INTEGRATING OPENSTREETMAP DATA AND SENTINEL-
2
IMAGERY FOR CLASSIFYING AND MONITORING
INFORMAL SETTLEMENTS
B enda Ayo
ii
INTEGRATING OPENSTREETMAP DATA AND
SENTINEL-2 IMAGERY FOR CLASSIFYING AND
MONITORING INFORMAL SETTLEMENTS
Disse a ion supe ised by
Joel Dinis Bap is a Fe ei a da Sil a, PhD
Ins i u o Supe io de Es a ís ica e Ges ão de In o mação,
Uni e sidade No a de Lisboa
Lisbon, Po ugal
Disse a ion co-supe ised by
P o . D . Hanna Meye ,
Ins i u e o Landscape Ecology,
Heisenbe gs . 2, D-48149 Müns e
Disse a ion co-supe ised by
Ignacio Gue e o
Ins i u e o New Imaging Technologies,
Uni e si a Jaume I
Cas ellón de la Plana, Spain
Feb ua y 2020
iii
DECLARATION OF ORIGINALITY
I decla e ha he wo k desc ibed in his documen is my own and no om someone
else. All he assis ance I ha e ecei ed om o he people is duly acknowledged and all
he sou ces (published o no published) a e e e enced.
This wo k has no been p e iously e alua ed o submi ed o NOVA In o ma ion
Managemen School o elsewhe e.
Lisbon, Po ugal, Janua y 2020
B enda Ayo
i
ACKNOWLEDGMENTS
P incipally, I would like o hank my Supe iso s, Joel Dinis Bap is a Fe ei a da Sil a
(PhD), P o . D . Hanna Meye and Ignacio Gue e o. You we e ins umen al om he
incep ion, h ough he execu ion and end o his Thesis. I am ex emely g a e ul and
hono ed o ha e collabo a ed wi h you.
I am ex emely g a e ul o all he p o esso s ha helped h ough his mas e ’s p og am
especially p o esso s Ma co Painho and Ch is oph B ox o hei guidance and suppo
h oughou he p og am du a ion.
Thank you o my coho and ellow MSc. GeoTech g adua e s uden s – especially
Damien and Chamodi - you ha e been inc edible esou ces, whose posi i e and
ha dwo king spi i has made my Mas e ’s expe ience so much mo e enjoyable and un.
Las ly, hanks o my iends and amily who we e o g ea suppo e en i you we e
housands o miles away.
In eg a ing Opens ee map Da a and Sen inel-2 Image y o
Classi ying and Moni o ing In o mal Se lemen s
ABSTRACT
The iden i ica ion and moni o ing o in o mal se lemen s in u ban a eas is an impo an
s ep in de eloping and implemen ing p o-poo u ban policies. Unde s anding when,
whe e and who li es inside in o mal se lemen s is c i ical o e o s o imp o e hei
esilience. This s udy aims a in eg a ing OSM da a and sen inel-2 image y o
classi ying and moni o ing he g ow h o in o mal se lemen s me hods o map in o mal
a eas in Kampala (Uganda) and Da es Salaam (Tanzania) and o moni o hei g ow h
in Kampala. Th ee building ea u e cha ac e is ics o size, shape and Dis ance o nea es
Neighbou we e de i ed and used o clus e and classi y in o mal a eas using Ho spo
Clus e analysis and ML app oach on OSM buildings da a. The esul an in o mal
egions in Kampala we e used wi h Sen inel-2 image iles o in es iga e he spa io-
empo al changes in in o mal a eas using Con olu ional Neu al Ne wo ks (CNNs).
Resul s om Op imized Ho Spo Analysis and Random Fo es Classi ica ion show ha
In o mal egions can be mapped based on building ou line cha ac e is ics. An accu acy
o 90.3% was achie ed when an op imally ained CNN was execu ed on a es se o
2019 sa elli e image iles. P edic ions o in o mali y om new da ase s o he yea s
2016 and 2017 p o ided p omising esul s on combining di e en open sou ce
geospa ial da ase s o iden i y, classi y and moni o in o mal se lemen s.

i
KEYWORDS
In o mal Se lemen s;
Remo e Sensing;
U baniza ion;
Machine Lea ning (ML);
Random Fo es (RF);
Con olu ional Neu al Ne wo ks (CNN);
OpenS ee Map (OSM)
Sen inel-2 sa elli e image y
ii
ACRONYMS
CNN - Con olu ional Neu al Ne wo ks (CNNs)
DEM - Digi al ele a ion model
DSM - Digi al su ace model
EO – Ea h Obse a ion
GEOBIA - Geog aphic objec -based image analysis
GPS - Global posi ioning sys em
GSO - Global slum on ology
HR - High Resolu ion Sa elli e Image y
ISDA - In o mal se lemen da abase a las
ISUP - In o mal se lemen upg ading P og amme
KOTAKU - Ko a Tanpa Kumuh
LiDAR - Ligh de ec ion and anging
MGD - Millennium de elopmen goals
MKL - Mul iple Ke nel Lea ning
MLTs - Machine Lea ning Techniques
OBIA - Objec -based image analysis
OHSA - Op imized Ho spo Analysis
OOA - Objec -o ien ed analysis
OSM – Open S ee Map
PSUP - Pa icipa o y slum upg ading P og amme
RF - Random o es
RGB – Red G een Blue
SDI - Shack/Slum Dwelle s In e na ional
SVM - Suppo Vec o Machine
UAV - Unmanned Ae ial Vehicle
iii
UN – Uni ed Na ions
UN-Habi a – Uni ed Na ions Human Se lemen s P og amme
VGG - Visual Geome y G oup
VHR - Ve y high esolu ion
VHR – Ve y High Resolu ion Sa elli e Image y
WDI - Wo ld De elopmen Indica o s
ix
INDEX OF THE TEXT
DECLARATION OF ORIGINALITY .................................................................. iii
ACKNOWLEDGMENTS ....................................................................................... i
ABSTRACT ..............................................................................................................
KEYWORDS ........................................................................................................... i
ACRONYMS ......................................................................................................... ii
INDEX OF THE TEXT .......................................................................................... ix
INDEX OF TABLES ............................................................................................... xi
INDEX OF FIGURES ........................................................................................... xii
1. INTRODUCTION .......................................................................................... 1
1.1. Backg ound and Mo i a ion........................................................................... 1
1.2. Resea ch Gap Iden i ica ion ........................................................................... 3
1.3. Resea ch Aim ................................................................................................ 4
1.4. Me hodology o e iew .................................................................................. 5
1.5. Thesis S uc u e ............................................................................................. 5
2. LITERATURE REVIEW ............................................................................... 6
2.1. Cha ac e is ics o In o mal Se lemen s .......................................................... 6
2.1.1. Building cha ac e is ics: ......................................................................... 6
2.1.2. Access Cha ac e is ics: ........................................................................... 7
2.1.3. Loca ion and neighbou hood Cha ac e is ics: ......................................... 7
2.2. OpenS ee Map o U ban Analy ics .............................................................. 7
2.3. App oaches o iden i ying and Mapping In o mal Se lemen s ...................... 8
2.3.1. Su ey and census-based app oach ......................................................... 8
2.3.2. Pa icipa o y-based app oach.................................................................. 9
2.3.3. GIS and Remo e sensing-based app oach ............................................... 9
2.4. Ho Spo Clus e ing Analysis ....................................................................... 11
2.5. An o e iew o Machine Lea ning Techniques ............................................ 11
2.5.1. Con olu ional neu al-ne wo ks.............................................................. 13
3. DATA AND CASE STUDY .......................................................................... 16
3.1. Case S udy .................................................................................................. 16
3.2. Da ase s and P e-P ocessing ........................................................................ 18
3.2.1. OSM Da a ............................................................................................. 18
3.2.2. Sa elli e Image y ................................................................................... 19
3.2.3. Re e ence da a....................................................................................... 20
4. RESEARCH METHODOLOGY ................................................................. 21
4.1. Da a En ichmen and Compu a ion o Pa ame e s ........................................ 21
4
polygons, and a he same ime iden i y in o mal egions ac oss he en i e ci y and no jus
pa icula places. The e is also need o examine whe he simila building polygons in in o mal
se lemen s - simila in he sense o geome ical p ope ies exhibi a spa ial pa e n. The in o mal
se lemen buildings cha ac e is ics coupled wi h Machine lea ning algo i hms could help o p edic
whe e in o mal se lemen s could be and also i hese cha ac e is ics a e exhibi ed in a ying
geog aphical loca ions. This will help educe he agg a a ion ha comes up du ing hei
de ec ion o in en o y o planning issues and also p o ide mo e insigh and unde s anding
on he g ow h o hese se lemen s.
The o he challenge wi h mapping in o mal se lemen s a ci y, egional o coun y le el is he
absence o adequa e da a o analysis a hese scales. The a ailabili y o open sou ce sa elli e
image y like Landsa and Sen inel da a p o ide a wide co e age bu he spa ial esolu ions
hey o e does no p o ide o adequa e dis inguishing be ween o mal and in o mal
se lemen s in ci ies. Howe e , wi h accu a e delinea ing o hese se lemen s a a ine scale,
hese sa elli e images can be used o moni o he g ow h o in o mali y o e ime since he
in eg a ion
o spa io- empo al analysis o in o mal se lemen s is s ill inadequa e.
