Flood Disaster Analysis Based On Remote Sensing Images

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 37

1

Running Head: FLOOD DISASTER ANALYSIS

Flood Disaster Analysis Based On Remote Sensing Images

A Thesis Submitted

To Faculty of Aeronautics

In Partial Fulfillment of the Requirements

For the Degree of Bachelor of Aerospace Engineering

By

Persis Mugisha Ainembabazi

Advisor: Liu Ganchao


2
FLOOD DISASTER ANALYSIS

Acknowledgement

My greatest appreciation goes to my parents Aggrey and Delphine Mugisha, for all the

motivational and financial support that they have given to me, as well as my relatives for their

constant moral support while doing this research.

Special thanks goes to my Research Supervisor, Mr. Liu Ganchao for his invaluable

constructive advice and guidance throughout the course of this study that has made it possible

for this paper to be a success.

I sincerely appreciate Mr. Shafiq Nedala, a GIS and Remote sensing specialist for his expert

advice while consulting during this research. I thank him for training and equipping me with

the required skills for using various software which assisted me in completing this research.
3
FLOOD DISASTER ANALYSIS

Abstract

Floods are a severe occurring natural calamity in both rural and urban areas, affecting the

livelihood of people and the structural infrastructure of an area. Remote sensing image

change detection (CD) has enhanced the understanding of significant land surface changes on

the Earth’s surface between bi-temporal images. In this study, sentinel 2 imagery was used to

identify and monitor differences by observing in the same geographical area across a period

of time. The images used date back to the years 2019 (before), 2020 (during) and 2022

(after), the flood occurred, and provided information on the land cover of the area and extent

of the flood which enabled bi-temporal flood detection. Using the ArcGIS software, different

tools were run to create micro-watersheds. A model was satisfactorily trained to ensure the

stream discharge from the catchment goes in the right direction for each cell in the image.

Tests were done on datasets to analyse and contextualize the changes, as well as evaluate the

effectiveness of the method for flood detection. The Support Vector Machine (SVM)

algorithm best known for producing accurate results in classifying datasets was used for

image analysis. The results obtained showed an increasing extent of water around the

catchment over the period of time during which and after the flood occurred. The results

indicated that apart from heavy rainfall leading to the increasing water levels and the

occurrence of floods in the catchment area, it had also undergone land use and land cover

changes over that period of time due to factors such as population growth and human

activities along the shores of the lake.

KEY WORDS: remote sensing (RS), flood disaster, support vector machine (SVM), sentinel-

2
4
FLOOD DISASTER ANALYSIS

Table of Contents

Title …...................................................................................................................................... 1

Acknowledgement .................................................................................................................... 2

Abstract ..................................................................................................................................... 3

Table of Contents …................................................................................................................ 4

Chapter One: Introduction .................................................................................................... 5

1.1 Related Works…….……………………………………………………………............. 8


Chapter Two: Methodology .................................................................................................. 11

2.1 Study Area .................................................................................................................... 13

2.2 Data acquisition ............................................................................................................ 14

2.3 Image preprocessing ...................................................................................................... 16

2.4 Image analysis .............................................................................................................. 19

2.5 Accuracy Assessment ................................................................................................. 22

Chapter Three: Results and Discussion ............................................................................... 23

Chapter Four: Conclusion .................................................................................................... 31

References ............................................................................................................................... 32
5
FLOOD DISASTER ANALYSIS

Chapter One: Introduction

Flooding is a natural disaster that can lead to mortality, and potential damage to property assets.

They mostly occur in very populated areas as a result of quick, irregular and incoherent urban

sprawl hence raising the level of risk to the population. Between mid- April to late June 2020,

there was an extremely severe rainfall in Uganda (Mawerere & Omuya, 2021) which led to an

increment in the waters of Lake Victoria and further resulted into flooding along the shores.

This rise occurred (Cheptoris, 2020) from October 2019 to April 2020 and in that time the water

continuously increased from a level of 12.00 meters to 13.32 meters. This caused an impact on

other countries in the region where the water also emanates across into the Mediterranean Sea.

The Murchison Bay catchment area (Kiggundu et al., 2018, p.1) situated on the northern

shoreline of Lake Victoria basin in Uganda. It has a highly valuable ecological system due to

the many human activities that take place within the area. Life and settlement were affected by

the water that inundated the shores, flooded the markets, submerged homes and businesses

even those that were out of the 200perimeter buffer zone (Mawerere & Omuya, 2021)

encouraged by local authority.

Figure 1

Homes submerged.
6
FLOOD DISASTER ANALYSIS

Figure 2

Markets and businesses flooded.

An environmental expert (Mawerere & Omuya, 2021), expressed that the people settling in the

area had intruded on the shores of the Lake, therefore interrupting the way of nature. He warned

that the flooding had further been quickened by environmental degradation, and this led to a

disturbance of the water sources. Debris that is poorly disposed of containing plastics and

polythene, are washed up with soil into the lake which increases its water levels and leads to

flooding on the shores. This contamination affects the water PH level and leads to the

accumulation of algae thus changing its color. It poses a direct threat not only to the people’s

health causing them to contract diseases, but also to aquatic life. These conditions have

deteriorated the eminence of the natural environment and threatened the crucial ecological

benefits it provides.
7
FLOOD DISASTER ANALYSIS
Figure 3

Water Pollution in the catchment area.

