INDEX OF THE TEXT

DECLARATION OF ORIGINALITY .................................................................iii

ACKNOWLEDGEMENTS ................................................................................ iv

ABSTRACT ...................................................................................................... v

KEYWORDS ................................................................................................... vi

ACRONYMS................................................................................................... vii

INDEX OF THE TEXT ..................................................................................... ix

INDEX OF TABLES ........................................................................................ xii

INDEX OF FIGURES..................................................................................... xiii

1 INTRODUCTION ....................................................................................... 1

1.1 Background and Motivation ..................................................................................... 1

1.2 Research Gap Identification ..................................................................................... 3

1.3 Research Aim ............................................................................................................. 5

1.3.1 Aim ..................................................................................................................... 5

1.4 Methodology Overview ............................................................................................. 5

1.5 Thesis Structure......................................................................................................... 6

2 LITERATURE REVIEW ............................................................................. 7

2.1 Flood Influencing Factors ......................................................................................... 7

2.1.1 Natural-caused Factors...................................................................................... 7

2.1.2 Anthropogenic Factors (Human-Induced) ....................................................... 9

2.2 Flood Susceptibility Modeling Approaches ............................................................ 10

2.2.1 Hydrological Approach (Physical-based) ....................................................... 10

2.2.2 Qualitative Approach ....................................................................................... 11

2.2.3 Statistical Approach ......................................................................................... 12

2.3 Machine Learning Modelling Approach ................................................................. 13

2.3.1 Decision Tree (DT)........................................................................................... 14

2.3.2 Random forest (RF) ......................................................................................... 14

ix
 2.3.3 Support Vector Machine (SVM) ...................................................................... 15

3. DATA AND CASE STUDY ......................................................................... 16

3.1 Case Study ............................................................................................................... 16

3.2 Datasets and Preprocessing .................................................................................... 17

3.2.1 Derivation of Flood Inventory Dataset ............................................................... 17

3.2.2 Derivation of Flood Influencing Factors Datasets.............................................. 19

4. METHODOLOGY ....................................................................................... 24

4.1 Feature Selection and Multi-collinearity Analysis ................................................. 24

4.1.1 Linear Support Vector Machine (LSVM) ........................................................ 25

4.1.2 Multi-collinearity Analysis .............................................................................. 25

4.2 Data Cleaning and Normalization .......................................................................... 25

4.3 Index of Entropy Modelling .................................................................................... 26

4.4 Machine Learning (ML) Algorithms ....................................................................... 27

4.4.1 Support Vector Machine .................................................................................. 27

4.4.2 Random Forest Model ..................................................................................... 28

4.4.3 Decision Tree Model ........................................................................................ 29

4.5 Hybrid Modelling .................................................................................................... 29

4.6 Creation of Flood Susceptibility Maps.................................................................... 29

4.7 Results Validation and Model’s Performance Assessment .................................... 30

4.7.1 Model Evaluation using the ROC Curve ......................................................... 30

4.7.2 Statistical Metrics ............................................................................................ 31

4.8 Software and Device Specifications ........................................................................ 32

5 RESULTS ................................................................................................ 33

5.1 Predictiveness of Flood Influencing Factors (Feature Engineering)..................... 33

5.1.1 Feature Selection ............................................................................................. 33

5.1.2 Multi-collinearity Analysis .............................................................................. 34

5.2 Flood Modelling Algorithms ................................................................................... 34

x
 5.2.1 Index of Entropy Flood Modelling .................................................................. 34

5.2.2 Support Vector Machine Flood Modelling ...................................................... 38

5.2.3 Random Forest Flood Modelling .................................................................... 39

5.2.4 Decision Tree Flood Modelling ....................................................................... 41

5.3 Accuracy Assessment and Validation of Flood Models.......................................... 41

6 DISCUSSION ........................................................................................... 44

6.1 Limitations and Recommendations ....................................................................... 47

7 CONCLUSION ......................................................................................... 49

8 BIBLIOGRAPHIC REFERENCES ............................................................. 51

9 ANNEXES ............................................................................................... 58

9.1 Relative Distribution of flood pixels within Flood Influencing Factors’ Classes .. 58

9.2 Descriptive statistics of the training and testing datasets ..................................... 60

xi
INDEX OF TABLES
Table 1: Spatial Datasets and Data sources .......................................................................... 22
Table 2 : Multi-collinearity Analysis...................................................................................... 34
Table 3: Frequency ratio and Index of Entropy coefficients values distribution within flood
influencing factors classes...................................................................................................... 37
Table 4: Performance metrics of Classifiers .......................................................................... 43

xii
INDEX OF FIGURES
Figure 1: Flooding in Lagos ..................................................................................................... 3
Figure 2: Case Study ............................................................................................................... 17
Figure 3: Flood Inventory Map .............................................................................................. 19
Figure 4: Flood influencing factors........................................................................................ 21
Figure 5: Methodology Flowchart ......................................................................................... 24
Figure 6: Factor’s predictive ability (average merit) result .................................................. 33
Figure 7: Flood probability index derived from: (a) RF-IOE, (b) DT-IOE, (c) Stand-alone DT,
(d) Stand-alone RF, (e) SVM-IOE, (f) Stand-alone IOE, (g) Stand-alone SVM. ................. 39
Figure 8: Flood susceptibility maps derived from (a) RF-IOE, (b) Stand-alone DT, (c) Stand-
alone RF, (d) DT-IOE, (e) Stand-alone SVM, (f) SVM-IOE, (g) Stand-alone (IOE). .......... 40
Figure 9: Area under Curve (AUC) showing the Success Rate ............................................. 42
Figure 10: Area under Curve (AUC) showing the Prediction Rate ....................................... 42

xiii
1 INTRODUCTION
1.1 Background and Motivation

Flooding is a natural phenomenon, considered as one of the major disasters overwhelming

the world today because of its catastrophic and devastating after-effects[1]. Flooding is a
resulting event of an excessive inundation of overland flow[2]. The hazard has been a threat
to all areas of human lives in various ways such as destruction of lives and properties,
wrecking of nations' economy, severe damages to transportation systems, and the alteration
of biodiversity live patterns[3]. The exponential and severe economic losses caused by
flooding every year globally amounts to over $20 billion with over 3000 fatalities and
losses[4]. Regrettably, flooding has been labelled the costliest natural hazard because of the
resulting high economic losses ranging to about 31%[5].

Similarly, the West African States (WASs) are not exempted from this cataclysmic hazard.
These states are not far-fetched from the hazard due to rapid indiscriminate development
and population increase[6], [7]. Based on past studies, these states are more vulnerable
because of bad infrastructural planning, low level of technology, and political instability
which leads to the creation of poorly resourced policies[2], [8]. However, these factors are
aggravated by climatic, topographic, geomorphologic, and anthropogenic factors[9]. It has
been recorded by past studies that fluvial and coastal flooding are the major flooding types
that are paramount within the region which is of major concern to the urban residents and
government authorities affected by this hazard[10]. According to the international disaster
(EM-DAT) database, over 1,803 deaths have resulted from flooding in the last 30 years in
Nigeria alone[5], [11].

Furthermore, studies have revealed that flooding occurs more in developing countries such
as the WASs due to the lack of understanding and poor knowledge about flood mitigation
measures and how to tackle it[12]. It makes it almost impossible to predict, mitigate, control,
and manage coastal, fluvial, and flash floods[13]. Based on EM-DAT database, floods in the
WASs last for 79 days on averagely and researchers have opined on the major roles of
anthropogenic activities on the occurrence of floods[11], [13], [14]. It is observed that the
unsystematic development within the cities, blocked drainage systems, and unsystematic
waste disposal methods are a few of the anthropogenic factors that cause flooding[12]. These
factors have been difficult to control because of the increasingly growing population with
spaces inadequate to contain the people which leads to the construction of houses on

1
waterways and the filling up of water channels[15]. Consequently, a series of developments
and urban expansion occurs in inappropriate locations which relate to the influence of
urbanization on the reshaping of the natural environment[16].

Considering the WASs, although several policies are being established and various
community-based approaches are being adopted, climate change continually fuel up the
occurrence of these flood events and urbanization upsurges the threats of these events[6],
[9]. As urbanization increases, so does imperviousness which increases run-off speed and
exacerbates flooding[14]. The focus of these countries has always been towards the high
amount of rainfall while other major factors that drive the occurrence of flooding are most
times neglected. The major approach often adopted deals with the rainfall-runoff approach
(hydrological modelling) which sometimes involves water percolation rate and drainage
systems while other major factors are not considered because of the lack of datasets which
could reveal the spatial and temporal variations of the major influencing factors of flood
occurrences[13].

Consequently, there has been a high limitation and low performance of flood susceptibility
modelling (FSM) in West Africa and the hydrological models utilized requires high-
resolution quality datasets such as LIDAR, and heavy, complex algorithms with high
computation costs in modelling flood occurrences[17]. Thus, accessibility to geospatial
datasets is a paramount issue in WASs as acquiring these datasets still seems limited, and to
fully obtain an accurate assessment of flooding requires the identification of flood
influencing factors characteristics which are largely based on remote sensing data to develop
a flood susceptibility map and identify areas prone to flooding[18]. However, there have
been recent developments that have provided the platforms in acquiring these datasets such
as the recently launched Sentinel satellite from the European Space Agency (ESA) and the
Global ASTER DEM from NASA which gave insights into this study and has provided the
opportunity for solving the problems faced in this part of the world.

Fortunately, machine learning models uplifts this burden due to its flexibility with non-
linear data such as floods and has proven superiority with success in the modelling of natural
hazards[19], [20]. More so, utilizing the ML approach is cost-effective and practical in data-
scarce areas[3]. Furthermore, it is ascertained that creating hybrid models through the
integration of statistical models with machine learning models saves time and reliable
results are attained[21], [22]. Therefore, this research seeks to fill this gap by investigating
not just the impact of natural factors but also the human-induced factors on flood

2
susceptibility and utilizing machine learning approach for modelling flood susceptibility in
the region.

Considering the study’s context, seven models were utilized in identifying and predicting
areas prone to flooding. These models are Index of Entropy (IOE) stand-alone model,
Decision Tree (DT) stand-alone model, Support Vector Machine (SVM) stand-alone model,
Random Forest (RF) stand-alone model, DT-IOE, SVM-IOE, and RF-IOE hybrid models.
Fifteen flood influencing factors, 139 flood locations, and 139 non-flood locations were used
in training the seven models. Subsequently, flood probability indexes generated from the
models were used in deriving the flood susceptibility maps. Thereafter, results attained were
validated using the ROC curve and several efficient statistical indices. This study was
conducted in Lagos city which is arguably the most affected city in the region affected by
flood incidences (Figure 1).

Figure 1: Flooding in Lagos[23], [24]

1.2 Research Gap Identification

Flood is a global disaster that has been studied by several researchers, focusing on its
mitigation and modelling. However, in recent years, through the utilization of geospatial
technologies (GI) which is composed of GIS, remote sensing, and Global Positioning System
(GPS), flood susceptibility mapping accuracy has increased. GI technologies have been
integrated with machine learning (ML) algorithms to model areas susceptible to floods[25],
[26].

To facilitate the reduction of flooding necessitates previous identification of factors

influencing the occurrence of floods and areas that are highly susceptible to flood risks[22].

3
Accordingly, highly accurate flood susceptibility maps should be considered as an essential
resource in managing flood risk. Therefore, the distinction of this research from other
studies is the full consideration of factors that influences the occurrence of floods. Different
studies have considered the geomorphologic, hydrological, and climatic factors which are
often categorized as natural-caused factors such as slope, aspect, Topographic Wetness
Index (TWI), Stream Power Index (SPI), curvature, altitude, rainfall, Land Use Land Cover
(LULC), NDVI, geology and while major anthropogenic factors are mostly not considered.
This is because flood influencing factors vary based on geo-environmental characteristics of
the study area and factors are often selected based on existing works done in the study
area[21].

However, while flooding is a cataclysmic and recurrent hazard in West Africa, no previous
studies have been performed paying attention to flood susceptibility prediction. Studies have
been performed emphasizing hazards description and awareness, risk, vulnerability, and
feasibility studies[7], [12]. Thus, there is low knowledge and limitation on the mitigation
capabilities of flood occurrences in the region. Also, these studies are conducted using
secondary data sources, social and descriptive analysis as research instruments, while a few
flood hydrological modelling studies performed, are conducted without a huge focus on
flood influencing factors[6], [17].

Therefore, it is of paramount importance to further investigate the impact of flood

influencing factors on flood occurrences to develop floodplains management approaches,
lay down more refined policies, and provide more knowledge on the influencing factors.
According to the United Nations, improper flood control techniques and land use, and bad
urban planning practices intensify flood occurrences[27]. On a global scale, few studies have
related the impact of urbanization on the occurrence of floods and an increase in runoff due
to rapid urbanization, high population, enormous deforestation which drives floods
occurrences.

As such, anthropogenic factors such as Normalized Difference Building Index (NDBI),

population density, drainage density, distance from roads while distance from river as a
natural-caused factor should be considered in flood susceptibility modelling. Past studies
have related that anthropogenic factors play a huge role in the region under study which
necessitates considering them in this study[28]. Also, apart from predicting the areas
susceptible to flooding through the implementation of machine learning models, it is of high

4
necessity to identify the impact of each flood influencing factor on flood occurrence relative
to the study area which is also implemented in this study.

In summary, this study attempts to assess the influence of flood influencing factors (natural-
caused and human-induced) on flooding to predict future flood occurrences thereby
providing key and valuable information on flood susceptibility zones and measures that can
be adopted in mitigating its occurrence.

1.3 Research Aim

1.3.1 Aim
The aim of this research study is to develop and utilized machine learning-based flood
susceptibility models in creating flood susceptibility maps to identify locations prone to
flooding considering the human-induced and natural-caused factors while also
acknowledging the impact of these factors on the area under study.

To achieve this main research aim, the following research sub-questions were addressed:
1. What is the impact of both natural-caused and human-induced influencing factors
on the occurrence of floods in the study area?
2. Which machine learning technique is appropriate based on accuracy to predict areas
susceptible to flooding when natural and human factors are both considered?

1.4 Methodology Overview

Within the context of the aim and research questions, the following methodology was
adopted:

• Creation of geospatial database through the derivation of flood predictors

(influencing factors) and flood inventory map.
• Feature selection of variables and multi-collinearity analysis to ensure each factor’s
predictive ability and significance.
• The implementation of bivariate and multivariate statistical model analysis
adopting the IOE technique to attain each factor’s influence on flood occurrence.