1.3. Resea ch Aim
The main aim o his esea ch wo k is o in eg a e OSM da a and sen inel-2 image y o
classi ying and moni o ing he g ow h o in o mal se lemen s.
To achie e he main esea ch Aim, he ollowing sub ques ions a e add essed:
1. How can we exploi he po en ial o using buildings ou line cha ac e is ics such as size
and shape o di e en ia e In o mal se lemen s om Fo mal Se lemen s?
2. Is i possible o p edic in o mal a eas in a ci y by unde s anding housing in o mali y in
o he ci ies o simila con ex using buildings ou line cha ac e is ics and machine
lea ning?
3. Wha is he mos app op ia e Machine Lea ning echnique based on accu acy o p edic
in o mal a eas in a ci y based on buildings ou line cha ac e is ics?
4. How can we exploi he po en ial o eely a ailable Sen inel-2 sa elli e image y wi h
ad anced machine lea ning o es ima e he g ow h o In o mal se lemen s?

5
1.4. Me hodology o e iew
Based on he esea ch ques ions, he ollowing me hodology wo k low was adap ed:
 P ep ocessing o OSM da a and Sa elli e Image y. This in ol ed da a cleaning and
p epa a ion sui able o Analysis.
 OSM Building polygons en ichmen wi h Fea u e geome ic and opologic
cha ac e is ics.
 Ho spo clus e ing analysis on building ea u es aining da ase o clus e In o mal
se lemen buildings om he o mal ones.
 Tes and selec he bes classi ie based on accu acy o p edic in o mal a eas in a ci y
based on simila i y con ex o buildings ou line cha ac e is ics.
 Sen inel-2 image classi ica ion. This s age in ol es slicing he images in o iles and
using he p edic ed In o mal se lemen egions om he s ep abo e o gene a e aining
se on image iles o co esponding da e pe iod. The ea e , a Con olu ional Neu al
Ne wo ks Model is ained o iden i y Image iles ha belong o In o mal se lemen s.
The ained model is hen used o p edic In o mal se lemen image iles o o he
empo al da ase s.
1.5. Thesis S uc u e
The esea ch is o ganized as ollows:
 Chap e 2 e iews he ela ed wo ks and heo e ical backg ound on in o mal
se lemen s, app oaches o mapping and iden i ying in o mal se lemen s, and Machine
Lea ning Algo i hms
 Chap e 3 desc ibes he s udy a ea, da ase s and ools used o he esea ch.
 Chap e 4 desc ibes he implemen a ion de ails o he me hodology and he expe imen s
conduc ed
 Chap e 5 p esen s he esul s;
 Chap e 6 add esses he analysis and discussion o he esul s including he limi a ions
o he esea ch;
 Chap e 7 p esen s a summa y o he conclusions by answe ing he main e-sea ch
ques ions and ecommenda ions o u u e wo ks.
6
2. LITERATURE REVIEW
This chap e gi es an o e iew on he undamen al concep s ela ed o his esea ch and he
cu en s a e o esea ch in he ield o mapping in o mal se lemen s I highligh s he physical
cha ac e is ics o in o mal se lemen s, a discussion on a ious app oaches used o map
in o mal se lemen s and he in eg a ion o GIS and in o mal se lemen s and Machine Lea ning
Algo i hms.
2.1. Cha ac e is ics o In o mal Se lemen s
The challenge wi h mapping and iden i y wha an in o mal se lemen is s a s wi h he absence
o a s anda d de ini ion which has esul ed in o ailu e o inco po a e hese se lemen s in o
census and demog aphic su eys a na ional o e en dis ic le el [23]. Based on li e a u e, hey
ake on names like Slum, In o mal, Squa e , Spon aneous, Ghe o, Illegal, I egula , o
desc ibed by local names such as Fa ela, bidon ille (F ench), mudun-sa i (A abic), bai os da
la a (Po uguese) ownship o gecekondu (Tu kish), o men ion bu a ew [24]. The UN-Habi a
quali a i ely de ined a slum as a household ha is lacking ei he o imp o ed wa e , imp o ed
sani a ion, enu e secu i y o o e c owded li ing en i onmen is based on a measu e o
dep i a ion indica o s. On he o he hand, he de ini ion o in o mal se lemen s diffe s om
slum as hose a eas ha de eloped h ough unau ho ized occupa ion o land ou side a legal,
egula o y, planned and p o essional amewo k o example [26, 27]. The e o e, quan i ying
and classi ying in o mal se lemen s/ slums in a uni e sally accep able way is pa icula ly
di icul because (a) Wha is e e ed o as a slum in one coun y may be a good quali y o
li ing and (b) o e ime and a se lemen can ge o malized h ough slum upg ading p og ams
[27]. Howe e , o e he yea s, some ea u es ha e been unique o in o mal se lemen s which
include:
2.1.1. Building cha ac e is ics:
Buildings in in o mal se lemen s end o be smalle (10 o 40 m2), ha e simple shapes ha a e
mos ly ec angula and he e ogeneous o ien a ion compa ed o hose in o mal se lemen s. The
building densi y in slums is usually high since he spacing and gap be ween buildings in almos
nonexis ence [29, 30].
7
2.1.2. Access Cha ac e is ics:
Slums gene ally ha e i egula oad access ne wo k wi h oads ha a y in ype, su ace and
wid h. [30]The oads a e always sho in leng h wi h many dead ends and dangles. The oads
and access ne wo k in In o mal se lemen s a e usually na ow wi h limi ed usage by mo o
ehicles because hey a e ei he wo na ow o le a ca pass o can only allow access o on
ehicle a a ime. Slums a e cha ac e ized wi h mos ly oo pa hs.
2.1.3. Loca ion and neighbou hood Cha ac e is ics:
Slums a e usually loca ed in haza dous a eas like nea dumping si es, we lands, along ailways,
sewe age canals. [31]This is because mos o he land whe e hese ea u es a e loca ed is public
land and in abandoned. They a e usually close o employmen oppo uni ies o unskilled and
low skilled jobs like manu ac u ing indus ies o ease access o employmen .
2.2. OpenS ee Map o U ban Analy ics
Access o spa ial da a has changed apidly o e he yea s om adi ionally p ohibi i e sou ces
o openly licensed con en and da a due o changes in In o ma ion Technology and
Communica ion (ITC) b ough abou by in e ne , social media and inexpensi e po able GNSS
de ices like mobile phones and Handheld GPS [10, 32]. OSM is he mos popula geospa ial
open da a sou ce pla o m con aining billions o en ies o VGI and is main ained by a massi e
communi y o mappe s om a ound he wo ld all wo king o-wa ds he goal o cu a ing
accu a e and comple e geospa ial da a. In he ecen yea s, he e has been an inc eased use o
VGI da a no only in GIScience bu also in o he ields like ecology, planning, compu e science
e c. [33,34,35]. o academic esea ch and s udies. OSM oads laye s a e associa ed wi h he
highes comple eness compa ed o buildings based on a ecen s udy [36] ha es ima es ha
he OSM oads ha e eached mo e han 80% o comple eness a a global scale. OSM da ase s
ha e been used o ex ac u ban in o ma ion and o analysis [37, 38]. These da ase s a e usually
used as a supplemen a y sou ce o in o ma ion o land-use mapping wi h emphasis on
classi ying a i icial su aces [39], by combining OSM and sa elli e images o ex ac
in o ma ion ela ed o he u ban en i onmen [39–42]. Few esea ch s udies exis on building
oo p in s da a en ichmen o u ban analysis [43, 44]. In hese s udies, au oma ic me hods a e
de eloped by using he geome ic and opological ea u es in oo p in da a, in o de o enhance
he maps wi h he building usage in o ma ion. Henn e al. [45] de i ed he a chi ec u al ypes
8
o buildings based on 3D coa se block models wi h e ical walls and la oo s using Suppo
Vec o Machines (SVMs), whe eby, geome ic ea u es such as leng h, wid h, a ea, and
deg ee o pe pendicula i y o building oo p in s, ypes o buildings, as well as heigh
in o ma ion o buildings a e equi ed o he classi ica ion p ocess.
2.3. App oaches o iden i ying and Mapping In o mal Se lemen s
S udies on he de elopmen and exis ence o in o mal se lemen s ha e been ca ied ou no
only by geog aphe s bu also social scien is s who pu emphasis on social and economic aspec s
o in o mal se lemen s. Due o hei complexi y and e e a ying na u e, in o mal se lemen s
a e usually inadequa ely ep esen ed on u ban and ci y maps which lea es hem ulne able o
neglec by conce ned go e nmen agencies. The e a e b oadly h ee me hods used o collec
da a o in o m cha ac e iza ion and classi ica ion o in o mal se lemen s: household su eys
and census, Pa icipa o y app oaches and Ea h Obse a ion image y analysis using geospa ial
Techniques [46].