The phenomenon of flooding is quite complex and subject to change. Satellites and remote

sensing (RS) devices are useful for tracking environmental disasters in huge areas. They have

a great spatial and temporal resolution making the capable of capturing multispectral images

of places that have been impacted. Intelligent automatic recognition (Yanbing et al., 2021, p.2)

of changes in ground features becomes more and more important for flood monitoring and

detecting dynamic changes with the use of bi-temporal remote sensing images. Remote sensing

images (Dey, et al., 2009, p.1) acquired from different satellite and aerial sensing devices give

data that is important for evaluating the flood extent. In this study, I use sentinel-2 imagery

acquired from the European Space Agency (ESA) which has free available optical datasets

from the Sentinel-2 sensor. Images of the years 2019 (pre-flood), 2020 (during the flood) and

2022 (post-flood) are used to carry out bi-temporal detection and monitor the extent of the

flood. Different algorithms have been used for automatic information extraction (Dey et al.,

2009, p.1). These include the usage of water indices like Normalized Difference Water index

(NDWI), the PCA transformation, as well as the difference between Land Surface Water Index
8
FLOOD DISASTER ANALYSIS

and Enhanced Vegetation Index (EVI). However, it can be difficult to separate the areas that

are completely inundated from those partially submerged by water during analysis. Therefore,

in this paper, I use an advanced machine learning algorithm known as the Support Vector

Machine (SVM) for image analysis. In comparison to other classification approaches (Tzotsos,

p.2) like Neural Networks, Nearest Neighbor, supervised Maximum Likelihood and Decision

Tree classifiers, it is superior in terms of accurateness (Huang et al 2002, Foody and Mathur

2004). It is based on statistical learning frameworks (Zhang, 2012, p.179) which are based on

a restricted sample number in the data included in the training dataset obtained for great output

after classification.

1.1 Related Works

D. Han et al. (2007), did a comparison of a variety of benchmarking models with the Support

Vector Machine model. It showed that it exceeded the rest of the models in the test data sets.

This comes at a high cost in terms of time and effort to accomplish it. Linear and nonlinear

kernel functions (i.e. RBF) performed superior opposed to one another based on various

situations within one catchment.

Costache et al. (2020), found there are models that are Machine learning methods for

estimating flood susceptibility such as Artificial Neural Network (ANN), support vector

machine (SVM) and decision tree-based models. There are deep learning neural networks as

well as excessive learning machines. These models after application got an accuracy above

80%. Ruslan et al. (2013), used Artificial Neural Network models for flood water level
9
FLOOD DISASTER ANALYSIS

prediction for early warning system using Back Propagation Neural Network with Neural

Network Inverse Model put at the output for enhanced outcomes. The Back Propagation

method was used in relation to dataset acquired to reduce the number of error function in

relation to the complication and working of Artificial Neural Networks. The outcome gave

bad prediction accomplishment. Therefore, a Neural Network Inverse Model was suggested

to be put at the result of the Back Propagation Neural Network. This led to an enhanced

performance.

Panapitiya et al. (2021) proposed a collective approach of Artificial Neural Network with

Internet of Things devices to make forecasting simpler and more dependable. Also the usage

of more variables increased accuracy of the predictions. An Artificial Neural Network model

is trained using information that is collected and combined with live information feed thus

predicting the water level. This enabled forecasting flood events based on present readings.

Opella et al. (2019), combined Convolution Network, a deep feed forward neural network

that deals in image processing and predicting non-linear data with Support Vector Machine. It

is a supervised machine learning algorithm for binary classification in order to get increased

accuracy and improved image map outcomes. The combination of both unique network

architecture gives a valuable flood map. The support vector machine and convolutional neural

network have displayed an advanced execution in image processing and prediction abilities.

Islam et al. (2020), did a comparison of different machine-learning methods used in flood

detection. These include multi-layer perceptron (MLP), support vector machine (SVM), deep

convolutional neural network (DCNN). A comparison was made of the performances of the

Semi-Supervised Domain Adaptation method combined with the aforementioned models in


10
FLOOD DISASTER ANALYSIS

fine-tuning. This was done on multi-temporal optical (Spot-5) and radar (SAR) sensor image

data for flood-detection. It was discovered that the methods achieved results that were alike

in flood detection. The SSDA method, which is a fusion of deep Convolution Neural

Networks with a semi-supervised domain adaptation strategy, performed effectively with the

least training samples in comparison with the other methods. Therefore the SSDA method is

valuable for flood detection.

Menon et al. (2021), used the following classifiers k-nearest neighbors, Logistic Regression,

Support Vector Classifier, Decision Tree, and Random Forest machine learning method.

Logistic Regression and Support Vector Machine surpassed the real accuracy level. More input

data and machine learning methods are required to be examined to find the most suitable

alternative for flood detection.