5
 • Implementation of machine learning algorithms for model training using DT, SVM,
and RF stand-alone models and hybrid models through the models’ integration
with IOE.
• Derivation of flood susceptibility maps.
• Model evaluation and performance metrics using Area Under Curve (AUC) and
statistical Indices namely, Accuracy, Sensitivity, and Specificity, NPV and PPV.

1.5 Thesis Structure

• Chapter 2 (literature review) explores the past related works on flood influencing
factors that drive flood occurrences and flood susceptibility modelling (FSM)
approaches.
• Chapter 3 describes the study area, the preprocessing of primary datasets required
to create the geospatial database for performing FSM and tools utilized for the
research.
• Chapter 4 (Methodology) details the implementation procedures performed to
produce the final susceptibility maps and fulfil the intents of the research.
• Chapter 5 presents the results acquired.
• Chapter 6 details the critical analysis of the results attained relating them to
literature, the limitations encountered and recommendations for future steps in the
research domain.
• Chapter 7 details the research’s summary with its main findings presented.

6
2 LITERATURE REVIEW
This chapter focuses on three sections which conceptualize within the framework of this
study. Section 2.1 is devoted to flood influencing factors and are further categorized into two
sections based on natural and man-made driven variables. Section 2.2 details various
approaches that have been utilized in the performance of flood modelling in the FSM
domain. The section further details the pros and cons of each approach and how the
approaches have been integrated to perform flood modelling. Finally, section 2.3 comprises
various machine learning techniques that have been utilized in flood modelling with their
strengths and weaknesses and weighing more on the algorithms adopted in the study.

2.1 Flood Influencing Factors

Flood influencing factors are referred to as triggers that enhances the occurrences of floods.
Zhao et al. (2019) noted that the identification of influencing factors is a major procedure in
flood susceptibility assessment[29]. Influencing factors are often chosen based on past
related work in the study area where the most important factors have been identified as
factors vary from one region to another based on the geo-environmental characteristics
(topology, geology, hydrology, and anthropology) of the study area[22]. Moreover, there is
no consensus on the set of influencing factors or the number of influencing factors enough
for FSM[30].

Therefore, Flood influencing factors are commonly chosen based on previous studies and
expert knowledge. However, it is of relative significance to acquire the geographical
information related to the catchment area and its environs in flood modelling as urban
catchment areas are composed of natural and artificial substances[31]. As such, each of the
factors relatively important is categorized into natural-caused and human-induced and each
factor is described below accordingly.

2.1.1 Natural-caused Factors

Natural influencing factors such as elevation, slope, curvature, stream power index (SPI),
topographic wetness index (TWI), land use land cover (LULC), normalized difference
vegetation index (NDVI), rainfall, lithology, and soil have been utilized in previous
studies[18], [32]. Elevation is crucial in flood occurrence as areas with high elevation
enhances an increase in runoff while flat areas are often more prone to flooding due to high
water discharge[1]. Dodangeh et al. (2020) pointed out that a negative correlation exists

7
between flooding occurrence and elevation[33]. Slope influences surface runoff and water
percolation rate into the soil as water flows from a higher elevation to a lower one[1].

Khabat et al. (2018) attested that the increase in slope causes a decrease in runoff
infiltration[34]. Consequently, runoff velocity is dependent on the slope impact as steep
slopes have low water percolation because of a high increase in runoff velocity[35].
Curvature is a morphometric factor that influences the occurrence of floods as divergent and
convergent runoff areas are identified through curvature. It is categorized into three classes
namely, concave (negative curvature), flat (zero curvature) and convex (positive
curvature)[4]. It is observed that flooding occurs mostly in flat and concave areas[33]. Areas
with concave and flat shapes are more susceptible to flooding as such areas retain water
longer than areas with convex shape[36]. Soil characteristics differ from one region to
another based on the different composition of particles which determines the level of water
percolation. Its texture, type, and structure account for the runoff rate and level of water
infiltration.

Furthermore, TWI describes the flow of water towards the pull of gravity within a watershed.
The factor accounts for the accumulation of water in lower slope areas[16]. K. Chapi et al.
(2017) defined TWI as the ratio of a specific basin area to the slope[34]. TWI identifies areas
within a watershed that are prone to flooding as areas with a steeper slope have lower
percolation rate unlike flat terrain[22]. TWI, therefore, indicates percolation status in a
region and areas prone to flooding. SPI measures runoff’s water flow erosive power[20].
Consequently, areas with highly concentrated surface runoff and high erosive power are
identified[20]. It identifies the strength of flood flowing towards gravity and the amount of
water accumulated in the watershed as the steeper the slope, the increase in velocity of the
water flow[37]. Therefore, areas with a high tendency for flow accumulation indicates high
value while low values indicate areas with low flow accumulation[38].

Also, LULC plays a vital role in flood occurrence; urban areas are more peculiar to flooding
through increase runoff rate due to imperviousness of its surfaces while vegetated areas are
often less flooded because of the high vegetation density. Consequently, an inverse
relationship exists between vegetation density and flood occurrences[16]. Previous studies
indicated that land-use patterns play a major role in flooding and should be considered in
flood studies as each LULC type performs a specified role in flooding[39]. Lithology is an
important factor considering spatio-temporal variation where high underlying resistant
rocks or highly penetrable particles determine the drainage density rate of the area. It also

8
controls and determines the amount of sediment transported, speed of runoff, and
percolation rate[20], [33].

Also, rainfall in most studies has been referred to as the most important influencing
factor[40]. Rainfall has a remarkable influence on flood occurrence through its spatial and
temporal patterns, accordingly, an increase in the amount of rainfall leads to a significant
tendency of flooding[41]. Tehrany et al. (2019), noted that flooding is mainly originated by
rainfall and further influenced by other factors, and the amount of rainfall drives water
inundation depending on the characteristics of the basin such as its expanse, altitude and
the LULC formations[22], [42]. Distance from river plays a major role in flooding and
significantly determine its extent and magnitude[16]. Haoyuan Hong et al. (2018)
maintained that river initiates flooding when the amount of water exceeds the amount the
river network can handle[20]. Esmaeel Dodangeh et al. (2020) also emphasized that the
closer it is to a river, the increase the risk of flood occurrence[33]. Kamran Chapi et al. (2017)
revealed that frequent locations most affected by floods are areas close to the river[34]. Thus,
it is necessary to consider distance to river as an influencing factor.

In conclusion, NDVI represents the vegetation density and indicates the vegetational
characteristics of the study area, it is observed that high vegetation reduces flooding[33]. A
higher NDVI increases the possibility of water percolation into the soil and reduces the
possibility of flooding[19]. Consequently, a decrease in NDVI automatically increases the
probability of flooding.

2.1.2 Anthropogenic Factors (Human-Induced)

Previous studies have related that population density, drainage density, normalized
difference building index (NDBI), and distance from roads as anthropogenic factors play a
role in flood occurrence[43], [44].

Drainage density is defined as the total steam length(m) by the total basin area (𝑘𝑚2 ) of a
watershed. As a result, a high tendency of flooding in areas are identified with high drainage
density[36]. Therefore, high drainage density is positively correlated with high flood peaks
and volumes[34]. H. Shafizadeh-Moghadam et al. (2018), noted that drainage density
describes how well-drained or poorly drained the watershed is[19]. Idowu et al. (2020) also
corroborated that substandard drainage networks aggravate the occurrence of flooding and
Augustine (2017) opined that poor drainage systems cause continual flooding and stream

9
overflow in an area[7], [12]. Also, Mahmoud et al. (2018) revealed that a dense drainage
network coupled with a steep slope often leads to continual flooding[36].

Previous studies have related population density to have a significant influence on flood
occurrence as an increase in population within a region causes a significant decrease in
pervious spaces and an increase in urban growth[9], [15]. The Global Assessment Report on
Disaster Risk Reduction (UNISDR, 2019) revealed that the high increase in population
growth has caused an increase in flood risk[27]. This has been attested by previous flood
modelling studies[45]. Population density is defined as the total number of people
occupying a given region per unit area[46]. Hamid Darabi et al. (2019) noted that flood
occurrences are highly associated with high population density and should be significantly
considered in flood studies[39].

Furthermore, Distance from road influences the occurrence of floods. Gang Zhao et al.
(2020) related that road enhances water inundation due to its imperviousness and forms a
pathway for water flow[31]. It reduces the percolation rate which increases the runoff
rate[22].

In conclusion, Normalized Difference Building Index (NDBI) indicates the building

attributes within a region[33]. It a major determinant of impervious areas as the
concentration of buildings increases run-off. Therefore, there exists a direct relationship
between flooding and building density. Thus, should be considered as a driver of flood
occurrence[38].

2.2 Flood Susceptibility Modeling Approaches

Flood Susceptibility Modelling (FSM) has been achieved through various approaches.
However, each approach has its pros and cons and generates different results. Therefore, no
universal consensus is laid down on model selection for FSM. Each model differ based on
sensitivity to outliers, prediction accuracy, processing time and presumptions on data
distribution[19]. Consequently, each of the flood modelling approaches is discussed below.

2.2.1 Hydrological Approach (Physical-based)

In the past decade, traditional hydrological and hydraulic modelling approach has been
adopted by researchers for susceptibility mapping[47]. This modelling approach is
categorized into three namely; one-dimensional model (MIKE 11, HECRAS), two-
dimensional model (TUTFLOW, SOBEK) and three-dimensional model (Navier-
Stroke)[13]. However, with this approach, fieldwork is essential and highly costly for data

10
gathering[47]. According to Balicia et al. (2013), highly comprehensive and detailed data is
needed to achieve significant accuracy[48]. Moreover, hydrological models have been
ensembled with GIS, which has proven its capability for flood modelling[47]. The model
requires high computational time based on the model’s dynamics and the 2D model
performs more accurately than the 1D model in flood modelling but requires very long-term,
high-resolution data which are burdensome to acquire and prevents short-term
prediction[31]. It is revealed that most authors do not use the three-dimensional model to
avoid unessential complex algorithms since some less complicated models can provide
reasonable solutions[13].

2.2.2 Qualitative Approach

The qualitative approach incorporates expert knowledge and qualitative techniques to relate
independent variables with flood occurrence based on numerical expressions[20]. Some of
the popular qualitative techniques are the Analytical Hierarchy Process (AHP)[44], fuzzy
logic[20], and Multi-Criteria Decision Analysis (MCDA)[47]. Tehrany et al. (2019), noted
that qualitative models incorporate expert’s opinion for its modelling considering the
influencing factors and their attributes which could generate bias in the prediction
modelling. The author noted that flooding is a global problem and should be predicted with
an efficient and robust modelling approach[35].

However, most studies optimize this approach through the integration of the qualitative
models with various decision analysis algorithms, statistical models and machine learning
models[19]. Hossein et al. (2016) applied GIS ensemble method of FR and SVM to create
flood susceptible mapping in Malaysia where each influencing factor is optimized by MCDA
technique to generate weights that serve as inputs for SVM model[47]. Rahmati et al. (2016)
utilized the MCDA technique for flood modelling and generated flood susceptibility
maps[49]. However as opposed by Tehrany et al. (2019), the method is unsuitable for flood
susceptibility studies as expert knowledge is integrated into the model, it was further
attested by Romulus et al. (2020), who utilized fuzzy AHP model which achieved low
performance based on expert’s judgement involved in the modelling[4].

Rahmati et al (2016) however noted that AHP is simple, budget-friendly, less time-
consuming and easier to develop for flood susceptibility studies and more suited for regional
studies[47]. Samantha et al. (2018) also adopted MCDA which was conducted and
compared with the FR model. The author related that FR model had a better performance
than MCDA[25]. Hong et al. (2018) adopted the fuzzy logic technique for FSM and revealed

11
that the technique chosen does not quantify the variables’ importance as the variables were
chosen based on expert judgement on the study area. He further argued that incorporating
field experience and expert judgements generate more accurate results[20].

2.2.3 Statistical Approach

The statistical approach has been utilized by researchers in recent times, some of these
methods are frequency ratio (FR)[21], logistic regression(LR)[50], weight of evidence
(WOE)[51], Index of Entropy (IOE)[52] and Evident Belief Function[22]. This approach
works on a presupposition that flood influencing factors are associated with, and drives the
occurrence of flood events[1]. Tehrany et al. (2019) opined that most statistical methods
rely on linear presumption while flood is a multidimensional phenomenon, the author
related that ensemble statistical methods augment this flaw such as the adaptive neuro-fuzzy
inference system (ANFIS) however, this model requires various parameters to perform FSM
accurately[35].

This was counteracted by Pradhan et al. (2015) who LR and noted that LR utilizes both
continuous and discrete variables in FSM thereby describing the flexibility of the model[30].
Statistical modelling techniques are categorized into bivariate and multivariate statistical
models. Bivariate statistical models such as FR and WOE are probabilistic models which
measure the occurrences of flood based on each class of the influencing factors[45].
Therefore, the bivariate probability is calculated by correlating each class of the influencing
factors with flood occurrence and the higher the bivariate probability the stronger the impact
of that factor on flood occurrence[53]. However, it is noted that the bivariate approach is
based on generalization as interaction among factors are not considered and weights not
assigned to each factor[16]. This presumes that flood occurrence is based on the same set of
factors with equal weights across the study area[21], [26].

On the other hand, multivariate statistical model such as LR correlates each influencing
factor directly with flood occurrence and performs correlation among the influencing
variables[50]. However, the weak point of statistical models is its inefficiency in handling
complex and multidimensional phenomena because of its linear structure[54]. Moreover, It
is noted that the combination of FR and LR increases the efficiency of the model and cover
up the weakness of both models[30]. Furthermore, hybrid models formed from the
integration of statistical and machine learning models have been ascertained by past studies
to yield more accurate results[4]. According to Hong et al. (2018), hybrid model’s flexibility
allows extensive assessment of influence on each flood-related independent variables in

12
each class[20]. Wei Chen et al. (2020) related that increase in accuracy of flood studies
requires the precise identification and greater prediction capabilities in forecasting future
flood occurrences which is achievable through the hybridization of statistical and machine
learning models[40].

2.3 Machine Learning Modelling Approach

Machine learning (ML) is an efficient approach to the prediction of natural hazards[55]. The
approach helps to detect and uncover insights, relationships among variables which makes
it suitable for predictive modelling. Various methods have been applied to flood
susceptibility modelling (FSM). Models such as random forest (RF)[38], decision
trees(DT)[21], support vector machine(SVM)[22], artificial neural network (ANN)[56],
multivariate adaptive regression splines (MARS)[33], neuro-fuzzy inference system
(ANFIS)[54], have been utilized for FSM studies. These models have been compared with
their performances noted. However, to date, there is no superiority among the models as
each has its pros and cons[26]. In recent times, ML approaches are being integrated with
geospatial technologies in flood mapping studies for handling more complex phenomena
and computing large amounts of data accurately[54]. ML techniques have the advantage of
predicting and modelling complex structures in a proficient manner[16].