2.3.1. Su ey and census-based app oach
These app oaches use da a collec ed h ough census and su eys like da a on demog aphics,
popula ion, social and economic aspec s o map dep i a ion and po e y. This da a maybe
coupled wi h da a om o icial si es like he wo ld Bank is used o map slums based on social,
economic and habi a /In as uc u e a iables. In [47], an explo a o y ac o analysis o de ine
he Slum Se e i y Index (SSI) Mexico Ci y based on measu ing he shel e dep i a ion le els
o households om 1990 o 2010. The esul s showed ha he SSI dec eased signi ican ly
be ween 1990 and 2000 as a esul o se e al policy e o ms bu inc eased be ween 2000 and
2010. Weeks e al. [48] quan i a i ely used census da a o Acc a, Ghana o c ea e a slum index
based on he UN slum indica o s o measu e he concen a ion o slums. High co ela ions we e
ound be ween he slum index, he socio-economic cha ac e is ics o neighbo hoods and ce ain
land co e me ics de i ed om VHR sa elli e image y. A usion o open sou ces physical and
socio-economic da a in [9] was used o de elop an indica o da abase o cha ac e izing slum
se lemen s by le e aging da a mining echniques o mapping slums in Kenya’s majo ci ies.
I ’s impo an o no e ha one o he majo challenges is ela ed o census unde -co e age due
o ime cons ain , inadequa e quali y assu ance and inaccu a e add esses. The absence o
9
censuses in in o mal se lemen s is la gely due o he inaccessibili y o make-shi s uc u es
due o poli ical in ole ance o gene al diso de [30].
2.3.2. Pa icipa o y-based app oach
The pa icipa o y app oaches in ol e he coope a ion and pa icipa ion o he in o mal
se lemen dwelle s in o de o gene a e bo h spa ial and non-spa ial in o ma ion o p o ile
in o mal se lemen s [30]. The pa icipa o y slum upg ading P og amme (PSUP), es ablished
by he UN-Habi a and he Slum Dwelle s In e na ional (SDI) was a pa icipa o y app oach
ha encou aged and empowe ed communi ies o become ac i e pa ne s wi h s akeholde s in
de ising s a egies o plan sus ainable in o mal se lemen upg ades [49]. The Ko a Tanpa
Kumuh (KOTAKU) pla o m and p og am is esponsible o he handling o in o mali y in
Indonesia h ough inc easing he ole o local go e nmen and communi y pa icipa ion.
In o mali y is de ined by buildings, local s ee s, d inking wa e supply, communi y d ainage,
was ewa e managemen , ga bage managemen , and i e p o ec ion [50]. Howe e ,
Pa icipa o y-based app oaches a e ex emely ime consuming and e o -in ensi e is
cha ac e ized by limi ed spa ial co e age making i di icul o co e la ge a eas like egional
o dis ic le els [30].
2.3.3. GIS and Remo e sensing-based app oach
The connec ion be ween in o mal se lemen s and geog aphy has been a ocus o s udy o a
long ime. Many esea ch s udies ha e in es iga ed he alue o using Image- based
iden i ica ion o in o mal se lemen s h ough emo e sensing and o he GIS analysis ools o
mo e han 2 decades [23,13,48,49, 50]. Th ee key c i e ia a e o en used in such s udies: small
g ain size, high g ound co e age and i egula access ne wo ks. O he ac o s such as p oximi y
o haza dous a eas, lack o ege a ion and low- quali y oo ing a e some imes included. Such
s udies o en con la e in o mali y wi h slums. Bu while mos such s udies map a bina y
dis inc ion be ween o mal and in o mal mo phologies, some adop a mo e nuanced app oach.
B adley e al [54] deploy he use o bo h HR and VHR sa elli e image y o map in o mal
se lemen s in de eloping coun ies using machine lea ning. They used he Canonical
Co ela ion Fo es s (CCFs) o lea n he spec al signal o in o mal se lemen s om HR sa elli e
image y. The second me hod used a CNN combined wi h VHR sa elli e image y o ex ac
ine g ained ea u es. Pe e and Gulnaz s udied he use o objec based change de ec ion and
objec acking o in o mal se lemen s using emo ely sensing da a in Cape own (Sou h

10
A ica) be ween he pe iod 2000 o 2015. The objec change de ec ion p o ided o de ec ion
o hema ic changes pe objec and o documen he changes o hei p ope y alues and
po en ial mo emen . In [52], he au ho s discuss and e alua e a ious machine lea ning
app oaches and a combina ion o a ious ea u es o de ec slums. The classi ica ions me hods
e alua ed included mul i-class and hie a chical o dis inguish u ban om o he classes such as
ege a ion and wa e . [55] p esen ed he use o CNNs o he de ec ion o in o mal se lemen s
om VHR i.e. 0.60m Quickbi d sa elli e image o he ci y o Da es Salaam, Tanzania acqui ed
in 2007 and compa ed he esul s wi h s a e o he a classi ie s ha use handc a ed ea u es.
[56] e alua ed and poin ed ou ha Objec Based Change De ec ion (OBCD) and OBIA a e he
mos p omising echniques o au oma ized iden i ica ion o new buildings in in o mal
se lemen s, bu hey need o wo k wi h VHR da a ec i ied wi h a DSM o he same o
compa ible spa ial esolu ion. [57] This pape aimed o documen and unde s and basic
in as uc u al condi ions and hei changes o e an eigh -yea pe iod o he Kibe a Slum using
VHR sa elli e images by analyzing he dynamics o physical ans o ma ions o building sizes
and heigh s, buil -up densi ies, and building a angemen s. A pionee ing s udy by Ho man
[58] used OBIA o iden i y in o mal se lemen s om IKONOS image y in he Ci y o Cape
Town. In o mal se lemen classi ica ion was unde aken using sub-classes ha desc ibed
se lemen o ms (dense, medium, new and b igh ) based on complex hie a chy and class
desc ip ions such as ex u al and spec al ea u es. The au ho ound ha he abili y o de ec
in o mal se lemen s was dependen on he spa ial esolu ion o he image y. No quan i a i e
esul s we e p esen ed as he indings o he s udy. This esea ch was la e imp o ed by
Ho mann e al. [59], who showed ha se e al modi ica ions we e equi ed when applying
ex ac ion me hods o a Quickbi d scene in B azil. The adap ions included simpli ied and
p uned class-hie a chies o make he chosen class desc ip o s in heo y mo e ans e able o
compa able scenes. The esul s o his s udy demons a ed ha he selec ion o a s a egy o
in o mal se lemen segmen a ion and classi ica ion is da a and con ex -speci ic. Shekha [60]
delinea ed in o mal se lemen s in Pune Ci y, India, using Quickbi d image y. The s udy
highligh ed he e icacy o he de eloped me hodology o disc imina e in o mal egions by
desc ibing ypical cha ac e is ics o hese se lemen s. Fuzzy membe ship unc ion o ex u e,
geome y, and con ex ual in o ma ion we e used o achie e an o e all accu acy o mo e han
87%. Kohli e al. [30] expanded upon he wo k o Ho mann e al. [59] and de eloped Gene ic
Slum On ology (GSO), which can be used as pa o a concep ual classi ica ion OBIA schema.
11
Howe e , echniques like OBIA equi e VHR Da a usually wi h a esolu ion o unde 5 me e s
o de ec slum objec s and is a e y ime consuming app oach and equi es an adop ion o
pa ame e s and alues o a AOI. This makes hem di icul o ans e o o he ci ies o
coun ies. I should be no ed ha mos o he ecen app oaches o iden i y and map in o mal
se lemen s in ol e he usion o wo o mo e app oaches.
2.4. Ho Spo Clus e ing Analysis
Ho spo analysis is a spa ial analysis and mapping echnique in e es ed in he iden i ica ion o
clus e ing o spa ial phenomena in o s a is ically signi ican ho spo s and cold spo s [61].These
spa ial phenomena a e depic ed as poin s in a map and e e o loca ions o e en s o objec s.
A ho spo is an a ea ha has highe concen a ion o e en s compa ed o he expec ed numbe
gi en a andom dis ibu ion o e en s [62]. Ho spo de ec ion has e ol ed om he s udy o
poin dis ibu ions o spa ial a angemen s o poin s in a space [63] o unde s and spa ial
pa e ns in ime. The applica ion o ho spo analysis wi hin public heal h, epidemiological
esea ch and c ime mapping and esea ch has inc eased signi ican ly in he pas couple o
decades mainly due o ad ancemen in GIS-based so wa e. Ho spo analysis usually in ol es
he inciden coun o poin s in a loca ion, A ibu es ha u he desc ibes poin s o Pe iod o
ime i.e. Da e o ime o e en s. Ve y ew s udies a e based on he use o ho spo analysis in
he mapping o In o mal se lemen s because i equi es mos ly ec o da a which is qui e
insu icien o slums compa ed o as e da a. Mohamed e . al in oduced a me hod o
iden i ying and p edic ing in o mal se lemen s using s ee in e sec ions da a using Ge is-O d
Gi Ho Spo Analysis and a machine lea ning app oach[20].