11
FLOOD DISASTER ANALYSIS

Chapter Two: Methodology

Flow Chart
12
FLOOD DISASTER ANALYSIS

The figure above shows the proposed method adopted throughout the study. Data collection is carried out

which involves an extensive review of various literature on papers about the different algorithms used for

flood disaster analysis. It also required planning and choosing the study area which is, the Murchison Bay

Catchment Area. Field data was then gathered by direct observation of the study area to obtain geographical

information and examine the existing physical condition. On a field visit, interviews were done with the

local people while recording audios and taking images of the areas impacted by the flood. Satellite imagery

was acquired through European Space Agency, for three Sentinel-2 images of the years 2019, 2020 and

2022. This enabled bi-temporal flood detection to be carried out. Watershed delineation of the catchment

area was performed using The ALOSPALSAR DEM of 12.5m resolution (elevation data). Image pre-

processing and classification was carried out using various tools accessible in the ArcGIS 10.5 software.

The Murchison Bay catchment area was extracted from the composite image for the classification of land

cover. A model was trained and tests were done on datasets to evaluate the effectiveness of the method for

flood detection. Training data was created using the simple random sampling method and it generated the

number of training samples which were 100. Since supervised classification is simple, it delivers a

satisfactory accuracy and gives a clear interpretation of outcomes, the support vector machine algorithm

was implemented for analysis. The support vector machine supervised classification tool in ArcGIS

classified the extracted data features. Three major classes were obtained: open water body, vegetation and

the built-up area. The classes for vegetation and the built-up area were then combined into one class and

the data was cleaned for the output results to be clearly represented as classified land cover maps. The

overall accuracy, user’s accuracy, producer’s accuracy were used for accuracy assessment of the final

change maps. If accuracy was less than 80%, then supervised classification was repeated in order to get

highly accurate classified output results. The accuracy of the Support Vector Machine in classifying water
13
FLOOD DISASTER ANALYSIS

and non-water pixels was also evaluated using the Precision, Recall, F-1 Score and Accuracy equations

whose values range from 0 to 1. The higher the value the greater the performance of the model.

2.1 Study Area

Figure 4

The Murchison Bay Watershed.

The Murchison Bay Catchment Area (Kiggundu et al., 2018, p.45) is an annex of Lake Victoria.

It is between latitudes 00˚10'00"N - 00˚30'00"N and longitudes 32˚35'00'E - 32˚50'00"E. It is

1224 m above sea level. The Bay (L. A. Anaba et al., 2017, p. 26) comprises of a mean total

catchment area of 282 km2. This contains 20% swamps and 80% upland. In Uganda, it is a

center of development and livelihood to the millions of people in the area (Kiggundu et al.,

2018, p.45). The people rely on it for food, water, employment, transport and recreation. The

Bay receives a bi-seasonal rainfall in the months of March to May and September to November.
14
FLOOD DISASTER ANALYSIS

However, the catchment area experienced intense flooding between mid- April to late June

2020 along the shores of the lake. The quick rise in water levels of Lake Victoria (Cheptoris,

2020) were also accelerated by human activities like swamp degradation, infringement on

wetlands, lakeshores and river banks. This also included poor land use practices that have led

to soil erosion and eventually siltation of the water sources. The fast movement of water into

the water bodies with much silt has minimized their water storing capabilities (Cheptoris, 2020).

The local council authority stated that people have tried to de-silt the water by pouring sand

but with no success in stopping the outcome of flooding.

2.2 Data acquisition

This comprised of gathering field data and direct observation of the study area to obtain

geographical information examine the existing physical condition during a field visit. Interviews

with the local people in the area while recording audios were carried out as well as taking images

of the areas impacted by the flood. Remote sensing image change detection technology is used

for observing and monitoring an area's environmental features. Satellite imagery was acquired

through European Space Agency, which provides free access to time series remotely sensed

multi spectral images from the Sentinel-2 sensor in order to carry out bi-temporal flood detection

on the catchment area. Satellite sensors give enhanced spectral and temporal resolutions for

multispectral images, crucial for classification. Three Sentinel-2 images of 10m spatial

resolution acquired from the years February 2019, September 2020 and March 2022 were used

to detect the flooded areas in the catchment and provide information on the extent of the flood.
15
FLOOD DISASTER ANALYSIS
Figure 5

Sentinel-2 image for before the flood (2019)

Figure 6

Sentinel-2 image for during the flood (2020)

Figure 7

Sentinel -2 image for after the flood (2022)


16
FLOOD DISASTER ANALYSIS

Different aspects were considered when choosing the Sentinel-2A and 2B images to be used.

These involved (Kiggundu et al., 2018, p.46) the year they were taken, the visibility due to the

amount of cloud cover. Also the obtainability of the images for the expected year and the

condition of the images in terms of quality were considered. Images with a maximum of 10%

cloud cover were selected in the search criteria. However, there were specific years with data

gaps and without clear images that had to be excluded. Therefore the nearest year with clearer

images to the relevant year either preceding or after it were selected. For example, although this

study required images showing land cover changes from 2019 to 2021, due to lack of clear

imagery and data gaps in the Sentinel-2 images of the year 2021, it was substituted for the year

2022. Having obtained the imagery remotely from sensors, the image data for pre-flood, during

the flood and post-flood will then be processed and analyzed using a software called ArcGIS

10.5 in order to establish the spatial relationship between satellite data and field measured

quality parameters.