Furthermore, in recent times hybridization of machine learning models are being adopted.
The aim of utilizing the hybrid models is to increase the predictive capability and precise
identification of areas susceptible to flooding[4]. Thereby, the influence of the conditioning
factors on flooding can be detailly assessed. According to Amir Morsavi et al. (2018), the
hybridization of machine learning models enhances performance and increases
accuracy[55]. Romulus et al. (2019) noted that no single method is appropriate for flood
modelling and revealed that the utilization of hybridized machine learning models
eradicates the weaknesses of ML models and generates more accurate results[52].

ML approach learns the relationship between the flood influencing factors and flooding
occurrences without subjecting to an expert opinion which often leads to bias[38]. ML
models are advantageous in assessing any kind of data type (categorical, nominal, and
continuous) which makes the algorithm more flexible[43].

However, as revealed by researchers in various instances, ANN and ANFIS have been
adopted in flood studies, and both techniques are robust in the presence of outliers and
efficient in handling errors in the input dataset. However, it is noted that these algorithms

13
are difficult to understand and implement even with high versatility on incomplete
datasets[35]. Sayed Tameh et al. (2018) supported this fact as tuning of function parameters
complicates the utilization of the algorithm and noted that it is essential to optimize the
parameters with other algorithms to increase its flexibility[16]. A brief theoretical
background and overview of a few notable ML algorithms are detailed below.

2.3.1 Decision Tree (DT)

DT is an efficient ML technique that has performed effectively in flood modelling[3]. Its
procedural approach is easy to create and interpret. Although decision trees consume a lot
of time in classifying and computing, the algorithm can deal with uncertainties to a
significant extent in a dataset[35]. The algorithm is flexible in handling data with various
scales, assumes no statistical distribution and has high efficiency in creating rules for
predicting complex relationships[16]. DT algorithm classifies the influencing factors in a
hierarchical manner and equivalently in accordance with the susceptibility levels, and create
decision rules based on an established tree structure built on the significant levels of the set
of independent parameters utilized[35].

Thus, the set of parameters are analyzed to generate an outcome. Decision trees have been
integrated with different algorithms such as the naïve bayes tree (NBTree)[3] and
alternating decision tree (ADT)[18] and other processing techniques such as the
Classification and Regression Trees (CART)[51], Chi-squared Automatic Interaction
Detection (CHAID)[57], Unbiased Efficient Statistic Tree (QUEST)[39]. Each of these DT
classifiers has been used in modelling studies and each has its pros and cons. The NBTree is
often combined with the J48 algorithm to increase the predictive capability of the
algorithm[34].

2.3.2 Random forest (RF)

RF is a classification and regression modelling approach. RF is based on a fusion of random
subspace method and bagging ensemble learning[38]. RF algorithm combines decision trees
in predicting an outcome by permuting each variable randomly and the prediction results
acquired is compared with each variable to obtain its significance. In this context, a training
dataset D = ((A1, B1), …, (𝐴𝑛 , 𝐵𝑛 )) which contains the n vectors. A ∈ X and B ∈ 𝑌 where X
represents numerical observations and Y represents the class labels (flood and non-
flood)[3], [58]. RF works based on two processes namely bagging and random selection.
This is performed to prevent overfitting within the dataset and to increase the predictive
ability of the model. The process is popularly referred to as the out-of-bag procedure where

14
samples within the dataset are randomly drawn and replaced till a most minimal error is
achieved[14]. RF is versatile in handling inconsistent and missing data and performs well
with multi-collinearity and the performance of the model is based on the number of decision
trees (𝑁𝑡𝑟𝑒𝑒) and variables attributes in the subset (𝑀𝑡𝑟𝑦)[34]. An increase in the value of
𝑁𝑡𝑟𝑒𝑒 increases the time consumed in modelling while the modelling becomes more prone
to errors when the 𝑁𝑡𝑟𝑒𝑒 is small[59].

2.3.3 Support Vector Machine (SVM)

SVM is a high predictive model with high versatility, till recent, SVM has not been explored
in FSM[55]. It relies on a statistical learning approach that generates output values from a
number of input values[22]. It is a supervised machine learning technique that converts non-
linear structures into linear by generating a hyperplane to simplify and distinguish classes
in the data while categorizing the data into the training and testing dataset[35]. SVM, based
on predictive accuracy, is suitable for FSM as it handles data independently of the
measurement scales and works efficiently with any data format[22]. Hong et al. supported
this fact and revealed the effectiveness of SVM in classifying linear and non-linear data[20].

Researchers, however, noted that the efficiency of SVM performance highly depends on the
kernel adopted and the influencing factors adopted[30]. The SVM kernels mostly used are
linear kernel (LN), polynomial kernel (PL), radial basis function (RBF), and sigmoid kernel
(SIG)[35]. Radial basis function kernel (RBF) has often been implemented in previous
studies because of its efficiency and high accuracy. Hong et al. (2018) noted that SVM is
disadvantageous because of its inability to measure the significance of attributes chosen[20].
However, this drawback was resolved by Tehrany et al. (2019) who utilized kappa index with
the SVM model to detect the significance of the attributes[22]. A kernel function is used in
distinguishing and transforming the data[21]. Afterwards, the main input of the training
dataset is mapped into a high dimensional space where the split hyperplane is created in the
original space of n coordinates to distinguish between points of two different classes (flood,
non-flood)[22].

According to Tehrany et al. (2019), several hazard modelling studies have revealed that DT
is moderately robust than SVM in modelling natural hazards[35]. However, other studies
reveal that both ML algorithms are both efficient and robust in hazard modelling and offer
similar results.

15
3. DATA AND CASE STUDY
The following chapter describes the area of study, datasets and tools utilized in the study.
Furthermore, it focuses on the geographical derivations and the creation of the geospatial
database that was used in performing the flood modelling. Section 3.1 defines the study area
while section 3.2 details the preprocessing and the creation of the flood inventory dataset,
section 3.3 details the flood influencing variables derivation and analyses majorly from
satellite imageries using ArcGIS Pro’s spatial analysis and spatial statistical tools.

3.1 Case Study

The study focuses on the entire region of Lagos state, a metropolitan coastal low-lying city
located in the South-western region of Nigeria, West Africa. It is regarded as the economic
hub of Africa. The city’s geographical coordinate ranges between 3.1º to 3.4º E longitude
and 6.5º to 6.8ºN latitude[12]. The city experiences an equatorial (humid and hot) climate
with a double-maxima rainfall all through the year. The region’s climate has two distinct
periods namely the rainy season (April – October) and the dry season (November –
March). The city is composed of mangrove swamp and forests where the mangrove
swamps dominate the south while the forests are majorly found in the northern areas of
the region[7].

The city is composed of highly dense road networks with inland waterways and her southern
boundaries are defined by about 180 km of Atlantic coastline and a border along the western
perimeter with the Republic of Benin. The city has a total landmass of about 2,345km2 which
represents about 0.4% of Nigeria’s total land area[10]. It is a highly populated region, and
its population density continues to increase due to its commerciality. Based on its rapid
urbanization, there has been a series of expansion towards the creeks and lagoon within the
city[28]. Recurrently, flood occurs within the city with a destructive impact on lives and
properties which forces the evacuation of people from their residences. Thus, regarded as
one of the West-Africa cities highly prone to flooding.

16
 Figure 2: Case Study

3.2 Datasets and Preprocessing

3.2.1 Derivation of Flood Inventory Dataset

Flood susceptibility mapping demands two sets of data. The first dataset constitutes the past
flood locations (flood inventory) which indicate the past flooded regions while the second
dataset constitutes the flood triggering parameters otherwise referred to as the flood
influencing factors[32]. Flood susceptibility is based on the impact of the flood influencing
factors on the occurrence of floods which entails assessing the significance of each
contributing parameter on flood occurrence. In this study, the flood inventory dataset was
provided by the Lagos State Emergency Management Agency (LASEMA) which comprises
past flood location events within the study area from 2010 to 2020 and was augmented with

17
satellite imageries, flood historical records and data derived from the international disaster
(EM-DAT) database. A total of 139 flood events were located within the study area and an
equal number of 139 non-flood points was created across the study area using the ‘create
random point’ tool. To verify the correctness of the flood locations and non-flood locations,
the Normalized Difference Water Index (NDWI) was calculated from Landsat 7 ETM+,
Landsat 8 OLI, and Sentinel 2 satellite imageries using the near-infrared (NIR) and short-
wave infrared (SWIR) bands to identify the past flood locations from 2010 to 2020. It is
calculated as follows:

𝑵𝑰𝑹−𝑺𝑾𝑰𝑹
NDWI= (1)
𝑵𝑰𝑹+𝑺𝑾𝑰𝑹

The flood locations were represented as points as polygon formats yields exaggerated results
and become complicated for the algorithms utilized[50]. Therefore, the points were
pinpointed on the centroid of the flooded areas. This has been further proven by hazards
modelling studies which utilized the point format for flood inventory and generated accurate
results[20]. The flood inventory map (Figure 3) was further divided into training and testing
dataset as required for the training and validation stages, respectively. There is no consensus
on how the inventory data is classified as it is highly dependent on the availability and quality
of data. Space robustness and time robustness are two standards for classifying flood
inventory. Time robustness involves dividing flood inventory data into two periods of past
occurrence and future occurrence which represents training and testing datasets,
respectively[22].

However, acquiring temporal data is burdensome as each flood occurrence is associated with
the precipitation that triggered it and goes for other spatio-temporal influencing parameters.
In space robustness, flood inventory data is randomly divided into two datasets namely
training and testing datasets [35]. In this study, both standards were integrated based on
the flood inventory data available. Therefore, the flood inventory data was divided into 70%
for training and 30% for testing based on the 2010 to 2020 flood data. Flooding is a binary
classification modelling; therefore, it was required to create equal 139 non-flood location
points to ensure consistency and accuracy. Consequently, the 139 flood and 139 non-flood
locations were divided using random selection technique into training and testing datasets,
respectively. Values 0 and 1 were assigned to the non-flood and flood points respectively,
where 0 represents flood non-existence and 1 represents flood existence.

18
 Figure 3: Flood Inventory Map

3.2.2 Derivation of Flood Influencing Factors Datasets

Satellite imagery has successfully proven to play a major role in hazard assessment based on
its combination of spatial, spectral and temporal resolution which identifies past and present
occurrences of hazards[4]. Sentinel 2 imagery was utilized in this study to derive the flood
predictors namely LULC, NDBI, NDVI. The Sentinel-2 imagery provided a spatial resolution
10m-60m), multi-spectral features (13 bands) and temporal resolution (five days with two
satellites at the equator)[60]. The LULC was obtained from the classification of the Sentinel
2 imagery using the maximum likelihood algorithm which achieved an overall accuracy of
89%. The Sentinel-2 imagery was acquired on October 15, 2020, and a spatial resolution of
10m and four spectral bands; red (B4), green (B3), blue (B2) and NIR (B8) were used in
deriving the LULC map. Google Earth imagery was used as the training data and a total of
14 LULC classes were distinguished within the study area.

On the other hand, the NDVI and the NDBI were also calculated using the SWIR (B11)
bands, NIR bands and red bands. The formula used for calculating both indices are as
follows:

𝑺𝑾𝑰𝑹−𝑵𝑰𝑹
NDBI= (2)
𝑺𝑾𝑰𝑹+𝑵𝑰𝑹

19
 𝑵𝑰𝑹−𝑹𝒆𝒅
NDVI= (3)
𝑵𝑰𝑹+𝑹𝒆𝒅

20
Figure 4: Flood influencing factors; (a) Distance from River, (b) Distance from Road, (c) Altitude,
(d) Curvature, (e) Geology, (f) LULC, (g) NDBI, (h) Drainage Density, (i) NDVI, (j) Rainfall, (k)
Population density, (l) Slope (m) Soil, (n) TWI, (o) SPI.

21
 Furthermore, a digital elevation model (DEM) for the watershed was obtained from the
advanced spaceborne thermal emission and reflection radiometer (ASTER Global DEM)
30m database. The following flood predictors such as elevation, curvature, slope, TWI, SPI,
distance from river (river network), and drainage density were derived from the DEM (Table
1).

Influencing Source of Data Data Scale and Data Source

Factor Type Resolution
Rainfall UEA climatic research unit GRID 30m https://sites.uea.ac.uk/cru/data

Altitude Derived from DEM GRID 30m https://earthdata.nasa.gov/learn/articles/new-

aster-gdem
Curvature Derived from DEM GRID 30m
“
Slope Derived from DEM GRID 30m
“
TWI Derived from DEM GRID 30m
“
SPI Derived from DEM GRID 30m
“
Drainage Derived from DEM GRID 30m
Density
“
Distance from Geofabrik Website Line 30m https://www.geofabrik.de/
Road Coverage
(Open Street Map)
LULC Classifying Sentinel-2 GRID 30m https://scihub.copernicus.eu/
Imagery.
Soil Digital Soil map (DSM) Vector 1:250,000 https://data.mendeley.com/datasets/zmrt6k83
database wk/draft?a=d9a35c1e- c19b- 4ddd- b34e-
69674a8ceb18
Geology Digital Soil map (DSM) Vector 1:250,000 https://data.mendeley.com/datasets/zmrt6k83
database wk/draft?a=d9a35c1e- c19b- 4ddd- b34e-
69674a8ceb18
Population City population Website GRID 30m http://www.citypopulation.de/

Distance from Derived from DEM GRID 30m https://earthdata.nasa.gov/learn/articles/new-

River aster-gdem
NDVI Derived from Sentinel-2 GRID 30m https://scihub.copernicus.eu/
Imagery
NDBI Derived from Sentinel-2 GRID 30m https://scihub.copernicus.eu/
Imagery
Table 1: Spatial Datasets and Data sources

22
Curvature was classified into three classes namely concave (positive value (+)), flat (value 0)
and convex ((negative value (-)). The spatio-temporal factors such as the SPI and TWI was
calculated from as DEM as follows:

TWI = In (𝐴𝑠/𝑡𝑎𝑛 𝛽) (4)

SPI = 𝐴𝑠𝑡𝑎𝑛 𝛽 (5)

Where 𝐴𝑠 equals the area of the catchment(𝑚2), and β (radians) equals slope gradient[22].
Furthermore, the drainage density was derived using the following equation stated below:

𝟏 𝒔
DD= ∑𝒊 𝑳 (6)
𝑺 𝒊

Where S equals the study area and Lsi equals the length of the river within the study area[36].
The population data was obtained from the city population website[61] which was processed
using an areal interpolation tool to derive the population raster map. The mean annual
rainfall data from 2010-2020 was obtained from a high resolution spatial gridded dataset
provided by the climatic research unit, university of East Anglia[62]. The road data was
obtained from the open street map through the Geofabrik website[63], while the river
network was derived from the DEM using the flow accumulation and flow direction tool and
distance to both attributes were derived using the Euclidean distance tool.