2.5. An o e iew o Machine Lea ning Techniques
Machine lea ning echniques (MLTs) (Figu e 2) gene a e knowledge in a lea ning phase by
means o aining da a and ans e his o new da a o p edic ion and gene aliza ion e en wi h
noise-con amina ed and incomple e da a [58,59].
MLTs ha e been success ully used in analyzing and modeling a ious complex en i onmen al
disciplines, including in medicine, inancial ma ke s, ecology, geog aphy, biomedicine, and
epidemiology [66–69]. The g owing popula i y o MLTs can be a ibu ed o hei abili ies o
app oxima e almos any complex non-linea unc ional ela ionship [70–72]. Machine
12
Lea ning algo i hms a e based on bo h supe ised and unsupe ised lea ning. Supe ised
Lea ning wo ks wi h labeled inpu da a, known as aining da a in a aining p ocess o
p edic ing unknown da a. Based on he known da a, he aining p ocess he eby con inues un il
he model achie es a desi ed le el o accu acy. Once he model is ained i can be used o
ans e his knowledge on new da a by gene alizing om he aining da a o unknown
examples. Supe ised lea ning p oblems can be u he g ouped in o classi ica ion and
eg ession p oblems. On he o he hand, o Unsupe ised Lea ning, he inpu da a is no
labeled and does no ha e a known esul . The model ies o ecognize a pa e n in he inpu
da a o lea n abou i and ex ac ules. This can be done h ough a ma hema ical p ocess o
sys ema ically educe edundancy o by o ganizing he da a by simila i y like h ough
Clus e ing and Associa ion.
Figu e 2: Machine Lea ning Chea Shee [73]
Some o he mos common ML algo i hms in Geospa ial Sciences and Mapping o In o mal
se lemen s a e Logis ical Reg ession, Decision T ees (DT), sel -o ganizing maps (SOM),
andom o es s (RF), suppo ec o machines (SVM) and a i icial neu al ne wo ks (ANN).
I ’s impo an o no e ha selec ing he bes algo i hm o a ask is qui e challenging since no
algo i hm ou pe o ms o he s in a gi en ask. Machine lea ning echniques ha e been and a e
s ill being used o ca y ou geospa ial analysis in a ious ields o clus e ing, classi ica ion
13
and p edic ion a ious en i onmen al and socioeconomic phenomena using di e en ypes o
da a.
2.5.1. Con olu ional neu al-ne wo ks
Con olu ional neu al-ne wo ks (CNNs) a e a b anch o a i icial neu al ne wo ks designed o
ecognize objec s in images based on hei abili y o de elop an in e nal ep esen a ion o a
wo-dimensional image [74]. The con olu ional neu al ne wo k is composed mainly o h ee
ypes o laye s i.e. Con olu ional Laye s, Pooling Laye s, Fully-Connec ed Laye s. The
Con olu ional laye s a e comp ised o il e s which a e he neu ons o he laye and a ea u e
map which is he ou pu o one il e applied o he p e ious laye . The pooling laye s down-
sample he p e ious laye s ea u e map by ollowing a sequence o one o mo e con olu ional
laye s wi h he in en ion o consolida e he ea u es lea ned and exp essed in hem. They a e
used o educe he size o he da a while main aining he mos impo an ea u es. Fully
connec ed laye s a e he no mal la eed- o wa d neu al ne wo k laye used a he end o he
ne wo k a e ea u e ex ac ion and consolida ion has been pe o med by he con olu ional
and pooling laye s, o c ea e inal non-linea combina ions o ea u es and o making
p edic ions by he ne wo k [75]. CNNs a e well sui ed o he ask o image classi ica ion
in he ield o emo e sensing, Howe e , CNNs also pose challenges when applied o
emo e sensing as i is o en di icul o imp ac ical o ob ain a la ge se o labelled
images, he model pe o ms poo ly and ends o o e i [74]. Such challenges can be cu bed
using echniques like Da a augmen a ion using o a ions and lips o inc ease he o al numbe
o aining images o u he a oid o e i ing p oblem [33].
20
In o mal se lemen s, hough mos in o mal se lemen buildings a e smalle han his spa ial
esolu ion o 10m*10m. The images o he da es in Table 2 we e selec ed based on pe cei ed
isually simila i y, quali y and minimal cloud co e .
Acquisi ion Da e
Bands Used
01-01-2016 4,3,2
23-08-2017 4,3,2
31-12-2019 4,3,2
Table 2: De ails o Sen inel-2 sa elli e images
3.2.3. Re e ence da a
The map selec ed o alida e egions mapped ou as in o mal egions in Kampala (Figu e 7)
was p epa ed by AcToge he (Ug) in 2014, does no include he slum se lemen s which ha e
ecen ly come up in he ci y. Howe e , i p o ides he loca ions o he slum se lemen s
manually delinea ed h ough Fields su eys and Public Pa icipa ion. The slums in Kampala
ci y a e no clus e ed a speci ic loca ions bu ha e sca e ed p esence all o e he ci y wi h
a ying sizes and shapes. The se lemen s a e loca ed along he main oads in elonga ed and
i egula shapes, along ailway lines and nea he lake. Un o una ely, ecen In o mali y maps
o Da -es Salaam we e di icul o come by.
Figu e 7: Dis ibu ion o In o mal Se lemen s in Kampala based on A e age Household Size
[97]

21
4. RESEARCH METHODOLOGY
The me hod consis s o h ee main phases (Figu e 8). Fi s phase consis s o en iching he
building ou line ea u es wi h geome ic and opologic pa ame e s,
The second phase en ails ho spo clus e ing and classi ica ion o hese cha ac e is ics o iden i y
in o mal se lemen s in simila con ex ci ies, and he inal phase in ol es spa io- empo al
moni o ing o in o mal se lemen s using sen inel-2 image y.
Figu e 8: Me hodology Flowcha
4.1. Da a En ichmen and Compu a ion o Pa ame e s
Be o e ca ying he clus e ing and classi ica ion p ocess, geome ic and opological measu es
o allow g ouping o ea u es in o di e en clus e s we e de e mined. The measu es chosen o
en iching he building ea u es we e size, shape and sho es dis ance o a neighbo ing building.
A building’s shape was desc ibed by he numbe o e ices i has i.e. e ex numbe (VN).
Mo e complex buildings ha e many e ices and is e sa. The size was desc ibed by
calcula ing i s a ea (A) [98,99]. The Nea es Neighbo Dis ance (NND) and was used o
desc ibe he opological ela ions o neighbo ing ea u es [100]. Buildings in in o mal
22
se lemen s end o be e y close o each o he wi h limi ed gap be ween hem. They a e also
small in size and ha e simple shapes like o ec angles o squa es houses. The no maliza ion
o compu ed a iables was ca ied ou since hey ha e a ying scales and o a oid in luence o
some a iables o e o he s du ing use in he machine lea ning algo i hms. The Min-max
no maliza ion app oach was used whe e all he alues we e mapped be ween [0–1].
𝑿󰆒= (𝑿𝑿𝐦𝐢𝐧)
(𝑿𝐦𝐚𝐱𝑿𝐦𝐢𝐧) (Min-Max No maliza ion Equa ion)
Whe e, X is an o iginal alue, X’ is he no malized alue, and Xmin, Xmax a e he minimum
and maximum alues o his pa icula p ope y.
4.2. In o mal Se lemen Buildings Clus e ing and Classi ica ion
In he i s s ep, clus e s a e de e mined based on simila i y o building ea u e cha ac e is ics
using ho spo clus e ing analysis. The clus e s a e hen eclassi ied o educe he numbe o
clus e s. I a building belongs o he same clus e based on he ea u es cha ac e is ics, i s
binned in one clus e and is e sa. This is done on he Kampala aining pa i ion. The second
s ep in ol es using di e en classi ie s o selec one ha sui s he aining pa i ion da a. The
selec ed classi ie is hem used o de e mine in o mali y egions in he Kampala Tes ing
Pa i ion and Da -es-Salaam da ase .
4.2.1. Clus e ing o In o mal Se lemen Buildings
As shown in chap e 2, in o mal se lemen a eas ha e building ea u es cha ac e ised by simple
shapes, small buildings and ha e limi ed gaps be ween hem.
In o de o de e mine he pa e n o spa ial dis ibu ion and whe e low and high alues o he
building ea u e cha ac e is ics a e g ouped, ho spo analysis based on he use o Ge is-O d Gi*
s a is ic was ca ied ou on he Kampala aining pa i ion da ase o iden i y s a is ically
signi ican ho and cold spo s [61]. Since ho spo analysis equi es he p esence o clus e ing,
he e is need o es o he p esence o clus e ing in he da ase by assessing spa ial
au oco ela ion o iden i y clus e ing wi hin he en i e da ase . The Spa ial Au oco ela ion
(Mo an’s) ool in A cGIS measu es spa ial au oco ela ion by simul aneously measu ing
ea u e loca ions and a ibu e alues. I ea u es ha a e close oge he ha e simila alues,
23
hen ha is said o be clus e ing, and is e sa. The Mo an’s I e u ns alues which include he
z-sco e and p- alue which will indica e i clus e ing is ound in he da a o no .