2.3 Image preprocessing

This is done using ArcGIS 10.5 software, which is a server software and online geographic

computer system for extracting information from either satellite images or remotely sensed

data. It enables end users to perform spatial queries, analyze, and edit spatial data. It also

consists of GIS tools and workflows for image pre-processing and classification. Water

shedding delineation of the catchment area was carried out using ALOSPALSAR DEM (Digital

Elevation Model) of 12.5m resolution (elevation data) which is obtained from the Alaska

satellite and downloaded from Alaska Spatial Data Portal (ALSK). The image is radio-
17
FLOOD DISASTER ANALYSIS

metrically and geometrically corrected. The process of delineation involved filling void values

in the DEM by filling sinks, re-projecting the dataset to Arc 1960 UTM 36N, creating

Murchison Bay micro catchment and monitoring centered in the micro catchment. Flow

direction and flow accumulation tools were run to create micro watersheds. The model needed

to ensure that the water goes in the right direction for each cell in the image.

Figure 8

Watershed delineation of the catchment area

Prior to image classification, pre-processing techniques like geometric correction were

implemented basically to remove some systematic errors in the data increasing their

intelligibility as a representation of the object sensed, and extracting meaningful patterns from

the data. This corrected the GPS coordinates to a particular area. Images were also corrected

using top over atmosphere method and composite bands are generated. The images were

projected, cut and clipped based on the area of interest then classification of land use / land

cover was performed. Pre-processing of data (standardization) is highly commendable to

enhance accuracy of classification.


18
FLOOD DISASTER ANALYSIS
Figure 9

Pre-processing for the image before the flood (2019)

Figure 10

Pre-processing for the image during the flood (2020)

Figure 11

Pre-processing for the image after the flood (2022)


19
FLOOD DISASTER ANALYSIS

2.4 Image analysis

Image classification, training a model and tests on data sets are done as part of data analysis to

evaluate the effectiveness of the method for flood detection. After the area of interest was

clipped out it is then classified based on land cover. A classification-based method (Chen &

Shi, 2020, is used to identify the change category by classifying the extracted bi-temporal data

features. The process involved categorizing pixels or areas of images into classes in order to

signify various kinds of physical objects. Training data was created using the simple random

sampling method and it generated the number of training samples. Here each item had an equal

and likely possibility of getting randomly selected in the sample. Data was then got from as

large a percentage as possible of this random selection. The number of pixels in the entire study

area were used to create the sample area.

Formulas used for random sampling

The total number of possible training samples (without replacement)


𝑁!
= 𝑛!(𝑁−𝑛)!
(1)

The sum of possible training samples (with replacement)

= Nn (2)

N is the entire population magnitude; n is the sample size.

The number of training samples got are 100, and are grouped into three classes based on the

land cover use: water boundaries/open water (40), vegetation (30), and built-up areas (30).

N (sampling frame) = 717 km2 in ArcGIS software.

Analysis is done with The Support Vector Machine (SVM) (Tzotsos, p.2) which is a machine

learning method with high classification accuracy of datasets. It is built on statistical learning
20
FLOOD DISASTER ANALYSIS

frameworks that also uses supervised learning models to analyze data for classification. SVM

is the classifier that maximizes the margin and creates a hyper plane or a group of them

(“Support-vector machine,” p.2) in a high- or limitless-dimensional space.

Figure 12

The optimal surface evolved from the linearly separable case.

Two hyper planes (Zhang, 2012, p.180) are selected that separate the two classes of samples

parted by H without errors, H1 and H2 are areas that go through the current point of H. The

region (“Support-vector machine,” p.4) enclosed by both the hyper planes is referred to as the

"margin". An ideal separation is got by the hyper plane that has the biggest distance to the

closest training-data point of any class. The maximum-margin hyper plane is the one that is

just in between the two hyper planes. Generally the bigger the margin, the lesser the

generalization error of the classifier (“Support-vector machine”, p.2).

The linear discriminate function (Zhang, 2012, p.180) in the n-dimensional area is given by: g

(x) = wTx + b in reference to pattern recognition. The classification hyper plane equation is

recorded as: (ω ⋅ x) + b = 0. For normalized datasets, the hyper planes can be expressed by the

following the formulas:

wTx + b = 1 (3)

(whatever is on or above this border belongs to one class, marked 1)


21
FLOOD DISASTER ANALYSIS

wTx + b = -1 (4)

(whatever is on or below this border belongs to the other class, marked −1)

The discriminate function g(x) was standardized (Zhang, 2012, p.180) therefore all the

training samples are satisfied when |𝑔(𝑥)| ≥ 1. Similar to meeting away from the classification

2
of the surface of the sample|𝑔(𝑥)| = 1. Hence, the class interval is equal to || 𝑤 || which is the

distance between both hyper planes. Therefore the interval on the equivalent becomes

||𝑤|| 𝑜𝑟 ||𝑤|| 2. The constraint i is included for each to avoid data points from lying in the

margin. Maximize the distance such that:

wTxi + b ≥ 1 for yi = 1 (5)

wTxi + b ≤ -1 for yi = -1 (6)

These constraints express that each data point should fall on the appropriate part of the margin.

