7.08 73 136.4 53.2 60.2 0.326 0.001 1.4 0.7 0.098 0.29 1.887 0.457 14.1 4.5 53.2
Jun 11, 2020 0.78 6.95 71 135.6 43.4 54.8 0.852 0.006 1.9 0.5 0.121 0.01 1.982 0.557 15.3 4.6 43.4
Jun 13, 2020 0.99 7.57 70 133.2 48.4 64.8 0.432 ND 0.36 1.8 0.198 0.15 0.228 0.023 8.3 6.8 48.4
Jun 16, 2020 1.1 7.75 74 134.1 42 52 0.162 0.0055 0.48 0.1 0.086 0.27 0.122 0.018 11.6 6.5 42

65
Appendix: C Data Analysis Results of contaminated water
Table 2 Results from conducted analyses of collected contaminated water samples.

Total Ca N N Mn SiO2 Cl
Parameters Turbidity PH TDS Alkalinity Hardness (Nitrite) (Nitrate) Fe (Iron) (Manganese) (Silica) (Chloride)
Units
Date NTU mg/l mg/l mg/l mg/l mg/l mg/l mg/l mg/l mg/l
Jan 8, 2020 33.7 7.75 524 63 40.6 0.006 1.1 1.982 0.644 11.6 6.7
Jan 12, 2020 62.2 7.61 574 53.2 35 ND 1.2 0.228 0.488 7.4 3.5
Jan 15, 2020 73.12 7.73 578 43.4 42.4 0.0055 0.89 0.122 0.457 ND 4.2
Jan 19, 2020 77.1 6.02 547 48.4 33.2 0.0022 57.12 0.064 0.434 1.4 5.4
Jan 21, 2020 57.41 7.71 562 42 35.8 ND 0.52 0.077 0.349 ND 2.3
Jan 24, 2020 64.52 7.15 521 47.4 32.2 0.0231 0.28 2.543 0.34 2.7 6.3
Jan 26, 2020 62.62 4.23 423 47 67.5 0.0056 0.13 22.1 3.2 100 300
Jan 28, 2020 59 6.76 200 51 50.2 ND 0.24 24.2 0.118 15.6 4.5
Jan 30, 2020 41.4 7.17 159 52.8 42.4 ND 0.57 1.993 0.116 8.3 5.7
Jan 31, 2020 71.3 6.98 320 64 39.2 0.0062 0.8 0.137 0.085 12.3 6.5
Feb 2, 2020 65.6 7.05 331 53 47 0.0045 3.7 2.077 0.032 14.7 5.4
Feb 4, 2020 87.9 7.37 323 48.8 44 0.0156 0.11 1.662 0.026 9.3 4.8
Feb 7, 2020 88.7 5.91 351 44.6 42.2 0.0085 62.3 16.5 0.025 10.6 7
Feb 10, 2020 83.3 7.65 372 45.4 34.2 0.0074 1.9 0.464 0.017 8.6 4.2
Feb 12, 2020 64.4 7.73 311 46 42.6 0.0045 0.36 0.122 0.012 2.3 4.1
Feb 15, 2020 77.6 7.76 301 62.2 33.2 0.0053 0.13 1.993 0.644 14.1 5.3
Feb 18, 2020 72.2 7.68 306 54.8 35.8 ND 0.24 17.64 0.488 15.3 7.1
Feb 21, 2020 66.8 7.21 323 61.2 32.2 0.001 0.57 1.771 0.434 8.3 3.6
Feb 24, 2020 87.8 7.54 410 49 60.4 0.006 0.8 1.887 9.1 11.6 4.1
Mar 1, 2020 99.12 7.51 387 47.2 50.2 ND 3.7 1.982 0.34 53.7 5.3
Mar 4, 2020 102.3 7.26 315 44 71.3 0.0034 0.11 13.56 0.228 ND 6.4

66
Mar 10, 2020 120.12 7.08 425 47.4 39.2 0.0017 1.4 0.122 0.118 1.4 5.2
Mar 13, 2020 92.1 6.95 362 45 47 ND 1.9 0.064 0.116 ND 6.3
Mar 16, 2020 78.4 7.57 303 42.2 44 ND 0.36 0.077 0.085 2.7 5.8
Mar 23, 2020 95.6 7.75 299 51 42.2 0.009 0.48 2.543 0.032 5.6 5.6
Mar 26, 2020 96.4 7.61 342 52.8 34.2 0.001 0.32 2.077 0.026 15.6 4.3
Mar 29, 2020 89.7 7.73 344 64 42.6 0.006 1.1 19.9 17.2 8.3 287
Apr 2, 2020 63.8 5.99 356 53 24 0.0035 69.2 1.993 0.017 12.3 5.2
Apr 5, 2020 77.1 7.71 361 57 37.4 0.0055 0.89 0.137 0.017 14.7 3.1
Apr 11, 2020 56.5 7.15 320 49 34.2 0.0022 0.48 2.077 0.012 9.3 5.9
Apr 14, 2020 65.3 7.54 350 52 38.4 0.0156 0.32 1.993 0.018 0.9 4.5
Apr 17, 2020 52.5 6.76 301 45 42.2 ND 0.2 0.137 0.488 1.9 4.6
Apr 23, 2020 56.1 5.3 298 47 40.6 0.0563 0.15 13.2 0.42 76.4 6.8
Apr 26, 2020 57.9 6.98 305 45.4 35 0.0085 0.42 1.662 0.349 5.6 4.8
Apr 29, 2020 58.7 7.05 302 46.2 66 0.0055 0.65 1.582 0.34 13 6.4
May 3, 2020 66.75 7.37 312 47 22 0.0022 0.8 0.464 0.228 13.2 6.5
May 6, 2020 56.6 6.05 300 45.4 36.4 ND 65.4 0.263 10.2 14.1 5.8
May 12, 2020 45.4 7.65 320 46.2 34.2 0.0014 0.42 0.164 0.116 15.3 5
May 15, 2020 41.2 7.73 273 48 34 0.0043 1.1 14.4 0.085 8.3 7
May 21, 2020 42.3 7.76 363 59 42.2 ND 1.2 2.245 0.058 11.6 3.6
May 24, 2020 36.1 7.68 356 56.8 40.6 0.0085 0.89 2.144 0.028 7.4 5.1
May 27, 2020 39.9 7.21 307 48 39.2 0.0074 0.57 2.011 0.026 ND 5.4
May 30, 2020 46.21 7.54 193 59 38.4 0.0045 0.8 1.993 0.017 1.4 6.3
Jun 3, 2020 78.7 7.51 255 56.8 42.2 0.0053 3.7 1.484 0.017 ND 3.6
Jun 8, 2020 98.9 7.26 314 60.4 87.5 ND 0.11 1.771 0.012 2.7 305.1
Jun 9, 2020 88.4 7.08 361 49 42.2 0.001 1.4 1.887 0.012 5.6 5.1
Jun 11, 2020 83.5 6.95 323 47.2 34.2 0.0055 1.9 0.122 1.5 15.6 6.4
Jun 13, 2020 95.1 5.6 311 63 42.6 0.0022 74.2 0.064 0.025 8.3 6.2
Jun 16, 2020 235.2 7.75 547 53.2 22 ND 0.48 0.077 0.032 65 5.9
Jun 19, 2020 195.3 7.61 465 43.4 36.4 0.0014 0.32 2.543 0.017 14.7 6.1
Jun 22, 2020 311.1 7.73 700 48.4 34.2 0.0043 0.2 2.077 0.434 9.3 5

67
68