Soil and Lithology at a scale of 1:250,000 were extracted from the free open adaptable digital
soil map database of Nigeria created by Ugonna et al(2020)[64]. The Soil was categorized
into eight classes identified as Oxic Rhodudalf, Rhodic Paleudult, Oxic Tropudalf, Typic
Tropudult, Typic Tropudult, Oxic Tropudalf, Aquic Paleudult, Orthic Tropaudalf and Typic
Tropaquent while lithology is categorized into nine classes identified as Coastal plain sand
(Alfisols), Coastal plain sands, Recent Alluvium, Sandstone (Abeokuta Formation),
Transitional materials of subrecent alluvium, Coastal plain sands II (Ultisols), Ewekoro
Formation( Upper coal measure), and Deltaic Basin and tidal flats respectively. All the other
predictors were reclassified to the desired classes using the quantile classification method to
create the flood database. The flood influencing factors were all re-processed to a 30 by 30m
pixel size that corresponds to the DEM’s spatial resolution and re-projected to the UTM zone
Minna 31N to create the geospatial database.

23
4. METHODOLOGY
The methodological approach for this research study is divided into six steps and
procedurally detailed as follows: (A) Data derivation, pre-processing and preliminary
analysis as laid out in chapter 3. (B) Feature Engineering (Feature Selection and Multi-
collinearity analysis). (C) Data Cleaning and Normalization. (D) Index of Entropy bivariate
modelling (E) Machine learning models training and the generation of flood susceptibility
maps (F) Models validation using Area under Curve and other statistical indices. A detailed
description of each step is presented in the following subsections (Figure 5).

Figure 5: Methodology Flowchart

4.1 Feature Selection and Multi-collinearity Analysis

This is a process to identify the predictive capabilities of the flood predictors as factors
selected fully depends on the geo-environmental characteristics of the study area and factors
that have low predictive power which could generate outliers and decrease the model’s
predicting ability can be identified and removed[4]. To ensure accuracy as no previous
studies have been done in the region, linear support vector machine was utilized in checking

24
each factor’s predictive ability and significance (Step B). Also, to avoid redundancy among
the factors, multicollinearity analysis was performed.

4.1.1 Linear Support Vector Machine (LSVM)

This is an efficient means of evaluating a predictor’s capability. It has been proven to
enhance the classification accuracy in modelling[4], therefore considered for this study.
Considering the training dataset and the fifteen influencing factors, the LSVM equation is
calculated as follows:

f(x) = sign (𝒘𝑻 a + b) (7)

Where f(x) represents the function upon which the linear support vector machine is derived,
𝑤 𝑇 represents the inverse matrix of weight associated with each flood influencing factor, a =
(a1, a2,….a15) represents the input vector that contains the flood influencing factors, and b
represents the offset of the hyperplane’s origin[22]. The factors are selected based on the
average merit for each factor which ranges from 0 – 1. Consequently, a factor equal to zero
is excluded from the analysis. The LSVM algorithm was implemented in the WEKA
environment using the attribute evaluation tool.

4.1.2 Multi-collinearity Analysis

This is to check if two or more influencing factors are highly correlated which could cause
redundancy and reduce the accuracy of the models utilized for the study[20]. All factors
introduced are very important in terms of anthropogenic, climatic, hydrological, and
geomorphological unique characteristics. Consequently, Tolerance (TOL) and Variance
Inflation Factor (VIF) approaches was utilized where VIF > 10 and TOL < 0.1 standard
indicates multicollinearity in the influencing factors as laid out in existing literature[35].
This approach has proven its efficiency in hazard modelling with a success record. The VIF
measures the correlation among variables by inflating the variance of each variable within
the regression’s coefficient while TOL is the inverse of VIF[20].

4.2 Data Cleaning and Normalization

After the creation of the geospatial database, data cleaning and normalization was required
before performing the flood modelling (Step C). Therefore, seven missing values found in
the database were replaced by the mean of each variable by computing differential statistics
for each of the influencing variable. Also, to avoid varying scales among the variables which
could alter the results of the machine learning algorithm, the min-max normalization

25
approach was utilized with values ranging between 0 and 1. The normalization equation is
represented as follows:

(𝑿−𝑿𝒎𝒊𝒏 )
𝑿′ = (15)
(𝑿𝒎𝒂𝒙 −𝑿𝒎𝒊𝒏

Where X represents the original value, 𝑋 ′ equals the normalized value, and 𝑋𝑚𝑖𝑛 , 𝑋𝑚𝑎𝑥
represents the minimum and maximum values for each influencing variable[60].

4.3 Index of Entropy Modelling

The index of Entropy is a bivariate statistical approach popularly utilized in the modelling
of natural hazards[53]. Previous studies have utilized this method and generated highly
accurate results in susceptibility modelling[51]. Consequently, the approach was adopted in
this study (Step D). Index of entropy measures variabilities and instabilities within a
database. Considering this study, the extent to which flood influencing factors triggers flood
occurrence is represented by the entropy of a flood event thereby generating the weight of
each influencing factors. Therefore, the weight of each influencing factor in the flood
probability index will be ascertained. To perform IOE, each influencing factor was classified
using the quantile classification technique to ensure even distribution of pixels across each
class and a reliable assessment of each class’s impact on flood occurrence is identified.
Furthermore, the IOE model generates two main outputs namely weights associated with
each factor and each factor’s classes. Each factor’s weight (𝑊𝑗 ) is calculated as follows:

𝑭𝑹𝒊𝒋
(𝑷𝒊𝒋 )= 𝑺𝒋 (8)
∑𝒋=𝟏 𝑭𝑹𝒊𝒋

Where 𝐹𝑅𝑖𝑗 represents the frequency ratio coefficient for each class of each influencing
factors; 𝑆𝑗 represents the number of classes, and (𝑃𝑖𝑗 ) represents the probability density[4].

𝑺𝒋
𝑯𝒋 =∑𝒊=𝟏(𝑷𝒊𝒋 ) 𝒍𝒐𝒈𝟐 (𝑷𝒊𝒋 ), 𝒋 = 𝟏, 𝟐, … , 𝒏 (9)

𝑯𝒋𝒎𝒂𝒙 = 𝒍𝒐𝒈𝟐 (𝑺𝒋 ) (10)

𝑯𝒋𝒎𝒂𝒙 − 𝑯𝒋
𝑰𝒋 = , I = (0,1), 𝒋 = 𝟏, … , 𝒏 (11)
𝑯𝒋𝒎𝒂𝒙

𝟏 𝑺𝒋
𝑷𝒋 = ∑𝒊=𝟏 𝑷𝒊𝒋 (12)
𝑺𝒋

𝑾𝒋 = 𝑰 𝒋 ∗ 𝑷 𝒋 (13)

26
Where 𝐻𝑗 𝑎𝑛𝑑 𝐻𝑗𝑚𝑎𝑥 equals the values of entropy; 𝐼𝑗 represents the information coefficient;
𝑃𝑗 represents the empirical probability; 𝑊𝑗 equals the weight values associated with each
flood influence factor.

Considering the stand-alone model, flood probability index final values is calculated, and
the equation is as follows:

𝒁
𝑭𝑺𝑰𝑰𝑶𝑬 = ∑𝒏𝒊=𝟏 ∗ 𝑪 ∗ 𝑾𝒋 (14)
𝒎𝒋

Where i equals the total number of conditioning factors; Z equals the number of classes of
the factor having the highest number of classes; 𝑚𝑗 represents the number of classes of each
factor; C equals the calculated rate of each class; and 𝑊𝑗 represents the final weight of each
factor. The advantage of IOE is it can be used for both BSA and MSA[52] which was
implemented in this study. The final weights derived from each factor were used as inputs
in modelling the hybrid machine learning models.

4.4 Machine Learning (ML) Algorithms

In this subsection, the three ML model’s implementation in this study is briefly explained.
The ML models are adopted to identify the correlation existing between the influencing
factors and flood occurrences and to forecast flood susceptibility in the study area (Step E).
Therefore, SVM, RF, and DT were utilized, and the implementation of each ML algorithm is
described below:
4.4.1 Support Vector Machine
The training dataset contains instance-label pairs (𝑥𝑖 , 𝑦𝑖 ), with 𝑦𝑖 ∈ 𝑅𝑛 , 𝑦𝑖 ∈ {1, −1}, and i
=1, …, m. x signifies the vector within the input space which incorporates all the influencing
factors (elevation, slope, curvature, rainfall, geology, soil, LULC, NDBI, NDVI, TWI, SPI,
population density, drainage density, river, and roads). SVM sets up an optimal hyperplane
that distinguishes and separates flood and non-flood pixels into {1, 0} in the training set.
Separating a hyperplane using linear separable data is defined as:
𝐲𝐢 (𝐰. 𝐱 𝐢 + 𝐛) ≥ 𝟏 − 𝛆𝐢 (16)

Where w represents the coefficient vector through which the hyperplane's orientation is
defined in the feature space, b represents the hyperplane's offset from its origin and
ε𝑖 represents the positive slack variable. Lagrangian multipliers are solved to find an optimal
hyperplane. It is calculated thus:

27
 𝟏
∑𝒏𝒊=𝟏 ∝𝒊 − ∑𝒏𝒊=𝟏 ∑𝒏𝒋=𝟏 ∝𝒊 ∝𝒋 𝒚𝒊 𝒚𝒋 (𝒙𝒊 𝒙𝒋 ), (17)
𝟐

∑𝒏𝒊=𝟏 ∝𝒊 𝒚𝒋 = 𝟎, 𝟎 ≤ ∝𝒊 ≤ 𝑪 , (18)
Where ∝𝑖 are Lagrange multipliers, C is the penalty and ε𝑖 represents the slack variables that
allow the penalized constraint violation. The step-by-step layout of SVM modelling is well
described in Tehrany et al (2014)[65]. The decision function of SVM classification is defined
as:
𝒈(𝒙) = 𝒔𝒊𝒈𝒏{∑𝒏𝒊=𝟏 𝒚𝒊 ∝𝒋 𝑲(𝒙𝒊 . 𝒙𝒋 ) + 𝒃} (19)
Where K (𝑥𝑖 , 𝑥𝑗 ) represents the kernel function. The kernel function is mathematically
expressed as:
K(𝒙𝒊 , 𝒙𝒋 ) = 𝒆𝒙𝒑(−𝜸||𝒙𝒊 − 𝒙𝒋 ||𝟐 ) (20)
The penalty (C) and the kernel width (𝛾) were optimized using the cross-validation approach
to actualize accurate results as opposed to the trial-and-error technique to prevent
overfitting. To perform cross-validation requires dividing the training dataset into n folds
where one-fold is reserved for testing and the remaining folds (n-1) fold is used for training.
The average accuracies of the validation are established and used in generating the final
flood susceptibility model. Therefore, ten-fold cross-validation was used in this study by
dividing the training dataset into 10 random groups till the best values of these parameters
were actualized. The optimized parameters (𝐶, 𝛾) were used in generating the final SVM and
𝑆𝑉𝑀𝐼𝑂𝐸 probability maps. RBF kernel was utilized in this study as its efficiency and advanced
interpolation and extrapolation potentialities have been proven in hazard modelling studies
and various literature sources[22].
4.4.2 Random Forest Model
The random forest modelling depends on the user-defined parameters namely as the 𝑚𝑡𝑟𝑦
and 𝑛𝑡𝑟𝑒𝑒 . The 𝑚𝑡𝑟𝑦 represents the number of variables randomly selected at the split of each
node and 𝑛𝑡𝑟𝑒𝑒 represents the number of trees contained in the model[20], [34]. Therefore,
the 𝑚𝑡𝑟𝑦 values are within the range of [ 1, 15] while the 𝑛𝑡𝑟𝑒𝑒 values are within [500, 1000,
1500, 2000, 2500] range. Ten- fold cross validation method was also utilized in tuning these
parameters to ensure their values are within the number of flood variables range. The tuned
parameters actualized were used in creating the final RF and 𝑅𝐹𝐼𝑂𝐸 probability indexes.

28
4.4.3 Decision Tree Model
The chi-squared Automatic Interaction Detection (CHAID) algorithm was chosen for the
flood susceptibility modelling as each factor used in building a branch shows a strong
correlation with the dependent variable and each new branch is created based on the relative
importance relationship between the influencing factors and flood inventory while factors
signifying the same influence are consolidated to form a branch. Therefore, this makes
CHAID a very suitable algorithm for modelling natural hazards based on its multifaceted
splitting in an optimum manner[35].
The splitting and merging parameters range between 0 and 1 and several parameters were
used until the optimum parameters were set. After this process, the Chi-square statistic was
applied by creating a tree structure which is initiated by the root node and further branched
into the internal nodes, and afterwards the terminal nodes. The Chi-square establishes a
binary decision that splits-up classes from other classes thereby creating a top-down
structural tree till the terminal nodes are concluded. This creates a structure that relates the
level of variable’s influence on the tree’s structure where some features are classified while
others are rejected by the algorithm based on their relative importance.

4.5 Hybrid Modelling

The flood conditioning factors were all re-classified using the IOE weights (𝑊𝑗 ). The
database derived was then used as inputs in training the machine learning models ( 𝑅𝐹𝐼𝑂𝐸 ,
𝑆𝑉𝑀𝐼𝑂𝐸 and 𝐷𝑇𝐼𝑂𝐸 ). Ten-fold cross-validation was also performed on each of the hybrid
machine learning models and the flood probability indexes were generated. To attest to the
accuracy of the hybrid models, stand-alone ML models were also used in training the
database in which all the influencing factors were all in a continuous data format and
unclassified without the influence of the IOE weights deriving the flood probability indexes
for each of the stand-alone machine learning models.