This can be rephrased as:

yi (wTxi + b) ≥ 1 for all 1 ≤ i ≤ n (7)

A classification of the surface of all samples appropriately classified is created, and it should

satisfy:

yi (wTxi + b) - 1 ≥ 0 (8)

Turn ||𝑤||2, the lowest classification surface (Zhang, 2012, p.180) is the optimum

classification surface. Whereas the max-margin hyper plane is totally influenced by the xi that

fall closest to it (“Support-vector machine”, p.5). The points on the hyper plane are called

support vectors and they sustain the optimum classification surface (“Support-vector

machine”, p.5). This method is very effective for high accuracy classification.
22
FLOOD DISASTER ANALYSIS

2.5 Accuracy Assessment

To assess accuracy, the user’s accuracy, producer’s accuracy, and overall accuracy (Opedes, et

al., 2022, p.7) were used and computed using the following equations:

Wi = (9)

Prop = Σ3 Wi (10)

User’s accuracy = (11)

Producer’s accuracy = (12)

1
Overall accuracy= ∑𝑛𝑖=1 𝑋𝑖𝑖 (13)
𝑁

Where: a is the sum of pixels per strata and b is the sum of pixels in the study area, do is the

appropriately classified pixels in the diagonal, r is the sum of row pixels, and c is the sum of

column pixels, x = singular cell values, xii = the sum of observations in row i and column i, n

= sum of classes, N = sum of samples

Table 1-1
Allocation of sample sites for validation data
Class Wi Ui Number of validation sites

Built-up area 0.062 0.91 30

Vegetation 0.055 0.67 30

Water Body 0.330 0.82 40

Wi is the mapped area proportions, Ui are values of user’s accuracies for year 2022.

Four other accuracy metrics are also used these are: (a) Precision, Recall, F1-score and

Accuracy to measure the classification accuracy of the support vector machine. The values
23
FLOOD DISASTER ANALYSIS

range from 0 to 1, and the higher the value, the more improved the model performance is.
𝑇𝑃
Precision =𝑇𝑃+𝐹𝑃

𝑇𝑃
Recall = 𝑇𝑃+𝐹𝑁

2∗𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛∗𝑅𝑒𝑐𝑎𝑙𝑙
F1- score = 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑅𝑒𝑐𝑎𝑙𝑙

𝑇𝑃+𝑇𝑁
Accuracy = 𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁

The true positive (TP) samples have pixels that the model correctly identifies the true classified

pixel type as the water pixel in the forecast. The true negative (TN) pixels are non-water pixels

in evaluation rightly identified as non-water pixels. The false positive (FP) samples are wrongly

identified as water pixels, which are non-water pixels in studied datasets. The false negative

(FN) samples are the non-water pixels identified by the support vector machine model, which

are water pixels in the studied datasets.

Chapter 3: Results and Discussion

In recent years, a great deal of research has been done in relation to flood detection using

remote sensing images. Optical images (Dey, et al., 2009, p.291) are quite simple to interpret.

However it can be difficult to delineate the land-water boundary, permeate challenges of

visibility due to cloud cover and establish the flood margin. A common approach used to solve

this is a threshold split-based method. Although, the optimal threshold is influenced by the

physical area, time, and atmospheric environment during which the image was captured.
24
FLOOD DISASTER ANALYSIS

Tanim et al. extracted flooding data from Sentinel 1 images and put it in different supervised

and unsupervised machine learning models. These include Random Forest (RF), Support Vector

Machine (SVM), and Maximum Likelihood Classifier (MLC), to identify pixels that are flooded

in the imagery and assess the working of the machine learning models. They came to the

conclusion that recent unsupervised flood imagery classification and detection algorithm had

improved outcomes with less requirements of data and computational time. This has increased

quick flood mapping.

Unsupervised machine learning algorithms proved to be tougher because of enhanced speed,

fewer conditions of training data, and computational running time. Therefore offering improved

computation efficacy. The incapability of machine learning approaches to find a crucial

threshold between water and non-water pixels made it difficult for unsupervised classification

algorithms to examine the Sentinel 1 imagery. Even for supervised classification algorithms.

Based on the implementation computations, the Support Vector Machine had the utmost

accuracy in all four models.

Bafitlhile & Li (2019), applied ε-Support Vector Machine (ε-SVM) and artificial neural network

(ANN) for simulating and stream flow forecasts of three distinct catchments. A comparison was

made of the implementation of both models and outcome was satisfactory for humid and semi-

humid areas. Support Vector Machine outperforms Artificial Neural Networks in stream flow

simulation and forecasting. A unique feature is both models showed a good performance while

carrying out the simulation. ANN performed badly in semi-arid catchments. The outcome of the

study proved that Support Vector Machine has an improved performance compared to Artificial

Neural Networks. The results revealed that the ANN model achieved high accuracy and
25
FLOOD DISASTER ANALYSIS

reliability in terms of stream flow forecasts of humid and semi-humid catchments. However,

the model faced difficulty in relation to underestimating the peak flow. The support vector

machine model obtained high accurate values for the minimum and maximum of the basins and

forecast times.