The Op imized Ho spo Analysis (OHSA) ool om A cGIS P o used o he ho spo analysis
is iden i ies s a is ically signi ican ho and cold spo s o mul iple es ing and spa ial
dependence using he using he False Disco e y Ra e (FDR) co ec ion me hod [101]. The
ea u es a e g ouped in o se en majo clus e s i.e. ea u es in he +/- 3 Gi_Bins we e s a is ically
significan a he 99% confidence le el; ea u es wi h 0 o he Gi_Bin field was no s a is ically
significan , hose in he +/2 bins eflec ed a 95% confidence le el, and ea u es in he +/1 bins
eflec ed a 90% confidence le el. Upon compu ing hese GiZsco es whe e cold spo s in all he
ea u e cha ac e is ics mos likely ep esen in o mal se lemen buildings, a c oss alida ion
es using Independen samples - es on he ho spo analysis esul s in o de o di e en ia e
be ween o mal and in o mal zones and o de e mine whe he he e is s a is ical e idence ha
he associa ed g oup means a e signi ican ly di e en be ween hem [102]. The GiZsco es o
bo h g oups; in o mal, and o mal egions a e compa ed and when he p- alue is less han
0.001, he null hypo hesis can be ejec ed and he esul s o GiZsco e can be illus a ed as
s a is ically signi ican [11].
The ea u es iden i ied as cold spo s we e hen eclassi ied and g ouped in o one class and all
he o he s as ano he class, which p oduces new eclassi ied esul s o he ea u e
cha ac e is ics o Size, shape and NND. The h ee eclassi ied esul s a e combined and i a
ea u e has all i s ea u e cha ac e is ics as cold spo s, i s g ouped as in o mal and i no , i s
clus e ed as o mal. This p oduces one esul an clus e da ase o he building ea u es which
shows bo h in o mali y and o mali y in he Kampala aining pa i ion da ase . These esul s
(especially he cold spo s) we e compa ed o a eas delinea ed as slums in Kampala gene a ed
by a collabo a ion be ween Slum Dwelle s In e na ional (SDI) and o he o ganiza ions [46, 89].
4.2.2. Classi ica ion o In o mal Se lemen s based on Building Cha ac e is ics
To p edic whe e in o mal se lemen egions a e based on building ea u es cha ac e is ics o
shape, size and NND, based on he clus e s om he aining pa i ion da ase , supe ised
machine lea ning classi ica ion was ca ied ou on da a o he same ci y (Kampala Tes ing
Pa i ion) and on da a o a di e en ci y (Da -es Salaam). Fi s ly, he clus e ing esul s we e
24
applied o ain, alida e and es se e al classi ie algo i hms in o de o selec one ha sui s
hese da ase s. Models we e buil using he py hon Sciki -lea n lib a y, aking ad an age o he
capabili ies and ea u es o his lib a y. These models included Logis ical Reg ession, Decision
T ee (DT), Mul ilaye Pe cep on (MLP), Random Fo es (RF) and Suppo Vec o Machine
(SVM). The choice o hese classi ie s in based on hei use in p e ious s udies o classi ie
buildings based on hei cha ac e is ics and also hei abili y o p o ide good accu acy wi h
e e ence o chap e 2. The sc ip in ol ed spli ing he T aining pa i ion da ase in o 70% o
aining and 30% alida ion. Classifica ion was pe o med i s on he aining da a and
accu acy measu es, including o e all accu acy, p ecision, ecall and -measu e we e no ed.
This was ca ied ou in o de o selec he bes classi ie o he da ase s. The selec ed classi ie
is hen used o p edic which ea u es belong o In o mal se lemen s in Kampala Tes ing
Pa i ion and Da -es Salaam.
4.3. Moni o ing he g ow h o In o mal Se lemen s
The sen inel- 2 image y a e p e-p ocessing and band composi ion we e clipped o he
Kampala Bounda y. The acqui ed in o mal se lemen egions om he classi ica ion phase
abo e p o ides o polygons ep esen ing in o mali y. We also assembled polygons o egions
ep esen ing h ee ca ego ies: “buil (wi hou in o mal)”, “ ege a ion”, and “wa e ”. All hese
egions we e gene a ed on he 2019 Image (Figu e 9). This is based on he assump ion ha he
OSM building ou lines used o classi y in o mali y we e as up- o-da e as o Oc obe 2019. All
he images we e hen sliced up in o equal-sized iles o 255*255 m o ensu e uni o mi y in he
da ase s and placed in olde s o “Tiles_2019”, “Tiles_2017” and “Tiles_2016” each wi h
5427 iles. The iles we e sa ed as i iles wi h a ached geo-in o ma ion which is equi ed o
plo he model p edic ions o he o iginal loca ions in he map.
25
Figu e 9: In o mal and Fo mal Tiles
The iles in olde Tile_ 2019 we e labelled based on he ou ca ego ies c ea ed based on hei
in e sec ion wi h he c ea ed polygons. All he iles belonging o “buil -up”, “ ege a ion”, and
“wa e ” we e g ouped in o one olde as Fo mal and he In o mal iles placed in an In o mal
Folde . Reg ouping he ou ca ego ies in o wo was mo e accu a e han aining a bina y
classi ie di ec ly om he a ailable da a. The iles we e hen spli in o aining (80%) and
es ing se s (20%) o model- alida ion o use in he CNN model.
A baseline CNN model based on gene al a chi ec u al p inciples o he Visual Geome y G oup
(VGG) on which o he models can be compa ed is es ablished. The CNN a chi ec u e in ol ed
s acking con olu ional laye s wi h small 3×3 il e s ollowed by a max pooling laye . These
laye s o m a block, and hese blocks can be epea ed whe e he numbe o il e s in each block
is inc eased wi h he dep h o he ne wo k such as 32, 64, 128, 256 o he i s ou blocks o
he model. Padding is used on he con olu ional laye s o ensu e he heigh and wid h shapes
o he ou pu ea u e maps ma ches he inpu s. This unc ion can hen be cus omized o de ine
di e en baseline models, e.g. e sions o he model wi h 1, 2, o 3 VGG s yle blocks. The
model was i ed wi h s ochas ic g adien descen , a conse a i e lea ning a e o 0.001 and a
momen um o 0.9. Since he p oblem was a bina y classi ica ion ask i.e Fo mal and In o mal,

26
he p edic ion o one alue o ei he 0 o 1 was equi ed. An ou pu laye wi h 1 node and a
sigmoid ac i a ion was used and he model op imized using he bina y c oss-en opy loss
unc ion. One-block (32 il e s), wo-block (32 and 64 il e s) and h ee-block (32,64 and 128
il e s) VGG models we e all compa ed on he da ase . Two app oaches o add ess o e i ing
o he aining da ase : d opou egula iza ion and da a augmen a ion we e explo ed o selec
he bes sui ed one. Upon selec ing he bes model wi h he adequa e model pa ame e s,
P edic ions we e pe o med on he Independen ile da ase s o “Tiles_2017” and
“Tiles_2016” o au oma ically p edic in o mal se lemen iles ou o all he iles. A polygon
laye was hen buil o he new in o mali y iles. This p edic ion s ep leads o iles ha a e
ei he o mal o no . Polygons o he image ile Bounda ies a e c ea ed and hose ha belong
o in o mal g oup a e o e laid on o he map o isual in e p e a ion.
27
5. RESULTS
This chap e desc ibes he implemen a ion de ails o he me hodology and p esen s he esul s
o his s udy. The sou ce code o his implemen a ion can be ound a
h ps://gi hub.com/Bakked9/Mas e sThesis.
5.1. Clus e ing Analysis
5.1.1. Clus e ing Indices o Building Cha ac e is ics
The Mo an's Indices o he A ea, shape and NND a e all close o 0.2, and ha e p- alues o 0.0
(Table 3). The z-sco e is a s anda d de ia ion which measu es how many SDs away an elemen
is om he mean, while he p- alue is a measu e o p obabili y ha a andom p ocess c ea ed
he obse ed pa e n. A small p- alue is an indica o ha he spa ial pa e n is no andom and
is e sa. A posi i e Mo an’s I esul indica es ha neighbo ing a eas o poin s a e simila wi h
espec o a ibu e alues, while nega i e Mo an’s I shows ha nea by egions a e less simila
in a ibu es han one would expec in a andom pa e n [103]. This goes o show ha clus e ing
exis s wi hin he da ase wi h espec o he measu emen a iables and hence Ho spo analysis
can be ca ied ou on he da a.