4.6 Creation of Flood Susceptibility Maps

The flood susceptibility models were used in deriving the flood susceptibility maps by
splitting up flood probability index generated in continuous data format into pixels
representing the susceptibility classes. The pixels are assigned a distinctive susceptibility
index by calculating flood susceptibility indices for each pixel to ascertain the possibility of
flood occurrences in the study area[22]. Each pixel obtained represents a value between 0
and 1, where 0 denotes no potential of flood susceptibility and 1 denotes a high potential of
flood susceptibility.

29
The probability index was further classified using the quantile technique to produce the final
susceptibility maps. The quantile technique was adopted in classifying the flood
susceptibility index due to its suitability in grouping an equal number of pixels in the same
classes without tampering with the values[34]. Quantile technique is a popular and efficient
classification technique based on its reliable performance in hazard modelling[29].
Subsequently, the flood susceptibility maps were classified into five classes namely: low, very
low, moderate, high, and very high based on literature[22].

4.7 Results Validation and Model’s Performance Assessment

The performance assessment and prediction capability of the flood models are evaluated
using both the training and the testing datasets through the ROC curve and statistical
metrices (Step F). Both evaluation measures are described below.

4.7.1 Model Evaluation using the ROC Curve

The Receiver Operating Curve (ROC) curve is a popular, comprehensive evaluation tool
used in hazard modelling studies[20], [52]. The ROC curve is conventionally utilized to
assess the performance of the ML models and its efficiency relies on the ranking model’s
performance in an organized manner and attractive visualization[35], [48]. The statistical
indicator of the ROC curve is represented by the Area Under Curve (AUC). Through the AUC,
the accuracy and the performance of the ML models are quantified using various thresholds
of the probability.
The curve is created by plotting ‘sensitivity’ on the Y-axis against ‘1-specificity’ on the X-axis.
The AUC ranges between 0 and 1 where 1 indicates that the observed and simulation data
are in a perfect spatial agreement[1]. Therefore, the closer the value is to 1 determines the
efficiency and the precision of the model. Consequently, an AUC value of <0.6 depicts weak
accuracy, 0.6 – 0.7 indicates moderate accuracy while 0.7 – 0.8 depicts good accuracy while
>0.8 indicates an almost perfect accuracy[66]. To plot the AUC curve and derive the
statistical indices is dependent on the following parameters namely: True Positive (TP), True
Negative (TN), False Positive (FP), False Negative (FN), (P) and (N). The AUC is derived as
thus:

∑ 𝑻𝑷+ ∑ 𝑻𝑵
𝑨𝑼𝑪 = (21)
(𝑷+𝑵)

Where TP denotes the number of pixels correctly classified, TN denotes the number of pixels
correctly classified as non-flood pixels, P is the total number of flood pixels, and N represents
the total number of non-flood pixels.

30
Considering this study’s context, success rate and prediction rate were constructed. The
success rate represents a ROC curve plot type that describes how the flood probability index
segregates the flood locations across the susceptibility zones and highlights the model fitting
rate to the training dataset[4]. On the other hand, the prediction rate reveals the
performance of the models in predicting locations prone to flooding and indicates how
efficient the model is in predicting floods[3]. The Success Rate is constructed based on the
training dataset which does not describe the efficiency of the model and the testing dataset
was used in constructing the prediction rate through the comparison of the testing dataset
to the flood susceptibility maps.
4.7.2 Statistical Metrics
Statistical measures were implemented in this study to augment the AUC curve in having a
detailed statistical analysis of the model’s predictive capabilities and to check the statistical
significance of the models[3]. The set of statistical metrics considered in this study were
Sensitivity (Recall), Accuracy and Specificity, Positive Predictive Rate (Precision), and
Negative predictive Rate (NPV).
Sensitivity is a statistical index that measures the proportion of flood pixels correctly
classified as flood pixels. Specificity index indicates and measures the proportion of non-
flood pixels correctly classified as non-flood pixels. The accuracy index measures the rate of
difference between the flood and non-flood pixels[4]. Each of the statistical indicators is
defined as[52]:

𝑻𝑷
𝑺𝒆𝒏𝒔𝒊𝒕𝒊𝒗𝒊𝒕𝒚 = * 100 (22)
𝑻𝑷+𝑭𝑵
𝑻𝑵
𝑺𝒑𝒆𝒄𝒊𝒇𝒊𝒄𝒊𝒕𝒚 = * 100 (23)
𝑭𝑷+𝑻𝑵

𝑻𝑷+𝑻𝑵
𝑨𝒄𝒄𝒖𝒓𝒂𝒄𝒚 = * 100 (24)
𝑻𝑷 `+𝑭𝑷+𝑻𝑵+𝑭𝑵
𝑻𝑷
𝑷𝑷𝑽 = *100 (25)
𝑻𝑷+𝑭𝑷

𝑻𝑵
𝑵𝑷𝑽 = * 100 (26)
𝑻𝑵+𝑭𝑵

Where FP equals the number of pixels incorrectly classified as flood events, and FN equals
the number of pixels incorrectly classified as non-flood. Thus, the higher the TP and the
lower the FP indicates the efficiency of the model.

31
4.8 Software and Device Specifications

The data derivation and pre-processing were implemented using the software ArcGIS Pro
2.7.0 from ESRI and ERDAS IMAGINE 2020. Afterwards, the IOE modelling was developed
using ArcGIS Pro 2.7.0 and Microsoft Excel through which each coefficient values were
derived.

WEKA 3.8.4 was used in performing the feature selection process while the multi-
collinearity statistical analysis was derived through the SPSS statistics 26 software from
IBM. The machine learning model training and classification was implemented using WEKA
3.8.4 which contains the required packages for conducting machine learning analysis as it
includes the Decision Tree, Support Vector Machine and Random Forest classification
algorithms and other packages which were used in performing the complete model training
process. SPSS statistics 26 was also used in deriving the Area under Curve (AUC) and other
performance metrics in validating the model.

The hardware (computer) utilized has an installed 8GB RAM and a processor Intel(R) Core
i5-6300U CPU @2.40GHZ 2.50GHz.

32
5 RESULTS
This chapter introduces the outputs attained from the methodology implemented in chapter
4. It should be noted that only selected tables and figures where appropriate were presented
while other results are attached in Annexes.

5.1 Predictiveness of Flood Influencing Factors (Feature Engineering)

5.1.1 Feature Selection

The predictive capability of each flood influencing factors on flood occurrences is germane
to performing flood susceptibility. Therefore, the linear support vector machine algorithm
was employed in performing the feature selection process and the Average Merit (AM)
values were attained (Figure 6). The AM values describe the strength of each influencing
factor in predicting flood occurrence and the value ranges between 0 and 1. Consequently,
the distance from river obtained the highest AM (0.850) and followed by the population
which obtained the average merit of 0.642, Distance from Road (0.583) while NDVI had the
lowest average merit of 0.089. Based on the results attained, all the influencing factors
achieved values greater than zero. Therefore, all the factors were considered in the flood
susceptibility modelling as each factor tend to have a certain influence on flood occurrence.
Furthermore, to ensure an outright assessment of flood susceptibility within the region, it is
of high necessity to give full consideration for each of the influencing factors no matter how
low the influence might be.

Average Merit (AM)

NDVI
Altitude
SPI
TWI
Slope
Geology
Drainage Density
Curvature
Rainfall
LULC
Soil
NDBI
Distance From Road
Population
Distance from River

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Figure 6: Factor’s predictive ability (average merit) result

33
5.1.2 Multi-collinearity Analysis
Tolerance (TOL) and Variance Inflation Factor (VIF) was utilized in performing this process
using the threshold of TOL > 0.1 and VIF < 10 critical values based on existing literature
(Table 2). Geology has the lowest tolerance (0.115) and the highest VIF (8.714). However,
even as it seems to have almost the same influence as soil with the TOI (0.133) and VIF
(7.504), both influencing factors were considered as previous studies have proved the
influence of geology on the occurrence of floods. Furthermore, all the influencing factors
considered met the threshold laid down as all factors exceed the theoretical critical values
for any evidence of multi-collinearity and were all therefore utilized in the modelling
process.

Factor Tolerance VIF

Rainfall 0.500 2.001
Altitude 0.569 1.758
Curvature 0.736 1.359
Slope 0.458 2.183
TWI 0.430 2.326
SPI 0.685 1.460
Drainage Density 0.883 1.133
Distance from Road 0.439 2.278
LULC 0.443 2.258
Soil 0.133 7.504
Geology 0.115 8.174
Population 0.495 2.021
Distance from River 0.350 2.857
NDVI 0.359 2.782
NDBI 0.413 2.419
Table 2 : Multi-collinearity Analysis

5.2 Flood Modelling Algorithms

5.2.1 Index of Entropy Flood Modelling

IOE modelling was utilized in this study as bivariate and multivariate statistical modelling.
This is to ensure the weights attributed to each class of an influencing factor and the overall
weight of each factor on flood occurrence was attained (Table 3). To perform IOE modelling
required calculating the FR values of the classes of each factor. However, it should be

34
mentioned that ratios > 1 signifies a high probability of flood occurrence while ratios < 1
signifies a low probability of flood occurrence. Population class between 21970 – 73351
attained the highest value of FR attaining 4.01. This was followed by grassland class of the
LULC factor which attained a value of 2.56. It should be mentioned that 22 classes attained
FR value equal to zero. The FR values were used to attain the probability density values (𝑃𝑖𝑗 )
which measures the probability of flood occurrence representing the weight of influence in
each class of influencing factors. Consequently, distance from river class between 0 –
2986.4m attained the highest value of 0.98, followed by distance from road class 0 - 440.3m
attaining a value of 0.85. This was then followed by the subrecent alluvium and coastal plain
sands class of the geology factor which attained a value of 0.43. It should be noted that as in
the case of FR values which attained the values of 0 also resulted in the probability density
of the classes’ values equaling to 0.
Subsequently, the weights of the flood influencing factors which ranges from 0 to 1 were
derived after completing the modelling (Table 3). Distance from river achieved the highest
weight with a value of 0.52 followed by the distance from road with a value of 0.34, Altitude
(0.26), SPI (0.20), NDBI (0.19), LULC (0.19), Geology (0.18), Population (0.17), Soil (0.17),
Slope (0.12), NDVI (0.12), Drainage Density (0.07), Rainfall (0.071), TWI (0.04), Curvature
(0.002). Thereafter, the susceptibility index 𝐹𝑆𝐼𝐼𝑜𝐸 was derived through the multiplication
of the influencing factor’s weight (𝑊𝑗 ) with the IOE coefficients earmarked to each class.
Thus, in deriving the flood susceptibility index, the final equation is as follows:
𝐹𝑆𝐼𝐼𝑜𝐸 = 0.51 ∗ [𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒 𝑓𝑟𝑜𝑚 𝑅𝑜𝑎𝑑 ] + 0.34 ∗ [𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒 𝑓𝑟𝑜𝑚 𝑅𝑖𝑣𝑒𝑟] + 0.26 ∗ [𝐴𝑙𝑡𝑖𝑡𝑢𝑑𝑒]
+ 0.20 ∗ [𝑆𝑃𝐼] + 0.19 ∗ [𝑁𝐷𝐵𝐼 ] + 0.19 ∗ [𝐿𝑈𝐿𝐶 ] + 0.18 ∗ [𝐺𝑒𝑜𝑙𝑜𝑔𝑦] + 0.17
∗ [𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛] + 0.17 ∗ [𝑆𝑜𝑖𝑙 ] + 0.12 ∗ [𝑁𝐷𝑉𝐼 ] + 0.07 ∗ [𝐷𝑟𝑎𝑖𝑛𝑎𝑔𝑒 𝐷𝑒𝑛𝑠𝑖𝑡𝑦]
+ 0.07 ∗ [𝑅𝑎𝑖𝑛𝑓𝑎𝑙𝑙 ] + 0.04 ∗ [𝑇𝑊𝐼 ] + 0.002 ∗ [𝐶𝑢𝑟𝑣𝑎𝑡𝑢𝑒]

35
Factor Class Flood Class Pixels 𝑭𝑹𝒊𝒋 (𝑷𝒊𝒋 ) 𝑾𝒋
Pixels

Altitude (m) 0-7 22 1621454 0.593882 0.158748 0.258

8 - 16 54 1652308 1.43049 0.382377
20 - 36 19 538174 1.545301 0.413067
37 - 57 1 255416 0.17137 0.045808
58 - 109 0 134622 0 0
Rainfall (mm) 0 - 9.6 7 948452 0.323004 0.089865 0.071
9.7 - 10.7 48 1671607 1.2567 0.349637
10.8 - 12.0 34 1167270 1.274771 0.354664
12.1 - 13.7 7 414088 0.739827 0.205833
Slope (Degree) 0 - 1.96 37 1921380 0.839859 0.21496 0.124
1.97 - 5.15 42 1281650 1.429216 0.365805
5.16 - 9.78 12 645367 0.810947 0.20756
9.79 - 16.71 5 263677 0.827019 0.211674
16.72 - 45.32 0 74793 0 0
Curvature Convex 11 482290 0.998313 0.320926 0.002
Flat 73 3255206 0.981582 0.315547
Concave 12 464478 1.130832 0.363527
Soil Aquic Paleudult 0 21150 0 0 0.169
Orthic Tropudalf 0 19449 0 0
Oxic Rhodudalf 15 976453 0.671283 0.200306
Oxic Tropudalf 0 22323 0 0
Oxic Tropudalt 16 629038 1.111499 0.331663
Rhodic Paleudult 0 15477 0 0
Typic Tropaquent 0 16716 0 0
Typic Tropudult 2 535389 0.16324 0.04871
Typic Ustifluvent 63 1959055 1.40527 0.419322
SPI 0 - 67626 96 4185641 1.000293 1 0.2
676267 - 287411 0 743 0 0
287411 - 625452 0 315 0 0
625542 - 1115832 0 117 0 0
1115833 - 2155584 0 51 0 0
Drainage Density 0 - 3896.93 52 2650331 0.858909 0.237492 0.073
3897 - 11690.78 34 944863 1.575266 0.435568
11691 - 22948.57 9 493284 0.79871 0.220847
22949 - 55206.47 1 114092 0.383697 0.106094
Distance from Road 0 - 440.3 90 2138540 1.834335 0.847763 0.337
(m)
441.4 - 9373.4 3 1242905 0.105205 0.048622
9374 - 19598.7 3 583243 0.224195 0.103615
19600 - 36214.5 0 219630 0 0
Land use Land Cover Irrigated Croplands 0 1040 0 0 0.185
Residential Areas 17 728714 1.020756 0.09349
Rural Areas 0 11631 0 0