Over the years, deep learning models (Bentivoglio, et al., 2021, p.1) have shown increased

accuracy, speed when compared to traditional approaches and increased speed when compared

to numerical methods for flood analysis. The current trend in studies on flood detection relies

on Neural Networks. The convolutional neural networks (Lianchong Zhang & Junshi Xia, 2021,

p.2) have been effectual in flood risk / damage evaluation, even more accurate in categorizing

land cover and image scene. Therefore they overcome the limitations of numerical models and

traditional approaches for flood mapping. Classification of the sentinel-2 images was performed

using supervised Support Vector Machine. This algorithm was found to be memory efficient,

and had robust tools. Support Vector Machine can evaluate important data appropriately. Its

value of following a structural risk minimization principle has enabled it in maximizing the

margin; therefore, its broad based use does not reduce. However this algorithm has faced a

number of difficulties in the field of change detection for example, the computing and storing

requisites have enhanced quickly as much as the number of training vectors. D. Han et al. (2007),

discovered that just as artificial neural network models, face a difficulty from over fitting and

under-fitting complications with the over-fitting being more problematic than under-fitting. The

same goes for the support vector machine. It was shown that an optimal choice from huge input

combinations and parameters is a difficulty for modelers with Support Vector Machines.
26
FLOOD DISASTER ANALYSIS

Bafitlhile & Li (2019) found that the major setback of the prediction accuracy of SVM was the

occurrence of lags, and the lag phase enhanced with the forecast lead time.

Support vector machine method is not appropriate for big data sets because the complication of

the algorithm’s training is increasingly impacted on by the size of the dataset. Therefore the

training time is higher with the dataset until it is unusable due to computation restrictions.

Support Vector Machine has poor performance and doesn’t execute well when the data set has

a lot of noise in other words the target classes are overlapping. The features can have very

similar or overlapping properties. Noise should not be the problem for kernels with high-bias,

such as the linear and polynomial kernels. The issue of noise should lie more for the low-bias

kernels such as the radial basis function (RBF).

When the number of properties for each data point surpasses the number of training data

specimen, the Support Vector Machine gives an underperformance. The vulnerability of the soft

margin optimization problem. It has resulted in the hyper planes being tilted to the lesser class

when the training data used is excessive.

To solve a nonlinear problem Support Vector Machine uses the kernel trick. The complexity of

this method based on classification could be reduced by the usage of the radial basis kernel

function therefore selecting the correct kernel function is important. If the data is not linearly

separable, and the linear kernel is used it leads to the algorithm not performing well. A common

technique used to select the appropriate kernel is changing the kernel function in the hyper

parameter search.
27
FLOOD DISASTER ANALYSIS

Since the support vector classifier operates by putting data points, over and beneath the

classifying hyper plane there isn’t a probabilistic explanation for the classification. Overall the

Support Vector Machine method achieves very high classification accuracy.

After training datasets with the Support Vector Machine classifier and the classify raster tool,

then the model was run. The data was further edited to combine the vegetation and build up

training sample classes into one class using the reclassify tool.

Figure 13

Classification of the image before the flood (2019)

Figure 14

Classification of the image during the flood (2020)


28
FLOOD DISASTER ANALYSIS

Figure 15

Classification of the image after the flood (2022)

The output results were then converted from raster to vector and cleaned by merging wrongly

classified layers to the correctly classified layers. The classified maps achieved high accuracy

values in terms of percentages. The overall classification accuracy (Kiggundu et al., 2018, p.48)

is the amount of appropriately categorized samples of an error matrix. It was computed by the

division of the sum of appropriately categorized samples by the sum of referenced samples.

Table 1-2
Classification accuracy assessment of the land cover at the catchment area.

Years 2019 2019 2020 2020 2022 2022

Classes UA (%) PA (%) UA (%) PA (%) UA (%) PA (%)

Water Body 88 88 83 83 91 91

Build up &
95 92 86 82 87 94
Vegetation
Overall 86.67% 88.67% 89%

Accuracy
29
FLOOD DISASTER ANALYSIS

Results from the performance evaluation of the Support Vector Machine model showed a

precision measurement of 0.85, a recall value of 0.85, the F1-score of 0.85 and therefore the

accuracy measurement of 0.87 is obtained. Basing on the final output results, computations were

made in order to discuss the changes that have been recognized and quantified. Experimental

results showed there was an increase in the extent of water over the period of time during which

and after the flood occurred. Initially the total area of the Murchison Bay catchment covered by

the lake was 187.3321 km2 in 2019, and when the flood occurred in 2020 it increased by

approximately 3.2 km2 to 190.5 km2. After the flood occurred, by 2022 the water extent of the

lake had decreased by 2.5km2 to 188.1 km2. Although this is still more than the initial area of

the lake before the flood occurred it indicated that the lake was returning to its normal area

coverage before the flooding took place. The results indicated that apart from experiencing

heavy rainfall leading to increasing water levels and the occurrence of floods in the catchment

area, it had also undergone land cover changes in that period of time. This is due to factors such

as population growth and human activities along the shores of the lake.

Table 1-3
Computation of Results

Row Labels Sum Of Area


Water Body 187.3321046

Table 1-4
AREA (sq.km)
YEARS OF FLOODING EXTENT INCREASE
BEFORE (2019) 187.3 0
DURING (2020) 190.5 3.2
AFTER (2022) 188.1 -2.5
30
FLOOD DISASTER ANALYSIS

Figure 16

Final Output Results


31
FLOOD DISASTER ANALYSIS

Chapter Four: Conclusion

Recently, suitable flood detection products developed from remote sensing imagery have

become efficient approaches of assessing flood disasters. Different sectors have benefitted from

them worldwide. Although recognizing permanent and temporary water in the event of a flood

is difficult, an efficient way to minimize flood damages is to improve flood risk mitigation and

evaluation. The results are essential for local and regional level planning studies in assessing

urban development trends in the catchment.