Va iable Mo an's Index Expec ed Index Va iance z-sco e p- alue
Nea es Dis ance
0.257656
-0.000017 0.000139 21.848243 0.000000
A ea 0.278887 -0.000017 0.000139 23.696816 0.000000
Shape 1.336825 -0.000017 0.000139 113.285603
0.000000
Table 3: Spa ial Au oco ela ion (Mo an's I) Resul s
5.1.2. Op imized Ho spo Analysis
The op imized ho spo analysis esul ed in o Gi_Bin fields which iden ified he s a is ically
significan ho spo s, along wi h he non-significan and cold spo s, indica ing he ype o
clus e s o he ea u e cha ac e is ics. In he case o he T aining pa i ion, he Gi_Bin field
showed ha pa ches sp ead a ound he selec ed a ea was cha ac e ized by cold spo s, non-
significan ypes, and ho spo s ypes ha a e dis ibu ed allo e (Figu e 10). The a eas
iden i ied as non-significan clus e ypes p o ide some so o bounda y be ween he cold and
28
ho spo s. Whe eas he cold spo s we e loca ed in a small a ea mainly on he no h and cen al
pa s, he ho spo clus e s we e sp ead ou in on he no h, sou h, eas and wes sides.
Figu e 10: OHSA Resul s Maps o Size, Shape and NND
5.1.3. The Associa ion be ween he Cold spo s and In o mal se lemen s
Table 4 illus a es he esul s o he - es analysis, using a 95% con idence le el o each ea u e
cha ac e is ic. Since he Sig (2-Tailed) alues i.e. he p- alue o he GiZsco e and Nneighbou s
p edic o s a e e y small i.e. p<0.001, we ejec he null hypo hesis o Le ene's es and
conclude ha he a iances in he GiZsco e and Nneighbou s p edic o s o Fo mal is
signi ican ly di e en om ha o In o mal. The In o mal g oups mean numbe s o neighbo s
a e la ge han hose o he o mal a eas. The mean GiZsco es a e he in o mal a eas a e also
smalle han hose o he o mal a eas. This shows ha he in o mal egions a e mo e likely o
be cold spo s wi h low alues o Shape, size and NND.
29
SIZE(AREA)
G oup
N
Mean
S d.
De ia ion
S d. E o
Mean
Sig.(2
-
ailed)
GiZSco e
ColdSpo (In o mal) 104884 -4.4855 2.0227 0.0062 0.000
Ho Spo (Fo mal) 36868 3.6859 1.8595 0.0097
Nneighbo s ColdSpo (In o mal) 104884 958.00 302.407 .934 0.000
Ho Spo (Fo mal) 36868 363.13 129.434 .674
NEAREST NEIGHBOUR DISTANCE (NND)
G oup
N
Mean
S d.
De ia ion
S d. E o
Mean
Sig.(2
-
ailed)
GiZSco e
ColdSpo (In o mal) 123467 -5.6911 2.8813 0.0082 0.000
Ho Spo (Fo mal) 41942 5.1848 3.3074 0.0161
Nneighbo s
ColdSpo (In o mal) 123467 936.37 281.848 .802 0.000
Ho Spo (Fo mal) 41942 344.97 119.569 .584
SHAPE (VERTEX NUMBER)
G oup
N
Mean
S d.
De ia ion
S d. E o
Mean
Sig.(2
-
ailed)
GiZSco e
ColdSpo (In o mal) 129377 -4.8803 1.9626 0.0055 0.000
Ho Spo (Fo mal) 34552 5.8729 4.6398 0.0250
Nneighbo s ColdSpo (In o mal) 129377 850.26 319.712 .889 0.000
Ho Spo (Fo mal) 34552 512.44 286.337 1.540
Table 4: Independen samples - es analysis Resul s
Cold spo s in his analysis ep esen a eas ha ha e low alues o A ea, Shape and NND.
In o mal se lemen s a e cha ac e ised by buildings ha ha e simple shape, small is size and
a e densely loca ed in a pa icula loca ion i.e. sho NNDs be ween polygons (see igu e 10).
Wi h his, he cold spo s esul ing om he OHSA a e conside ed o be loca ions wi h
in o mali y based one he building ypologies used o cha ac e ise hem.
To simpli y he da a o u he analysis, all he OHSA esul s we e classi ied o bin he da a
poin s in o wo classes i.e. “In o mal” and “Fo mal”. This was done h ough eclassi ica ion o
he Gi_bin esul s wi h a h eshold o -2 o lowe (95% cold spo con idence le el) as
In o mali y and he es as o mali y. As shown in igu e 11 and 12, he egions ha iden i ied
as In o mal sha e common loca ions ac oss he map wi h espec o he building ypologies.
The esul s we e combined o highligh only hose a eas ha had been classi ied as in o mali y
in all he ypologies. The esul an da ase was hen used as a aining se o he Classi ica ion
s ep.
36
Figu e 16: Change in In o mal Se lemen Regions in Kampala. This shows In o mal iles ha exis ed in
Janua y 2016, new in o mal iles ha had de eloped by Augus 2017 and also by Decembe 2019.

37
6. DISCUSSION
6.1. Discussion
This s udy p esen ed app oaches o classi y in o mali y egions based on building ou line
cha ac e is ics and moni o ing g ow h o in o mal se lemen s o e ime by execu ing CNNs
on High esolu ion sa elli e image y. Fi s ly, we highligh ha OSM building ea u e
cha ac e is ics i.e. size and shape o building ou line, and dis ance o nea es building can be
used o di e en ia e be ween o mal and in o mal buildings. In he app oach, he geome ic
and opologic measu es p esen ed in his esea ch we e clus e ed based on he alues high o
low. As an impo an aspec o he e alua ion, he ans e abili y o he ob ained clus e s om
aining se o es da a was in es iga ed. Fi s ly, he esul s we e applied o he es ing da a o
he same ci y and hen o ano he ci y. 108,390 o 429,270 buildings (25.25%) we e classi ied
as in o mal se lemen buildings in Kampala while 188,728 o 971,008 buildings (19.44%) in
Da -es Salaam we e classi ied as in o mal se lemen buildings (Figu e 13). The e was a
dec ease in Random Fo es classi ie model pe o mance on he Kampala Tes Pa i ion and
Da -es Salaam may be due o di e ences in he cha ac e is ics o buildings in e y u ban a eas,
o he ac ha mo e da a we e a ailable in Da -es -Salaam. I ’s impo an o no e ha he
accu acy alues we e consis en ly abo e 0.75, sugges ing e y good disc imina o y
pe o mance i espec i e o he ci y. The a ia ion in building ypologies used du ing his
analysis (shape, size and spacing) played an impo an ole du ing he classi ica ion o
in o mali y and o mali y e en hough hese cha ac e is ics a y acco ding o geog aphical
loca ion [93]. Though he mo phological na u e o in o mal se lemen s a ies ac oss he
global, hey s ill ba e some simila i ies in cha ac e is ics. This could be he eason why he
same model could be applied in a di e en ci y and p oduce dis inc esul s. Compa ed o i s
a eal co e age, he e a e ewe in o mal se lemen buildings in Da es Salaam. O e he yea s,
a dec ease in he popula ion li ing in Slums in Tanzania has been dec easing wi h a sha p
dec ease obse ed om 2011 e en wi h apid popula ion g ow h o e ime (Figu e 17). This is
p obably due o s a egies by he Tanzania go e nmen o shi om egula ly demolished
homes in in o mal se lemen s o awa ding i les could be one o he easons o a dec ease in
in o mali y in he ci y. The MKURABITA P og amme i s launched in Da es Salaam, aims
a ans o ming p ope y and businesses held in he in o mal sec o in o legal en i ies ha a e
oo ed i mly in he o mal sec o [104].
38
Figu e 17: Tanzania - Popula ion Li ing In Slums (% O U ban Popula ion) [[105]]
The expe imen a ions o Using CNN on Sen inel-2 e alua es shows p omising esul s on
disc imina ing In o mal om o mal se lemen s o e ime. Howe e , Seasonal a ia ions
be ween he images p e en ed he deep lea ning model om ecognizing ce ain In o mal
Se lemen s. Tes ing he model on new mul i- empo al da ase s causes he model pe o mance
o dec ease as we mo e away om he aining se ’s sensing ime. The pe o mance o he
model is bound o deg ade as we mo e u he away om he aining pe iod. Upon compa ing
he de elopmen o he slums in Kampala ci y using he ime slide in Google Ea h om he
yea 2014 o he da es a which sa elli e images a e downloaded, change obse a ions can be
made. I can be seen ha since 2014, he buil densi y o slums has inc eased a p e-exis ing
loca ions and smalle new pa ches ha e eme ged a subu ban a eas o he ci y. Compa ed o
he la ge a eas o al eady exis ing slums, hese new se lemen s a e qui e smalle and hence
could be igno ed du ing p edic ion on he image iles due o mixed classes wi h ege a ion and
o mal a eas.
To e alua e he esul s ob ained, he p edic ion Accu acy ob ained om he Baseline VGG3 +
Da a Augmen a ion CNN Model (90.3%) on he sen inel-2 image y is a e y good esul s
based on he spa ial esolu ion o he images. In [106], Fede ico .B and Damián .S also de ec
in o mal se lemen s using sen inel-2A image iles. They employed he use o Remo e Sensing
Classi ie o P edic iles ha belong o in o mal se lemen s. They ecommend he use o deep
lea ning echniques, like con olu ional neu al ne wo ks o imp o e he accu acy o he
classi ica ion. This indica es ha he CNN p o ides be e classi ica ion accu acy especially
39
because i ’s an image based classi ie . On ano he no e, he use o building ou line
cha ac e is ics like size, shape and Nea es Neighbou dis ance p o ides o a lo mo e insigh
on he na u e and mo phology o an in o mal se lemen .