36
 Broadleaved Evergereen Forest 15 871338 0.753242 0.068989
Semi-deciduous Woodland 8 403307 0.867929 0.079493
Broadleaved Deciduous Woodland 0 47792 0 0
Grassland 1 17097 2.559236 0.234398
Shrubland 6 239933 1.094187 0.100216
Hebaceous Vegetation 3 229139 0.572865 0.052468
Broadleaved Forest regularly flooded 0 44018 0 0
Broadleaved Forest 3 101851 1.288802 0.11804
Woody Vegetation 0 1557 0 0
Central Business District 38 665013 2.500251 0.228996
Water Bodies 5 838074 0.261046 0.023909
Geology Coastal Plain Sands 15 997603 0.657052 0.19921 0.175
Coastal Plain Sands (Alfisols) 2 537657 0.162551 0.049284
Coastal Plain Sands (Ultisols) 16 651361 1.073406 0.325444
Deltaic Basin and Tidal Flats 0 15477 0 0
Ewekoro Formation (Upper Coal Measure) 0 16077 0 0
Recent Alluvium 0 17181 0 0
Sandstone (Abeokuta Formation) 0 639 0 0
Subrecent Alluvium and Coastal Plain Sands 63 1959055 1.40527 0.426062
Population 0 - 486 52 3334648 0.682253 0.061751 0.173
487 - 1053 13 336658 1.689454 0.152912
1054 - 6010 15 287226 2.284859 0.206802
6011 - 21969 9 165197 2.383594 0.215738
21970 - 73351 7 76405 4.008373 0.362797
Distance from 0 - 2986.4 95 2469057 1.665271 0.97788 0.513
River(m)
3000.2 - 7700.3 1 1148992 0.037668 0.02212
7701.4 - 40071.7 0 536880 0 0
NDVI 0 - 0.04160 1 298954 0.146425 0.031126 0.122
0.04200 - 0.25140 38 873942 1.903355 0.4046
0.25150 - 0.43219 21 801044 1.147576 0.243943
0.43220 - 0.61302 13 656931 0.866248 0.18414
0.61400 - 0.84449 23 1571457 0.640684 0.136192
NDBI -1 - -0.4814 1 487770 0.089744 0.018046 0.192
-0.4813 - -0.2320 10 1251171 0.349866 0.070354
-0.2319 - -0.0884 15 724405 0.906418 0.182269
-0.0883 - 0.0552 32 682483 2.05247 0.412726
-0.0553 - 1.0000 38 1056499 1.574466 0.316605
TWI 0 - 6.8 31 1348395 1.00275 0.210479 0.042
6.9 - 8.9 29 1154504 1.095597 0.229968
9.0 - 11.6 21 690239 1.326991 0.278538
11.7 - 14.3 11 480114 0.999302 0.209755
14.4 - 24.2 4 513908 0.339487 0.071259
Table 3: Frequency ratio and Index of Entropy coefficients values distribution within flood
influencing factors classes.

37
5.2.2 Support Vector Machine Flood Modelling
To perform the SVM flood modelling requires the tuning of parameters gamma (𝛾) and
penalty (∁) to train the SVM stand-alone and SVM-IOE hybrid model. This was achieved
through the ten-fold cross-validation. In consideration of the SVM stand-alone model,
𝛾(0.11) and ∁(1.4) were attained and used for optimizing the database and deriving the flood
susceptibility index (Figure 7) while for the SVM-IOE model, 𝛾(0.19) and ∁(1.6) were
attained and used in generating the flood susceptibility index. The stand-alone SVM flood
susceptibility map ranges from 0.152 to 0.836 and using the quantile technique, the map
was classified into five classes of very low, low, moderate, high, and very high (Figure 8).
The lowest class (0.152 – 0.369) occupies 31.73% of the total study area which indicates the
low flood susceptibility areas of the watershed and the moderate class (0.495 – 0.635)
occupies 20.47% of the watershed while the very high susceptibility class (0.680 – 0.836)
which indicates areas highly prone to flooding occupies 22.81% of the watershed. On the
other hand, SVM-IOE susceptibility map ranges from 0.196 to 1.374. SVM-IOE lowest
susceptible class (0.196 – 0.450) occupies 21.56% of the total study area while the moderate
flood susceptible class (0.810 – 1.032) occupies 18.86% of the watershed while the highest
flood susceptible class (1.110 – 1.374) occupies 22.41% of the study area.

38
Figure 7: Flood probability index derived from: (a) RF-IOE, (b) DT-IOE, (c) Stand-alone DT, (d)
Stand-alone RF, (e) SVM-IOE, (f) Stand-alone IOE, (g) Stand-alone SVM.

5.2.3 Random Forest Flood Modelling

The ten-fold cross-validation technique was also used in tuning the 𝑚𝑡𝑟𝑦 and 𝑛𝑡𝑟𝑒𝑒
parameters to train the random forest model. Thereafter, optimal parameters were set to be
𝑛𝑡𝑟𝑒𝑒 = 2000 trees and 𝑚𝑡𝑟𝑦 = 10. The out of bag error procedure was also implemented,
which is based on the uniformity of the nodes and leaves within the RF model[58]. This is to
ensure accuracy within the model as the model’s accuracy decreases based on the exclusion
of important variables. Based on these metrics, distance from road, distance from river,
population, LULC and soil demonstrates high importance in the flood modelling.
The RF susceptibility map (Figure 8) derived ranges from 0.130 to 0.675. The lowest class
(0.130 – 0.327) which signifies areas less susceptible to flood occupies 36.15% of the study
area while about 15.60% of the watershed signifies areas moderately susceptible to flood
within the watershed. The highest class (0.547 – 0.675) signifies areas highly susceptible
and occupies 23.49% of the total study area. On the other hand, the optimal parameters
achieved through the cross-validation for the RF-IOE was 𝑛𝑡𝑟𝑒𝑒 = 2500 𝑡𝑟𝑒𝑒𝑠 and 𝑚𝑡𝑟𝑦 =
11 and the out of the bag procedure was also implemented for modelling the algorithm.
Thus, the final RF-IOE susceptibility index ranges from 0.055 to 0.615. The lowest
susceptible class (0.055 – 0.171) occupies 20.71% of the study area while the moderate

39
susceptible class (0.275 – 0.483) occupies 37.14% of the watershed having the highest
percentage of the study area. The highest susceptible class (0.518 – 0.615) which signifies
areas highly susceptible to flood occupies 17.32% of the study area.

Figure 8: Flood susceptibility maps derived from (a) RF-IOE, (b) Stand-alone DT, (c) Stand-alone
RF, (d) DT-IOE, (e) Stand-alone SVM, (f) SVM-IOE, (g) Stand-alone (IOE).

40
5.2.4 Decision Tree Flood Modelling
The splitting and merging parameters were set with the values of 0.8 and 0.001 for the
stand-alone DT model which were arrived at after continuous trial and error to achieve the
best accuracy. DT algorithm reduces parameters in performing its modelling as the most
important variables are selected and used in generating the model. Therefore, in the stand-
alone DT model, TWI, geology, altitude, NDVI and slope were rejected by the algorithm
while the other 10 influencing factors were utilized for the modelling. In terms of the DT
top-down structure, the influencing factors higher on the tree structure signifies a higher
influence on flood occurrence. Therefore, distance from road, distance from river, drainage
density, LULC and rainfall were first selected for splitting by the algorithm.
Thus, the DT tree generated contains 10 variables, 458 nodes and 108 leaves making it
impossible to present as a tree in the document and each leaf describes a certain degree of
flood potentiality. The final susceptibility index generated for the DT model ranges from
0.114 to 0.649. The lowest susceptible class (0.114 – 0.313) occupies 37.20% of the total
study area while the moderate class (0.412 – 0.489) occupies 14.39% of the watershed. The
highest susceptible class (0.529 – 0.649) indicating areas highly prone to floods occupies
22.43% of the study area. On the other hand, the splitting and merging parameters for the
DT-IOE model was set to 0.7 and 0.005, respectively. NDVI, TWI, slope, and altitude were
rejected by the hybrid model.
Therefore, 11 influencing factors were utilized for the modelling and the final tree contains
648 nodes and 189 leaves. The flood susceptibility index derived ranges from 0.057 to 0.600.
The lowest class of susceptibility (0.057 – 0.168) occupies 18.50% of the study area while
the third class (0.265 – 0.467) which signifies areas moderately susceptible to flood occupies
37.20% of the study area having the highest percentage of the study area. The highest
susceptible class (0.503 -0.600) which indicates high flood susceptibility occupies 17.86% of
the study area.

5.3 Accuracy Assessment and Validation of Flood Models

The seven flood susceptibility maps were evaluated by AUC technique and other statistical
metrics namely Sensitivity, Specificity, Accuracy, Positive Predictive Value (PPV), and
Negative Predictive Value (NPV). This was to test the accuracy of the models and the higher
the AUC values the better the model in terms of success rate and prediction rate. The DT
and RF stand-alone model has the highest success rate of 0.900 while the IOE model and
the SVM model both had the success rate of 0.899 (Figure 9).

41
 1.0

0.9

0.8

0.7
Sensitivity

0.6
DT (AUC = 0.900)
0.5 DT-IOE (AUC = 0.890)
IOE (AUC = 0.899)
0.4
RF (AUC = 0.900)
0.3 RF-IOE (AUC = 0.895)
0.2 SVM (AUC = 0.899)
SVM-IOE (AUC = 0.894)
0.1

0.0
0.0 0.2 0.4 0.6 0.8 1.0
1 - Specificity

Figure 9: Area under Curve (AUC) showing the Success Rate

1.0

0.9

0.8

0.7

0.6
Sensitivity

DT (AUC = 0.770)
0.5 DT-IOE (AUC = 0.871)

0.4 IOE (AUC = 0.871)

RF (AUC = 0.801)
0.3
RF-IOE (AUC = 0.867)
0.2 SVM (AUC = 0.861)
SVM - IOE (AUC = 0.864)
0.1

0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
1 - Specificity

Figure 10: Area under Curve (AUC) showing the Prediction Rate

This was then followed by the RF-IOE (AUC = 0.895), SVM-IOE (AUC = 0.894), and DT-
IOE (AUC = 0.890) had the lowest performance.

42
In consideration of the hybrid models, DT-IOE model produced the highest prediction rate
of 0.871 followed by the RF-IOE model with a prediction rate of 0.867 while SVM-IOE had
a prediction rate of 0.864 (Figure 10). On the other hand, the DT stand-alone model
produced the lowest prediction rate of 0.770, followed by the RF stand-alone model with a
prediction rate of 0.801 while the stand-alone SVM model had the highest prediction of
0.861. Percentagewise, using the hybrid technique improved the prediction rate as the DT-
IOE is 10% higher than the stand-alone model, and the RF-IOE model is 7% higher than RF
stand-alone model while the SVM-IOE slightly performed than the SVM stand-alone model
in predicting areas susceptible to flood.

Classifier Sensitivity Specificity Accuracy PPV NPV

DT-IOE 83.2 77.6 80.3 78.2 82.6

DT 67.4 69.2 68.3 70.7 65.9
IOE 74.4 72.1 73.2 70.7 75.6
RF 67.4 69.2 68.3 70.7 65.9
RF-IOE 79.4 80.6 80.3 81.1 78.9
SVM 78.0 69.8 73.8 71.1 76.9
SVM-IOE 82.4 84.1 78.3 77.0 80.0
Table 4: Performance metrics of Classifiers

The authenticity and reliability of the results were further checked with other statistical
metrics (Table 4). Overall, the DT-IOE model had the highest performance assessing the
accuracy of the classifiers based on the validation dataset with a specificity (83.2%), and this
was followed by SVM-IOE with the sensitivity of 82.4%, RF-IOE (79.4%), SVM (78%), IOE
(74.4%), RF (67.4%) and DT (67.4%). Based on the accuracy, the RF-IOE and DT-IOE had
the overall highest performance with an accuracy of 80.3%, followed by the SVM-IOE
(78.3%), SVM (73.8%), IOE (73.2%), RF (68.3%) and DT (68.3%).

43
6 DISCUSSION
The major focus of this research is to develop and utilize machine learning-based flood
susceptibility models in deriving flood susceptible maps considering a diverse range of
factors relative to the study area and to identify the impact of the factors on flood occurrence
in the study area. This was performed taking account of the human-induced factors and
other natural-caused factors acknowledged in previous studies and proven to have a certain
influence on flood occurrence in the study area. Moreover, feature engineering was
performed to investigate the predictive ability, significance, and interrelationship among the
influencing factors before the main modelling and all the factors had a certain influence and
were therefore utilized for the modelling. Furthermore, feature engineering was performed
as there are no existing works related to flood susceptibility in the area which necessitates
the identification of the factors prevalent in the region and their significance.
Also, it is worthy to note the importance of remote sensing which facilitated the derivation
of most of the influencing factors, verification of the past flood events and generally plays a
huge role in assessing potential areas susceptible to natural hazards. Subsequently, in
achieving accurate results and high prediction based on the novelty of FSM in the region,
novel hybrid models of efficient ML algorithms namely DT, SVM and RF integrated with
IOE, and each model as a stand-alone model was implemented in deriving flood
susceptibility maps for the study area (Figure 8). The ML models were implemented as an
effectual flood risk assessment is essential due to the increasing incessant flooding in the
region, the complex nature of the flood influencing variables resulting in its occurrence and
its long devastating effects on the region.
Based on the modelling’s novelty in the area, the study utilized all the fifteen influencing
variables considering a certain degree of influence each variable has on the spatial
distribution of flood occurrence in the region. The utilization of the IOE statistical model
provided the platform in assessing the influence of each class of each influencing factor and
the overall weight of each factor on flood occurrence through bivariate and multivariate
statistical analysis (Table 3). The acquired weights were then used in reclassifying the
factors. IOE was further utilized based on its superiority to other BSA/MSA models as it does
not presume a linear model and make no assumptions with regards to the distribution of
variables[53].
According to Costache et al. (2020) and Tehrany et al.(2019), adopting this hybridization
approach of machine learning and statistical techniques approach increases the accuracy of
the classification algorithms significantly which was further proven by the outcome of this