Due to environmental degradation, there is an increased rate of runoff into the catchment

lowering the quality of water and reducing plant growth within the Murchison Bay. Steps toward

extensive water management ought to be considered in order to minimize surface overflow into

the lake. Communities should be sensitized on the importance of environmental conservation

and swamp reclamation. They need to be informed about the dangers of poorly disposed debris

to their health and the aquatic life as well. These conditions have deteriorated the value of the

natural environment and threatened vital ecological amenities that it provides to the people

within the Murchison Bay Catchment area.

Therefore, there is a great need for the government to implement urgent and effective policies

to control as well as monitor water and land use, in order to preserve nature and the eco system.
32
FLOOD DISASTER ANALYSIS

References

Cheptoris, Sam. “Statement on the Rising water levels of Lake Victoria and the Nile System.”

Uganda Media Center. May 1st, 2020. https://www.mediacentre.go.ug/media/ statement-

rising-water-levels-lake-victoria-and-nile-system.

Nicholas Kiggundu, Listowel Anaba, Noble Banadda, Joshua Wayama & Isa Kabenge.

“Assessing Land Use and Land Cover Changes in the Murchison Bay Catchment of Lake

Victoria Basin in Uganda.” Journal of Sustainable Development. vol. 11, no. 1, 2018.

https://doi.org/10.5539/jsd.v11n1p44

Mawerere Sarah, and Omuya Noah. “Lake Victoria Breathes Amidst COVID-19 Pandemic:

Recaptures Its Tributaries.” Pulitzer Center. July 24th, 2020. https://pulitzercenter.org/

stories/lake-victoria-breathes-amidst-covid-19-pandemic-recaptures-its-tributaries.

V. Vanama, Y. Rao & C. Bhatt “Change detection-based flood mapping using multi-temporal

Earth Observation satellite images: 2018 flood event of Kerala, India.” European Journal

of Remote Sensing. vol. 54, no. 1: 42-58, 2021. DOI: 10.1080/22797254.2020.1867901.

Y. Bai, W. Wu, Z. Yang, J. Yu, B. Zhao, X. Liu, H. Yang, E. Mas, and S. Koshimura.

“Enhancement of Detecting Permanent Water and Temporary Water in Flood Disasters by

Fusing Sentinel-1 and Sentinel-2 Imagery Using Deep Learning Algorithms:

Demonstration of Sen1Floods11 Benchmark Datasets.” Remote Sens., vol. 13, no. 2220,

2021, https://doi.org/10.3390/rs13112220.
33
FLOOD DISASTER ANALYSIS

D. Chandrama, X. Jia, D. Fraser, and L. Wang. “Mixed Pixel Analysis for Flood Mapping Using

Extended Support Vector Machine.” Digital Image Computing: Techniques and

Applications, 2009. DOI 10.1109/DICTA.2009.55

T. Angelos, “A Support Vector Machine Approach For Object Based Image Analysis.” July

2006.https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.222.5004&rep=rep1&ty

pe=pdf.

G. Foody and A. Mathur, “A Relative Evaluation of Multiclass Image Classification by Support

Vector Machines.” IEEE Transactions On Geoscience And Remote Sensing, vol. 42, no. 6,

June 2004.

C. Huang, L. Davis, and J. Townshend, “An assessment of support vector machines for land

cover classification,” Int. J. Remote sensing, vol. 23, no. 4, pp. 725–749, 2002.

Y. Zhang. “Support Vector Machine Classification Algorithm And Its Application.” ICICA

Part II, CCIS 308, pp. 179–186, 2012.

L. Anaba, L. Banadda, N. Kiggundu, N. Wanyama, J. Engel, and D. Moriasi, “Application of

SWAT to Assess the Effects of Land Use Change in the Murchison Bay Catchment in

Uganda.” Computational Water, Energy, and Environmental Engineering, vol. 6, pp. 24-

40 2017. http://dx.doi.org/10.4236/cweee.2017.61003.

T. Madhuri. “Simple Random Sampling.” https://www.wallstreetmojo.com/simple-random-

sampling/.

H. Chen and Z. Shi, “A Spatial Temporal Attention Based Method and a New Dataset for

Remote Sensing Image Change Detection.” Remote Sens., 2020, doi:10.3390/rsxx010005.


34
FLOOD DISASTER ANALYSIS

“Support-vector machine.” Wikipedia. https://en.wikipedia.org/wiki/Support-vector_machine.

“Support Vector Machines.” Scikit Learn. https://scikit-learn.org/stable/modules/svm.html.

A. Saini, “Support Vector Machine (SVM): Oct 2021. A Complete guide for beginners.”

https://www.analyticsvidhya.com/blog/2021/10/support-vector-machinessvm-a-

complete-guide-for-

beginners/#:~:text=Margin%20in%20Support%20Vector%20Machine,and%20b%20is%

20an%20offset.&text=If%20the%20value%20of%20w,it%20is%20a%20negative%20po

int

R. Bentivoglio, E. Isufi, S. Jonkman, and R. Taormina, “Deep Learning Methods for Flood

Mapping: A Review of Existing Applications and Future Research Directions.” Dec 2021,

https://doi.org/10.5194/hess-2021-614.