In es iga ions on p edic ing in o mal se lemen s using oad in e sec ions om OSM da a [20]
showed ha OSM da a can be used in his scena io. Howe e , some o he limi a ions o his
s udy is ha some in o mal se lemen s ha e oads ha a e no mapped on OSM which makes
hei de ec ion di icul . This is whe e he use o Building ea u e cha ac e is ics, as shown in
his esea ch a e use ul. The cha ac e is ics help be e delinea e in o mal se lemen s. An e en
be e al e na i e would be he combina ion o building ea u e cha ac e is ics and densi y o
oad in e sec ions o iden i y hese in o mal se lemen s. The use o Mul ilaye Pe cep on
(MLP) on ho spo analysis GiZsco e and Nneighbou s o p edic in o mal se lemen s p o ed
a good enough classi ie , howe e in his esea ch, hough MLP pe o med well, Random
Fo es Classi ie was a be e classi ie o Building ea u e cha ac e is ics o p edic
in o mali y.
6.2. Limi a ions
While he s udy p oduced good esea ch indings, he e a e se e al limi a ions and
oppo uni ies o u he esea ch Addi ional ec o da a o land pa cels, small adminis a i e
dis ic uni s, ci y s unc ional zones, o land use da a could be in eg a ed o imp o e he
c ea ion o desi ed op imal shapes o image objec s o compac homogeneous buil -up a eas
o buil -up densi y analysis. I a ailable in accu a e up- o-da e o m, ec o laye o building
oo p in s could be used o buil -up densi y calcula ion, wi hou he need o land co e
classi ica ion and building ex ac ion om he sa elli e image da a. This laye would ep esen
only buildings, wi hou oads o o he man-made u ban s uc u es. Howe e , his da a is no
always eely a ailable in accu a e and up- o-da e o m, and his app oach akes ha in o
conside a ion. The lack o adequa e da a o es he model on o he ci ies besides he wo
p esen ed was expe ienced. This model wo ked o hese wo ci ies because he e bo h ha
OSM da a almos accu a ely a ailable. This is no he case o mos o he ci ies in he global
sou h ha a e engul ed by in o mali y.
40
7. CONCLUSION AND FUTURE WORKS
The main esea ch Aim o his esea ch wo k was o in eg a e OSM da a and sen inel-2
image y o classi ying and moni o ing he g ow h o in o mal se lemen s. To achie e his,
he esea ch ques ions we e in es iga ed:
1. How can we exploi he po en ial o using buildings ou line cha ac e is ics such as size
and shape o di e en ia e In o mal se lemen s om Fo mal Se lemen s?
The esul s om he ho spo clus e ing analysis demons a ed ha in o mal se lemen s
can be di e en ia ed om o mal a eas based on he use o Building ea u e
cha ac e is ics ha is shape, size and NND. The in o mal se lemen buildings a e
cha ac e ized by low alues (cold spo s) o each cha ac e is ic.
2. Is i possible o p edic in o mal a eas in a ci y by unde s anding housing in o mali y in
o he ci ies o simila con ex using buildings ou line cha ac e is ics and machine
lea ning?
Also i s obse ed ha hough mo phologically di e en , slums sha e some
cha ac e is ics like he na u e o building ou lines. This is shown by he ans e abili y
o he classi ie ained on one da ase o p edic in o mal se lemen s in new da ase s
in he same ci y and in a di e en ci y.
3. Wha is he mos app op ia e Machine Lea ning echnique based on accu acy o p edic
in o mal a eas in a ci y based on buildings ou line cha ac e is ics?
The mos app op ia e ML echnique is he Random Fo es classi ie . I p o ed i s
e ec i eness in a b oad ange o applica ions o Geospa ial echnologies. As a non-
pa ame ic me hod, he key p ope y o Random Fo es is i s capabili y o handle
di e en s a is ical dis ibu ions o ea u es, which was one o he main challenges o
his s udy. This make Random Fo es a sui able algo i hm o ou bina y classi ica ion
and p edic ion p oblem.
4. How can we exploi he po en ial o eely a ailable Sen inel-2 sa elli e image y wi h
ad anced machine lea ning o es ima e he g ow h o In o mal se lemen s?
41
Sen inel 2 images, despi e ha ing a spa ial esolu ion o 10m, can be used o in o mal
se lemen mapping when coupled wi h Ad anced Machine Lea ning Algo i hms like
CNNs.
We sugges ha his esea ch a ea could be explo ed mo e and u u e esea ch could be done,
building and imp o ing on he esul s wi h emphasis on a be e accu a e ep esen a ion o
In o mal se lemen s. Fu he mo e, o he OSM ea u es and pa ame e s such as Roads and
accessibili y wi hin he in o mal se lemen s o o he s could be used o be e iden i y whe e
he in o mali y egions could be. This is because In o mal se lemen s a e usually cha ac e ized
by na ow and sho oad segmen s his nume ous dangles an he end o he oads. O he
pa ame e s like p oximi y o haza dous a eas like we land and sewe age channels should be
conside ed. This howe e , would howe e equi e much mo e de ailed and complex land co e
classi ica ion scheme as a sou ce o in o ma ion o his analysis, including iden i ica ion o ,
ege a ion ypes, di e ence be ween ag icul u al ba e soil and ba e soil a eas in u ban a ea o
a cons uc ion si es and so on. I a ailable, Digi al Su ace Model (DSM) could be used o mo e
accu a ely iden i y and map in o mal se lemen s because inclusion o he hi d dimension o
Ele a ion would help highligh mo e dis inguishing cha ac e is ics o in o mal se lemen
buildings.
Ne e heless, he con ibu ion o he empo al dimension o in o mal se lemen mapping
in es iga ed in his esea ch equi es u he es ing wi h la ge da ase s and longe ime se ies.
Wi h his mind, we hink he posi i e esul s achie ed in his esea ch a e wo h building upon
in u u e s udies o u he in es iga e he empo al domain as an inpu o in o mal se lemen
mapping and classi ica ion models.

42
BIBLIOGRAPHIC REFERENCES
1. A ica’s U ban T ans o ma ion - Essays | U ban Age [In e ne ]. [ci ed 2020 Feb 6].
A ailable om: h ps://u banage.lseci ies.ne /essays/a ica-s-u ban- ans o ma ion
2. 68% o he wo ld popula ion p ojec ed o li e in u ban a eas by 2050, says UN | UN
DESA | Uni ed Na ions Depa men o Economic and Social A ai s [In e ne ]. [ci ed
2020 Feb 6]. A ailable om:
h ps://www.un.o g/de elopmen /desa/en/news/popula ion/2018- e ision-o -wo ld-
u baniza ion-p ospec s.h ml
3. Da es Salaam is g owing apidly—and i s planne s a e s uggling o keep up
[In e ne ]. [ci ed 2020 Feb 6]. A ailable om:
h ps://www.na ionalgeog aphic.com/en i onmen /2019/04/ anzanian-ci y-may-soon-
be-one-o - he-wo lds-mos -populous/
4. Abou Slum Upg ading | Ci ies Alliance [In e ne ]. [ci ed 2020 Feb 6]. A ailable om:
h ps://www.ci iesalliance.o g/abou -slum-upg ading
5. In oduc ion o u ban po e y | In e na ional Ins i u e o En i onmen and
De elopmen [In e ne ]. [ci ed 2020 Feb 6]. A ailable om:
h ps://www.iied.o g/in oduc ion-u ban-po e y
6. Roy D, Lees MH, Pala alli B, P e e K, Sloo MAP. The eme gence o slums: A
con empo a y iew on simula ion models. En i on Model So w [In e ne ].
2014;59(2014):76–90. A ailable om: h p://dx.doi.o g/10.1016/j.en so .2014.05.004
7. File:Ae ial iew o Makoko Slum in Lagos Nige ia.png - Wikimedia Commons
[In e ne ]. [ci ed 2020 Feb 6]. A ailable om:
h ps://commons.wikimedia.o g/wiki/File:Ae ial_ iew_o _Makoko_Slum_in_Lagos_
Nige ia.png
8. The 20 Wo s Slums in A ica [In e ne ]. [ci ed 2020 Feb 6]. A ailable om:
h p://www.a ica anking.com/20-wo s -slums-in-a ica/6
9. Mahabi R, Agou is P, S e anidis A, C oi o u A, C ooks AT. De ec ing and mapping
slums using open da a: a case s udy in Kenya. In J Digi Ea h [In e ne ]. 2018;0(0):1–
25. A ailable om: h ps://doi.o g/10.1080/17538947.2018.1554010
10. Suha ini N, Jones P. U ban Go e nance and In o mal Se lemen s [In e ne ]. 2019.
235 p. A ailable om: h ps://www.sp inge .com/gp/book/9783030060930
43
11. Ho mann P, Taubenböck H, We hmann C. Moni o ing and modelling o in o mal
se lemen s - A e iew on ecen de elopmen s and challenges. 2015 J U ban Remo e
Sens E en , JURSE 2015. 2015;1–4.