44
study[35]. Furthermore, the influence of each class of each factor’s influence on the
occurrence of floods is identified and given uttermost consideration.
Regarding the first research question, distance from river network has the highest
influence on flood occurrence in the area. This confirms that areas highly susceptible to
floods are very close to the river network, Furthermore, the lowest class (0 – 2986.4m) of
the distance from river has the highest influence (0.98) which indicates areas within 3000m
to the water bodies are highly susceptible to flooding. This is followed by distance from road
which had the second overall influence on flood occurrence within the region. This was more
significant in the factors first class (0 – 440.3m) which signifies the confluence of roads
within a 440m radius highly influences the occurrence of floods.
Furthermore, altitude does have a significant influence on the occurrence of floods in the
area having the highest third overall influence. Water flows from high altitudes to lower ones
and therefore upsurges the occurrence of floods in the lower areas. The geologic composition
of the region influences flooding as the subrecent alluvium and coastal plain lands which
comprises 85% of the Lagos landform is composed of fossils, sedimentary rocks which
reduces water permeability, allowing water inundation, and exacerbates flooding. This
reveals the impact of geology on flooding within the region.
Consequently, the overall results reveal that the factors majorly distance from river, distance
from road, NDBI, altitude, soil, LULC, rainfall and SPI play a critical and huge role in the
spatial distribution of flood occurrences in the study area. On the other hand, slope has a
low influence on flooding in the study area as the slope reflects a downward trend thereby
having a huge percentage of flat curvature which increases runoff speed and reduced water
percolation into the soil. However, this affected the SPI which describes the erosion capacity
to be greatly increased thereby having a very moderate impact on flood occurrence. Also, the
NDVI which describes the vegetation density has a low impact on flood occurrence in the
region due to the low vegetation density in the urban and residential areas. Curvature had
the lowest value signifying the least influence on the occurrence of floods.
Based on previous studies, curvature often generates low values which have, in turn, led to
various contrasting assertions on the role of curvature in FSM[20]. Figure 9. 1(Annex)
details the relative distribution of flood pixels within the classes of each flood influencing
factor which describes the flood potential observed in each class. Based on the flood
susceptibility maps derived, areas highly prone to flooding occupied approximately 21% of
the study area, which are around the lagoon, also in the plain region near river networks and
majorly in the residential areas and the central business district of the city. This reveals the
impact of urbanization on flood occurrences in the study area and confirms the influence of
45
NDBI and LULC on the occurrence of floods in the region. This can be attributed to Bahram
Choubin’s study that revealed 181,000 people inhabiting residential and urban areas around
rivers in the world are consistently affected by flooding[45].
Regarding the second research question based on accuracy assessment and predictive
performance of the models, all the models performed relatively well with slight differences
in terms of the success rate which describes the model fitting rate to the training dataset.
The DT and RF models had the highest performance followed by SVM and IOE. however,
hybrid models (DT-IOE, SVM-IOE, and RF-IOE) attained the lowest performance which
could be attributed to quantile classification technique implemented in classifying each of
the influencing factors in performing the IOE modelling though quite insignificant based on
the success rates achieved.
On the other hand, considering the prediction rate, the DT-IOE had the highest overall
prediction performance over the six other models outperforming the SVM-IOE and the RF-
IOE models. This signifies that the DT-IOE susceptibility map had the highest accuracy in
predicting flood locations prone to flooding. Also, this proves the rejection or addition of
additional influencing factors does not necessarily increase the accuracy of the models as
factors with similar influence may be rejected by the algorithm and does not suggest the
factor’s significance invalid[35].
Tehrany et al. (2019) also attained similar findings where DT model performed slightly
more than SVM model. This relates to the SVM’s ability in handling multi-linear data with
high precision and low error rates. Fotovatikhah et al. (2018) explored over a hundred
articles on floods and attested to this fact[43]. However, SVM algorithm lacks the ability in
evaluating the significance of variables utilized. Consequently, based on literature, past
flooding studies often utilizes SVM with various statistical models and integration with other
machine learning models in addressing the importance of variables utilized[4], [35].
However, DT stand-alone model attained the lowest prediction performance.
Furthermore, the advantage of utilizing hybrid machine learning models is further proven
as the DT-IOE obtained a higher accuracy than the DT stand-alone model. Additionally, it
should be mentioned that the utilization of hybrid models quickened the modelling process
as the processing time remarkably reduced due to the pre-processing of the flood influencing
factors as one of the major drawbacks of the ML models utilized is the required time for
analysis.
On the other hand, to ensure statistical significance and overall efficiency, statistical metrics
namely Sensitivity, Specificity, Accuracy, NPV and PPV were utilized in also assessing the
performance of the models. The DT-IOE performed best slightly than the RF-IOE and the
46
SVM-IOE in terms of specificity of over 80%. The RF had the lowest performance of
approximately 70% sensitivity which suggest a good indiscriminatory performance of all the
models utilized for the study.
Also, each model produced partly varied different susceptibility patterns with relation to the
susceptibility maps (Figure 8) generated even with similar statistical performances. This
could be relatively attributed to the selection procedure and the utilization technique of the
variables implemented by each ML algorithm. The RF permutes each variable randomly and
DT (CHAID) selects the variables hierarchically based on their relative importance on flood
occurrence while SVM incorporates all the variables and sets up an optimal plane to
distinguish the classes of the variables based on flood and non-flood origin. Thus, the
variation of susceptibility patterns based on selection and combinations of set of factors by
each ML algorithm.
However, it should be mentioned that selecting the best model for FSM is quite challenging,
even though hybridization of models is powerful, variations exist depending on the region
which could induce uncertainty in spatial prediction. Therefore, changing input data based
on future conditions could alter the model’s accuracy[19].

6.1 Limitations and Recommendations

There were some remarkable limitations encountered in this study. Foremostly, there was
data paucity (spatial and temporal) in terms of flood influencing variables, flood inventory
data, and resolution of imageries acquired for the study. It is ascertained that the prediction
abilities of the factors will increase if the factors are derived from higher resolution
imageries. Also, regarding the temporal dimension of flood occurrence in the region, which
is fully based on rainfall that initiates flooding, there is a limitation based on the availability
of rainfall data from rainfall stations which could reveal the unceasing influence of rainfall
on flooding in the region. However, rainfall’s influence was quite significant in the study
based on the data utilized which relates that more reliable data will further reveal the
increasing influence of rainfall in the study area over time.
Furthermore, acquiring more accurate and detailed inventory data is very fundamental to
the entire process and would help in enriching and optimizing the model’s parameters for
the study area and more accurate models can be attained. Even though multiple iterations
of random points for the inventory map were performed, there was no increase in the
precision of models which addresses the stability of the results attained.

47
However, it is needed to acquire more inventory points where one inventory dataset (map)
can be used for training and another inventory dataset for testing and multiple interactions
can be performed with their outcomes compared. This will help to attain a more significant
impact on the temporal dimension of flood occurrence in the region. In essence, there is a
huge absence of a comprehensive spatial data infrastructure (SDI) in the region which is
essential in bringing new insights into the flood susceptibility domain.
Also, based on feasibility studies that have been carried out in the region, wastes generation
and disposal is a key factor in the occurrence of floods due to the blockage of water channels
and decrease in the level of water percolation[12]. Therefore, more investigation is needed
towards the exploration of this factor by acquiring tons of waste generated per block radius
in each flood susceptibility zone.
In conclusion, performing accurate flood susceptibility mapping requires updated and
accurate flood historical data, high resolution derived flood influencing variables and a
powerful modelling algorithm to achieve highly sustainable results.

48
7 CONCLUSION
Flood Susceptibility Modelling (FSM) is one of the most popular research areas in natural
hazard studies. It is a significant domain where accuracy and time are essential in mitigating
and preventing flood occurrences. This study was implemented to investigate drivers of
flooding and identify areas prone to flooding in the West Africa region using Lagos as a case
study. This was achieved by implementing a novel hybrid and stand-alone machine learning
algorithms specifically DT-IOE, SVM-IOE, RF-IOE, DT, SVM and RF to train the geospatial
database composed of 15 influencing factors and 139 flood locations and 139 non-flood
locations. The hybrid models were created to enhance the accuracy of the stand-alone
models.
Thereafter, the models were then validated using the AUC and statistical metrics to check
the statistical significance and the overall efficiency of the model. Based on the AUC through
the evaluation of the success rate and the prediction rate, the DT-IOE (AUC = 0.899)
achieved better goodness of fit to the training dataset and the highest prediction accuracy
(AUC = 0.871). Percentagewise, the DT-IOE was approximately 10% higher than the DT
stand-alone model and the RF-IOE was 7% higher than RF stand-alone model which is a
significant improvement based on the model’s prediction accuracy. On the other hand,
checking for the statistical significance of the models and overall efficiency in terms of
Accuracy, Sensitivity and Specificity, DT-IOE had the best performance based on the
validation dataset which concludes DT-IOE as the most appropriate ML algorithm when
natural-caused and human-induced factors are concerned with regards to the study area.
The results revealed that human-induced factors play a huge role in the occurrence of floods
such as distance from road, NDBI and population density while natural-caused factors such
as distance from river, LULC, geology proved to be very significant drivers of flood
occurrences in the region. Also, the performance of feature selection process was
implemented to identify the most significant factors before performing the main modelling.
As a novel-based study in the region, the susceptibility maps generated would assist the
urban planners to prevent the increasing urbanization in the identified susceptible regions
thereby mitigating flood impact. Also, more inclined floodplains management approaches
and refined policies can be developed. Over time, it is realized that LULC, SPI and rainfall
as spatio-temporal factors influences the occurrence of flood and significant attention
should be given to the factors as a change in LULC over time determines a significant impact
of SPI and rainfall in the region.

49
In summary, the contribution of this research is emphasized as follows:
1. The proposed integration of IOE with DT, SVM and RF is a powerful modelling tool
in the classification and identification of flood locations prone to flood risk.
2. The adoption of feature engineering technique is a considerable approach before the
performance of FSM to identify the significance of the flood drivers as these factor’s
influence varies with time as to the adoption of user-defined factors from previous
research.
3. Human-Induced factors should be given full consideration in any region as
significant drivers of flood occurrences.
Finally, it should be mentioned that adopting the utilization of machine learning and
geospatial technology is very efficient in performing FSM based on time, costs, and
accuracy without expert judgement in the modelling. Also, the study gave insights into
the carrying out of FSM within West Africa as the results attained is relevant to the
national and local governments of flood-prone countries within the region which proves
FSM can be carried out successfully in the region.

50
8 BIBLIOGRAPHIC REFERENCES

[1] Y. Wang, Z. Fang, H. Hong, and L. Peng, “Flood susceptibility mapping using
convolutional neural network frameworks,” J. Hydrol., vol. 582, no. March, p.
124482, 2020, doi: 10.1016/j.jhydrol.2019.124482.
[2] R. Mind’je et al., “Flood susceptibility modeling and hazard perception in Rwanda,”
Int. J. Disaster Risk Reduct., vol. 38, no. April 2018, p. 101211, 2019, doi:
10.1016/j.ijdrr.2019.101211.
[3] W. Chen et al., “Modeling flood susceptibility using data-driven approaches of naïve
Bayes tree, alternating decision tree, and random forest methods,” Sci. Total
Environ., vol. 701, 2020, doi: 10.1016/j.scitotenv.2019.134979.
[4] R. Costache et al., “Spatial predicting of flood potential areas using novel
hybridizations of fuzzy decision-making, bivariate statistics, and machine learning,”
J. Hydrol., vol. 585, no. December 2019, p. 124808, 2020, doi:
10.1016/j.jhydrol.2020.124808.
[5] I. E. Olorunfemi, A. A. Komolafe, J. T. Fasinmirin, A. A. Olufayo, and S. O. Akande,
“A GIS-based assessment of the potential soil erosion and flood hazard zones in
Ekiti State, Southwestern Nigeria using integrated RUSLE and HAND models,”
Catena, vol. 194, no. January, p. 104725, 2020, doi: 10.1016/j.catena.2020.104725.
[6] P. T. Padi, G. Di Baldassarre, and A. Castellarin, “Floodplain management in Africa:
Large scale analysis of flood data,” Phys. Chem. Earth, vol. 36, no. 7–8, pp. 292–
298, 2011, doi: 10.1016/j.pce.2011.02.002.
[7] I. Ajibade, G. McBean, and R. Bezner-Kerr, “Urban flooding in Lagos, Nigeria:
Patterns of vulnerability and resilience among women,” Glob. Environ. Chang., vol.
23, no. 6, pp. 1714–1725, 2013, doi: 10.1016/j.gloenvcha.2013.08.009.
[8] J. Ntajal, B. L. Lamptey, I. B. Mahamadou, and B. K. Nyarko, “Flood disaster risk
mapping in the Lower Mono River Basin in Togo, West Africa,” Int. J. Disaster Risk
Reduct., vol. 23, no. October 2016, pp. 93–103, 2017, doi:
10.1016/j.ijdrr.2017.03.015.
[9] I. Douglas, “Flooding in African cities, scales of causes, teleconnections, risks,
vulnerability and impacts,” Int. J. Disaster Risk Reduct., vol. 26, no. September, pp.
34–42, 2017, doi: 10.1016/j.ijdrr.2017.09.024.
[10] C. C. Olanrewaju, M. Chitakira, O. A. Olanrewaju, and E. Louw, “Impacts of flood
disasters in Nigeria: A critical evaluation of health implications and management,”