M. Anastasia, S. Andreadis, I. Gialampoukidis, S. Vrochidis, V. Sitokonstantinou, D. Hoppe,

M. Gienger, L. Zhong. “Change detection techniques in Earth Observation” EOPEN Open

Interoperable Platform for unified access and analysis of Earth Observation data;

H2020776019 D4.1.

L. Zhang and J. Xia, “Flood Detection Using Multiple Chinese Satellite Datasets during 2020-

China Summer Floods.” Remote Sens., vol. 14, no. 51, 2022.

G. David , M. Torres, and E. Reinoso “Flood mapping through principal component analysis of

multitemporal satellite imagery considering the alteration of water spectral properties due

to turbidity conditions.” Geomatics, Natural Hazards and Risk, vol. 8, no. 2, pp. 607-623,

2017, DOI:10.1080/19475705.2016.1250115.
35
FLOOD DISASTER ANALYSIS

N. Tkachenko, R. Procter, and S. Jarvis “Predicting the impact of urban flooding using open

data.” R.Soc. open sci., vol. 3, no. 160013, 2016. http://dx.doi.org/10.1098/rsos.160013.

H. Opedes, H. Mücher, S. Baartman, S. Nedala, F. Mugagga, “Land Cover Change Detection

and Subsistence Farming Dynamics in the Fringes of Mount Elgon National Park, Uganda

from 1978–2020.” Remote Sensing, vol. 44, no. 2423, 2022.

https://doi.org/10.3390/rs14102423.

C. Yuan, F. Wang, S. Wang, and Y. Zhou, “Accuracy evaluation of flood monitoring based on

multi-scale remote sensing for different landscapes,” Geomatics, Natural Hazards and Risk,

vol. 10, no. 1, pp. 1389-1411, DOI: 10.1080/19475705.2019.1580224.

Pontus Olofsson, Giles M. Foody, Martin Herold, Stephen V. Stehman, Curtis E.Woodcock, M.

Wulder. “Good practices for estimating area and assessing accuracy of land change.”

Remote Sensing of Environment, vol. 148, pp. 42–57, 2014.

C. David, M. Speck, B. Devereux, G. Schumann, J. Neal, P. Bates. “Flood Detection In Urban

Areas Using TerraSAR-X,” IEEE Transactions on Geoscience and Remote Sensing, vol.

48, no. 2, 2010.

Fazlina Ahmat Ruslan, Nur Khalidah Zakaria, Ramli Adnan. “Flood Modelling using Artificial

Neural Network.” IEEE 4th Control and System Graduate Research Colloquium, pp. 19 –

20, 2013.

Roshni R Menon, Sandhra Simon, Maria A S, Rosemaria Shaju, Sachin. “Detection of Flood

Images using Different Classifiers.” International Journal of Innovative Science and

Research Technology, Volume 6, Issue 6, ISSN No:-2456-2165.


36
FLOOD DISASTER ANALYSIS

Thabo Michael Bafitlhile and Zhijia Li, “Applicability of ε-Support Vector Machine and

Artificial Neural Network for Flood Forecasting in Humid, Semi-Humid and Semi-Arid

Basins in China.” Water 2019, 11, 85; doi:10.3390/w11010085

Tanim, A.H.; McRae, C.B.; Tavakol-Davani, H.; Goharian, E. “Flood Detection in Urban Areas

Using Satellite Imagery and Machine Learning.” Water 2022, 14, 1140.

https://doi.org/10.3390/w14071140

Romulus Costache, Alireza Arabameri, Ismail Elkhrachy, Omid Ghorbanzadeh & Quoc Bao

Pham (2021) Detection of areas prone to flood risk using state-of-the-art machine learning

models, Geomatics, Natural Hazards and Risk, 12:1, 1488-1507, DOI:

10.1080/19475705.2021.1920480 https://doi.org/10.1080/19475705.2021.1920480

D. Han, L. Chan and N. Zhu. “Flood forecasting using support vector machines.” Journal of

Hydroinformatics 09.4, doi: 10.2166/hydro.2007.027, 2007.

http://iwaponline.com/jh/article-pdf/9/4/267/392906/267.pdf

Kazi Aminul Islam, Mohammad Shahab Uddin, Chiman Kwan and Jiang Li. “Flood

Detection Using Multi-Modal and Multi-Temporal Images: A Comparative Study.”

Remote Sens. 2020, 12, 2455; doi:10.3390/rs12152455

www.mdpi.com/journal/remotesensing

Joe Marlou A. Opella and Alexander A. Hernandez. “Developing a Flood Risk Assessment
37
FLOOD DISASTER ANALYSIS

Using Support Vector Machine and Convolutional Neural Network: A Conceptual

Framework” 2019 IEEE 15th International Colloquium on Signal Processing & its

Applications (CSPA 2019), 8 -9 March 2019, Penang, Malaysia.

Panapitiya P.K.D.C.R, Dhammearatchi D, Perera.R. “Flood Early Warning and Prediction

System for Tributary Streams.” CINEC Academic Journal Volume 5 Issue 1 2021.

Shruthi J, Sumathi M S, Srivatsa Raju S, Vidya R Pai. “Forecasting & Detection of Flood Using

Random Forest Learning Method.” European Journal of Molecular & Clinical Medicine

ISSN 2515-8260 Volume 07, Issue 08, 2020.

You might also like