12. A imah BC. Slums as exp ession o social exclusion: Explaining he p e alence o
slums in A ican coun ies. Uni ed Na ions Hum Se lemen s P og am [In e ne ].
2001;1–33. A ailable om: h p://www.oecd.o g/de /pgd/46837274.pd
13. Kohli D, Sliuzas R, S ein A. U ban slum de ec ion using ex u e and spa ial me ics
de i ed om sa elli e image y. J Spa Sci [In e ne ]. 2016;61(2):405–26. A ailable
om: h p://dx.doi.o g/10.1080/14498596.2016.1138247
14. Kohli D, S ein A, Sliuzas R. Unce ain y analysis o image in e p e a ions o u ban
slums. Compu En i on U ban Sys [In e ne ]. 2016;60:37–49. A ailable om:
h p://dx.doi.o g/10.1016/j.compen u bsys.2016.07.010
15. Wu m M, Taubenböck H, Weigand M, Schmi A. Slum mapping in pola ime ic SAR
da a using spa ial ea u es. Vol. 194, Remo e Sensing o En i onmen . 2017. 190–204
p.
16. S asolla M, Gamba P. Exploi ing Spa ial Pa e ns o In o mal Se lemen De ec ion in
A id En i onmen s. In A ch Pho og amm Remo e Sens Spa In Sci. 2007;36:31–6.
17. Ranguelo a E, Weel B, Roy D, Ku e M, P e e K, Lees M. Image based
classi ica ion o slums, buil -up and non-buil -up a eas in Kalyan and Bangalo e, India.
Eu J Remo e Sens [In e ne ]. 2019;52(sup1):40–61. A ailable om:
h ps://doi.o g/10.1080/22797254.2018.1535838
18. De ec ing In o mal Se lemen s using Sa elli e Image y and Con olu ional Neu al
Ne wo ks [In e ne ]. [ci ed 2020 Feb 6]. A ailable om:
h ps://blog.goodaudience.com/de ec ing-in o mal-se lemen s-using-sa elli e-image y-
and-con olu ional-neu al-ne wo ks-d571a819b 44
19. Jean N, Xie M, E mon S, Ko ea N. Combining Sa elli e Image y and Machine
Lea ning o P edic Po e y Inpu s : Day # me sa elli e image y P edic < ons :
Economic indica o s Lea ned ea u es co esponding o buildings ,. 2015;4:2015.
20. Ib ahim MR, Ti he idge H, Cheng T, Hawo h J. p edic SLUMS: A new model o
iden i ying and p edic ing in o mal se lemen s and slums in ci ies om s ee
in e sec ions using machine lea ning. Compu En i on U ban Sys . 2019;76:31–56.
21. Ami Pa el, And ew C ooks and NK. Spa ial Agen -based Modeling o Explo e Slum
44
Fo ma ion Dynamics in Ahmedabad, India. 2018;37–54. A ailable om:
h p://link.sp inge .com/10.1007/978-3-319-59511-5
22. C ooks AT. Cons uc ing and implemen ing an agen -based model o esiden ial
seg ega ion h ough ec o GIS. In J Geog In Sci. 2010;24(5):661–75.
23. Uni ed Na ions Human Se lemen s P og amme. The Challenge o Slums. The
Challenge o Slums. 2012.
24. Ghani E. U baniza ion and ( In ) Fo maliza ion. U ban Po e y Reduc Con 2013
B idg Ru al U ban Pe spec . 2013;
25. Bäh J, Me ins G. Ma ginal ie el in g osss ad en de d i en wel . Geog Rundsch.
2000;
26. The con inuing challenge o in o mal se lemen s An in oduc ion - Technische
In o ma ionsbiblio hek (TIB) [In e ne ]. [ci ed 2020 Feb 6]. A ailable om:
h ps://www. ib.eu/en/sea ch/id/BLCP%3ACN063768801/The-con inuing-challenge-
o -in o mal-se lemen s/
27. Taubenböck H, K a NJ, Wu m M. The mo phology o he A i al Ci y - A global
ca ego iza ion based on li e a u e su eys and emo ely sensed da a. Appl Geog .
2018;92(Janua y):150–67.
28. Da is M. Plane o slums. New Le Re iew. 2004. 5–34 p.
29. Jamieson J. Shadow Ci ies: A Billion Squa e s, a New U ban Wo ld. J Soc Poli Econ
S ud. 2006;
30. Kohli D, Sliuzas R, Ke le N, S ein A. An on ology o slums o image-based
classi ica ion. Compu En i on U ban Sys [In e ne ]. 2012;36(2):154–63. A ailable
om: h p://dx.doi.o g/10.1016/j.compen u bsys.2011.11.001
31. Kohli D. Iden i ica ion and Cha ac e iza ion o In o mal Se lemen s Using Sa elli e
Images in Suppo o Land Adminis a ion.
32. Joka A sanjani J, Zip A, Mooney P, Helbich M. OpenS ee Map in GIScience.
OpenS ee Map GIScience Exp Res Appl. 2015;
33. Ge asoni L, Fene S, Pe ie R, S u m P. Con olu ional neu al ne wo ks o
disagg ega ed popula ion mapping using open da a. In: P oceedings - 2018 IEEE 5 h
In e na ional Con e ence on Da a Science and Ad anced Analy ics, DSAA 2018.
2019.
34. G ippa T, Geo ganos S, Za ougui S, Bognounou P, Diboulo E, Fo ge Y, e al.
45
Mapping U ban Land Use a S ee Block Le el Using OpenS ee Map, Remo e
Sensing Da a, and Spa ial Me ics. ISPRS In J Geo-In o ma ion. 2018;
35. Mahabi R, C oi o u A, C ooks A, Agou is P, S e anidis A. News co e age, digi al
ac i ism, and geog aphical saliency: A case s udy o e ugee camps and olun ee ed
geog aphical in o ma ion. PLoS One. 2018;
36. Ba ing on-Leigh C, Milla d-Ball A. The wo ld’s use -gene a ed oad map is mo e
han 80% comple e. PLoS One. 2017;
37. Fon e: Using opens ee map o c ea e land use and... - Google Schola [In e ne ]. [ci ed
2020 Feb 6]. A ailable om:
h ps://schola .google.com/schola _lookup? i le=Using+Opens ee map+ o+C ea e+La
nd+Use+and+Land+Co e +Maps:+De elopmen +o +an+Applica ion&au ho =Fon e,
+C.C.&au ho =Pa ia ca,+J.A.&au ho =Minghini,+M.&au ho =An oniou,+V.&au ho
=See,+L.&au ho =B o elli,+M.A.&publica ion_yea =2019&pages=1100–1123
38. OpenS ee Map: Use -Gene a ed S ee Maps - IEEE Jou nals & Magazine [In e ne ].
[ci ed 2020 Feb 6]. A ailable om:
h ps://ieeexplo e.ieee.o g/abs ac /documen /4653466
39. Johnson BA, Iizuka K. In eg a ing OpenS ee Map c owdsou ced da a and Landsa
ime-se ies image y o apid land use/land co e (LULC) mapping: Case s udy o he
Laguna de Bay a ea o he Philippines. Appl Geog [In e ne ]. 2016;67:140–9.
A ailable om: h p://dx.doi.o g/10.1016/j.apgeog.2015.12.006
40. Yang D, Fu CS, Smi h AC, Yu Q. Open land-use map: a egional land-use mapping
s a egy o inco po a ing OpenS ee Map wi h ea h obse a ions. Geo-Spa ial In Sci
[In e ne ]. 2017;20(3):269–81. A ailable om:
h p://doi.o g/10.1080/10095020.2017.1371385
41. Kang L, Wang Q, Yan HW. Building ex ac ion based on OpenS ee Map ags and
e y high spa ial esolu ion image in u ban a ea. In A ch Pho og amm Remo e Sens
Spa In Sci - ISPRS A ch. 2018;42(3):715–8.
42. Luo N, Wan T, Hao H, Lu Q. Fusing high-spa ial- esolu ion emo ely sensed image y
and OpenS ee Map da a o land co e classi ica ion o e u ban a eas. Remo e Sens.
2019;11(1).
43. Huang H, Kiele B, Ses e M. U ban building usage labeling by geome ic and con ex
analyses o he oo p in da a. P oceeding 26 h In Ca og Con . 2013;
52
8. ANNEXES
8.1. Independen Samples T-Tes s On Ho spo Analysis GiZSco es and Nneighbo s

53
8.2. Plo s o Accu acy and Loss o Some CNN Models - T ain (blue) and Tes
(o ange) da ase s.
8.2.1. 1-VGG
8.2.2. 3-VGG
54
8.2.3. Baseline VGG3 + Da a Augmen a ion
8.2.4. P e-T ained VGG16
55
8.3. In o mali y Regions o Kampala in 2016 ( op le ), 2017 ( op igh ) and 2019
(bo om)
01-01-2016
23-08-2017
31-12-2019
56