51
 Jamba J. Disaster Risk Stud., vol. 11, no. 1, pp. 1–9, 2019, doi:
10.4102/jamba.v11i1.557.
[11] E. I. D. Database, “EM-DAT.” https://www.emdat.be/ (accessed Oct. 13, 2020).
[12] A. O. Israel, “Nature, the built environment and perennial flooding in Lagos,
Nigeria: The 2012 flood as a case study,” Urban Clim., vol. 21, pp. 218–231, 2017,
doi: 10.1016/j.uclim.2017.06.009.
[13] U. C. Nkwunonwo, M. Whitworth, and B. Baily, “A review of the current status of
flood modelling for urban flood risk management in the developing countries,” Sci.
African, vol. 7, p. e00269, 2020, doi: 10.1016/j.sciaf.2020.e00269.
[14] K. Chapi et al., “A novel hybrid artificial intelligence approach for flood
susceptibility assessment,” Environ. Model. Softw., vol. 95, pp. 229–245, 2017, doi:
10.1016/j.envsoft.2017.06.012.
[15] N. U. Whitworth Malcolm, “Flooding and Flood Risk Reduction in Nigeria: Cardinal
Gaps,” J. Geogr. Nat. Disasters, vol. 05, no. 01, pp. 1–12, 2015, doi: 10.4172/2167-
0587.1000136.
[16] S. V. Razavi Termeh, A. Kornejady, H. R. Pourghasemi, and S. Keesstra, “Flood
susceptibility mapping using novel ensembles of adaptive neuro fuzzy inference
system and metaheuristic algorithms,” Sci. Total Environ., vol. 615, pp. 438–451,
2018, doi: 10.1016/j.scitotenv.2017.09.262.
[17] C. Ugonna, “A Review of Flooding and Flood Risk Reduction in Nigeria,” Glob. J.
Human-Social Sci., vol. 16, no. 2, 2016.
[18] S. Janizadeh et al., “Prediction success of machine learning methods for flash flood
susceptibility mapping in the Tafresh watershed, Iran,” Sustain., vol. 11, no. 19,
2019, doi: 10.3390/su11195426.
[19] H. Shafizadeh-Moghadam, R. Valavi, H. Shahabi, K. Chapi, and A. Shirzadi, “Novel
forecasting approaches using combination of machine learning and statistical
models for flood susceptibility mapping,” J. Environ. Manage., vol. 217, pp. 1–11,
2018, doi: 10.1016/j.jenvman.2018.03.089.
[20] H. Hong, P. Tsangaratos, I. Ilia, J. Liu, A. X. Zhu, and W. Chen, “Application of fuzzy
weight of evidence and data mining techniques in construction of flood
susceptibility map of Poyang County, China,” Sci. Total Environ., vol. 625, pp. 575–
588, 2018, doi: 10.1016/j.scitotenv.2017.12.256.
[21] M. S. Tehrany, B. Pradhan, and M. N. Jebur, “Spatial prediction of flood susceptible
areas using rule based decision tree (DT) and a novel ensemble bivariate and
multivariate statistical models in GIS,” J. Hydrol., vol. 504, pp. 69–79, 2013, doi:
52
 10.1016/j.jhydrol.2013.09.034.
[22] M. S. Tehrany, L. Kumar, and F. Shabani, “A novel GIS-based ensemble technique
for flood susceptibility mapping using evidential belief function and support vector
machine: Brisbane, Australia,” PeerJ, vol. 2019, no. 10, 2019, doi:
10.7717/peerj.7653.
[23] P. Times, “What Lagos govt, residents must do to check flooding in Lekki,” 2017.
https://www.premiumtimesng.com/regional/ssouth-west/241903-lagos-govt-
residents-must-check-flooding-lekki-nimet.html (accessed Jan. 12, 2021).
[24] TheGuardian, “Lagos and the coming rains,” 2020.
https://guardian.ng/opinion/lagos-and-the-coming-rains/ (accessed Jan. 12,
2021).
[25] S. Samanta, D. K. Pal, and B. Palsamanta, “Flood susceptibility analysis through
remote sensing, GIS and frequency ratio model,” Appl. Water Sci., vol. 8, no. 2, pp.
1–14, 2018, doi: 10.1007/s13201-018-0710-1.
[26] M. Shafapour Tehrany, L. Kumar, M. Neamah Jebur, and F. Shabani, “Evaluating
the application of the statistical index method in flood susceptibility mapping and its
comparison with frequency ratio and logistic regression methods,” Geomatics, Nat.
Hazards Risk, vol. 10, no. 1, pp. 79–101, 2019, doi:
10.1080/19475705.2018.1506509.
[27] O. of D. R. R. (UNDRR), “Global Assessment Report on Disaster Risk Reduction,”
2019. https://digitallibrary.un.org/record/3825375?ln=en#:~:text=The Global
Assessment Report on,efforts to reduce disaster risk.&text=The GAR aims to
focus,economic support for risk reduction.
[28] U. C. Nkwunonwo, M. Whitworth, and B. Baily, “Review article: A review and
critical analysis of the efforts towards urban flood risk management in the Lagos
region of Nigeria,” Nat. Hazards Earth Syst. Sci., vol. 16, no. 2, pp. 349–369, 2016,
doi: 10.5194/nhess-16-349-2016.
[29] G. Zhao, B. Pang, Z. Xu, D. Peng, and L. Xu, “Assessment of urban flood
susceptibility using semi-supervised machine learning model,” Sci. Total Environ.,
vol. 659, pp. 940–949, 2019, doi: 10.1016/j.scitotenv.2018.12.217.
[30] M. S. Tehrany, B. Pradhan, S. Mansor, and N. Ahmad, “Flood susceptibility
assessment using GIS-based support vector machine model with different kernel
types,” Catena, vol. 125, pp. 91–101, 2015, doi: 10.1016/j.catena.2014.10.017.
[31] G. Zhao, B. Pang, Z. Xu, D. Peng, and D. Zuo, “Urban flood susceptibility assessment
based on convolutional neural networks,” J. Hydrol., vol. 590, no. February, p.
53
 125235, 2020, doi: 10.1016/j.jhydrol.2020.125235.
[32] M. Rahman et al., “Flood Susceptibility Assessment in Bangladesh Using Machine
Learning and Multi-criteria Decision Analysis,” Earth Syst. Environ., vol. 3, no. 3,
pp. 585–601, 2019, doi: 10.1007/s41748-019-00123-y.
[33] E. Dodangeh et al., “Integrated machine learning methods with resampling
algorithms for flood susceptibility prediction,” Sci. Total Environ., vol. 705, 2020,
doi: 10.1016/j.scitotenv.2019.135983.
[34] K. Khosravi et al., “A comparative assessment of decision trees algorithms for flash
flood susceptibility modeling at Haraz watershed, northern Iran,” Sci. Total
Environ., vol. 627, pp. 744–755, 2018, doi: 10.1016/j.scitotenv.2018.01.266.
[35] M. S. Tehrany, S. Jones, and F. Shabani, “Identifying the essential flood
conditioning factors for flood prone area mapping using machine learning
techniques,” Catena, vol. 175, no. December 2018, pp. 174–192, 2019, doi:
10.1016/j.catena.2018.12.011.
[36] S. H. Mahmoud and T. Y. Gan, “Urbanization and climate change implications in
flood risk management: Developing an efficient decision support system for flood
susceptibility mapping,” Sci. Total Environ., vol. 636, pp. 152–167, 2018, doi:
10.1016/j.scitotenv.2018.04.282.
[37] H. Shahabi, A. Shirzadi, K. Ghaderi, and E. Omidvar, “Flood Detection and
Susceptibility Mapping Using Sentinel-1 Remote Sensing Data and a Machine
Learning Approach : Hybrid Intelligence of Bagging Ensemble Based on K-Nearest
Neighbor Classifier,” MDPI, 2020, doi: Remote Sens. 2020, 12, 266;
doi:10.3390/rs12020266.
[38] G. Zhao, B. Pang, Z. Xu, J. Yue, and T. Tu, “Mapping flood susceptibility in
mountainous areas on a national scale in China,” Sci. Total Environ., vol. 615, pp.
1133–1142, 2018, doi: 10.1016/j.scitotenv.2017.10.037.
[39] H. Darabi, B. Choubin, O. Rahmati, A. Torabi Haghighi, B. Pradhan, and B. Kløve,
“Urban flood risk mapping using the GARP and QUEST models: A comparative
study of machine learning techniques,” J. Hydrol., vol. 569, no. December 2018, pp.
142–154, 2019, doi: 10.1016/j.jhydrol.2018.12.002.
[40] A. Arabameri et al., “Modeling spatial flood using novel ensemble artificial
intelligence approaches in northern Iran,” Remote Sens., vol. 12, no. 20, pp. 1–30,
2020, doi: 10.3390/rs12203423.
[41] C. Cao, P. Xu, Y. Wang, J. Chen, L. Zheng, and C. Niu, “Flash flood hazard
susceptibility mapping using frequency ratio and statistical index methods in
54
 coalmine subsidence areas,” Sustain., vol. 8, no. 9, 2016, doi: 10.3390/su8090948.
[42] K. Uddin, M. A. Matin, and F. J. Meyer, “Operational flood mapping using multi-
temporal Sentinel-1 SAR images: A case study from Bangladesh,” Remote Sens., vol.
11, no. 13, 2019, doi: 10.3390/rs11131581.
[43] F. Fotovatikhah, M. Herrera, S. Shamshirband, K. W. Chau, S. F. Ardabili, and M. J.
Piran, “Survey of computational intelligence as basis to big flood management:
Challenges, research directions and future work,” Eng. Appl. Comput. Fluid Mech.,
vol. 12, no. 1, pp. 411–437, 2018, doi: 10.1080/19942060.2018.1448896.
[44] M. Vojtek and J. Vojteková, “Flood susceptibility mapping on a national scale in
Slovakia using the analytical hierarchy process,” Water (Switzerland), vol. 11, no. 2,
2019, doi: 10.3390/w11020364.
[45] B. Choubin, E. Moradi, M. Golshan, J. Adamowski, F. Sajedi-Hosseini, and A.
Mosavi, “An ensemble prediction of flood susceptibility using multivariate
discriminant analysis, classification and regression trees, and support vector
machines,” Sci. Total Environ., vol. 651, pp. 2087–2096, 2019, doi:
10.1016/j.scitotenv.2018.10.064.
[46] A. J. Echendu, “The impact of flooding on Nigeria’s sustainable development goals
(SDGs),” Ecosyst. Heal. Sustain., vol. 6, no. 1, 2020, doi:
10.1080/20964129.2020.1791735.
[47] O. Rahmati, H. Zeinivand, and M. Besharat, “Flood hazard zoning in Yasooj region,
Iran, using GIS and multi-criteria decision analysis,” Geomatics, Nat. Hazards Risk,
vol. 7, no. 3, pp. 1000–1017, 2016, doi: 10.1080/19475705.2015.1045043.
[48] S. F. Balica, I. Popescu, L. Beevers, and N. G. Wright, “Parametric and physically
based modelling techniques for flood risk and vulnerability assessment: A
comparison,” Environ. Model. Softw., vol. 41, pp. 84–92, 2013, doi:
10.1016/j.envsoft.2012.11.002.
[49] O. Rahmati, H. R. Pourghasemi, and H. Zeinivand, “Flood susceptibility mapping
using frequency ratio and weights-of-evidence models in the Golastan Province,
Iran,” Geocarto Int., vol. 31, no. 1, pp. 42–70, 2016, doi:
10.1080/10106049.2015.1041559.
[50] M. Shafapour Tehrany, F. Shabani, M. Neamah Jebur, H. Hong, W. Chen, and X.
Xie, “GIS-based spatial prediction of flood prone areas using standalone frequency
ratio, logistic regression, weight of evidence and their ensemble techniques,”
Geomatics, Nat. Hazards Risk, vol. 8, no. 2, pp. 1538–1561, 2017, doi:
10.1080/19475705.2017.1362038.
55
[51] R. Costache and D. Tien Bui, “Spatial prediction of flood potential using new
ensembles of bivariate statistics and artificial intelligence: A case study at the Putna
river catchment of Romania,” Sci. Total Environ., vol. 691, pp. 1098–1118, 2019,
doi: 10.1016/j.scitotenv.2019.07.197.
[52] R. Costache and D. Tien Bui, “Identification of areas prone to flash-flood
phenomena using multiple-criteria decision-making, bivariate statistics, machine
learning and their ensembles,” Sci. Total Environ., vol. 712, 2020, doi:
10.1016/j.scitotenv.2019.136492.
[53] H. Hong, W. Chen, C. Xu, A. M. Youssef, B. Pradhan, and D. Tien Bui, “Rainfall-
induced landslide susceptibility assessment at the Chongren area (China) using
frequency ratio, certainty factor, and index of entropy,” Geocarto Int., vol. 32, no. 2,
pp. 139–154, 2017, doi: 10.1080/10106049.2015.1130086.
[54] Y. Wang et al., “Flood susceptibility mapping in Dingnan County (China) using
adaptive neuro-fuzzy inference system with biogeography based optimization and
imperialistic competitive algorithm,” J. Environ. Manage., vol. 247, no. July, pp.
712–729, 2019, doi: 10.1016/j.jenvman.2019.06.102.
[55] A. Mosavi, P. Ozturk, and K. W. Chau, “Flood prediction using machine learning
models: Literature review,” Water (Switzerland), vol. 10, no. 11, pp. 1–40, 2018,
doi: 10.3390/w10111536.
[56] M. B. Kia, S. Pirasteh, B. Pradhan, A. R. Mahmud, W. N. A. Sulaiman, and A.
Moradi, “An artificial neural network model for flood simulation using GIS: Johor
River Basin, Malaysia,” Environ. Earth Sci., vol. 67, no. 1, pp. 251–264, 2012, doi:
10.1007/s12665-011-1504-z.
[57] S. Lee and I. Park, “Application of decision tree model for the ground subsidence
hazard mapping near abandoned underground coal mines,” J. Environ. Manage.,
vol. 127, pp. 166–176, 2013, doi: 10.1016/j.jenvman.2013.04.010.
[58] S. Lee et al., “Spatial prediction of flood susceptibility using random-forest and
boosted-tree models in Seoul metropolitan city , Korea,” Sci. Total Environ., vol.
5705, p. 20, 2017, doi: 10.1080/19475705.2017.1308971.
[59] Y. Wang et al., “A hybrid GIS multi-criteria decision-making method for flood
susceptibility mapping at Shangyou, China,” Remote Sens., vol. 11, no. 1, 2019, doi:
10.3390/rs11010062.
[60] B. Ayo, “INTEGRATING OPENSTREETMAP DATA AND SENTINEL-2 IMAGERY
FOR CLASSIFYING AND MONITORING INFORMAL SETTLEMENTS,” 2020.
[61] P. Data, “City Population,” 2020. http://www.citypopulation.de/ (accessed Sep. 14,
56
 2020).
[62] U. of E. Anglia, “Rainfall Data,” 2020. http://www.cru.uea.ac.uk/data (accessed Sep.
18, 2020).
[63] O. S. Map, “Road Data obtained from Geofabrik Website.”
http://www.geofabrik.de/ (accessed Oct. 02, 2020).
[64] U. C. Nkwunonwo, F. I. Okeke, E. S. Ebinne, and N. E. Chiemelu, “Free, open,
quantitative and adaptable digital soil map data and database for Nigeria,” Data Br.,
vol. 31, p. 105941, 2020, doi: 10.1016/j.dib.2020.105941.
[65] M. S. Tehrany, B. Pradhan, and M. N. Jebur, “Flood susceptibility mapping using a
novel ensemble weights-of-evidence and support vector machine models in GIS,” J.
Hydrol., vol. 512, no. May, pp. 332–343, 2014, doi: 10.1016/j.jhydrol.2014.03.008.
[66] H. Mojaddadi, B. Pradhan, H. Nampak, N. Ahmad, and A. H. bin Ghazali,
“Ensemble machine-learning-based geospatial approach for flood risk assessment
using multi-sensor remote-sensing data and GIS,” Geomatics, Nat. Hazards Risk,
vol. 8, no. 2, pp. 1080–1102, 2017, doi: 10.1080/19475705.2017.1294113.

57