Advanced optimization and data modeling techniques to improve accesibility and thermal comfort in urban plannning

Uni e si y o he Basque Coun y
UPV/EHU
Doc o al Thesis
Ad anced Op imiza ion and Da a
Modeling Techniques o imp o e
Accessibili y and The mal Com o in
U ban Planning
Au ho :
Iñigo Delgado Enales
Supe iso s:
P o . D . Ja ie Del Se
D . Pa icia Molina Cos a
A Thesis submi ed in ul illmen o he equi emen s
o he deg ee o Doc o o Philosophy
in he
Depa men o Communica ions Enginee ing
Oc obe 28, 2024
(cc) 2024 Iñigo Delgado Enales (cc by-nc-nd 4.0)
iii
“The mo e I lea n, he mo e I ealize how much I don’ know.”
Albe Eins ein
UNIVERSITY OF THE BASQUE COUNTRY UPV/EHU
Abs ac
Bilbao Facul y o Enginee ing
Depa men o Communica ions Enginee ing
Doc o al Deg ee
Ad anced Op imiza ion and Da a Modeling Techniques o
imp o e Accessibili y and The mal Com o in U ban Planning
by Iñigo Delgado Enales
In he 21s cen u y, u ban a eas ha e es ablished hemsel es as he mos
common o m o human se lemen , a ou numbe ing u al a eas in e ms
o popula ion. The end o ci ies is ha hei popula ion will con inue
o g ow, un il hey accoun o 70% o he wo ld’s popula ion in 2050.
The s a egic ole ha ci ies play and will play o socie y in he u u e
is so g ea , ha since 2015, his impo ance is e lec ed in one o he 17
Sus ainable De elopmen Goals (SDGs) o he Uni ed Na ions, which aims
o make ci ies mo e inclusi e, sa e, esilien and sus ainable.
The impac o ci ies and he decisions made when planning hem, no
only a ec he u ban a ea in ques ion bu also ha e a global impac . One
example o he wide impac o u ban planning is i s con ibu ion o pollu-
ion in u ban a eas, u he exace ba ing clima e c isis. Achie ing a mo e
sus ainable ci y, ha is bo h esilien and accessible o i s inhabi an s,
depends on p ope u ban planning. U ban planne s usually ace complex
p oblems, as se e al ac o s come in o play. This is why decision suppo
sys ems play an impo an ole in acili a ing he planning and design o
mo e sus ainable and li able ci ies. These sys ems, which a e based on
digi al ools such as Geog aphical In o ma ion Sys em (GIS) o 3D mod-
eling, can help o analyze and isualize da a, e alua e di e en scena ios
and s a egies, and o esee he possible consequences o u ban planning
decisions. In he las decades, wi hin he new digi al e a wi h mo e access
o loads o da a sou ces, A i icial In elligence (AI) is pos ula ed as he
new leading suppo ool in u ban-decision making p ocesses. This Thesis
explo es he possibili ies o di e en AI echniques o u ban planning om
di e en pe spec i es.
In he i s pa o he Thesis, accessibili y and u ban com o a e
add essed. Speci ically, he Thesis p oposes a new decision-suppo sys em
based on mul i-c i e ia op imiza ion combined wi h a geo e e enced g aph
model. The amewo k makes i possible o design u ban p ojec s ha
imp o e accessibili y and educe exposu e o noise and/o ai pollu ion

i
h ough he ins alla ion o u ban elemen s ( amps, escala o s and so o h)
aking in o accoun he o al economic cos o he ins alla ion. In his
way, u ban planne s a e p o ided wi h a new ool ha akes in o accoun
he simul aneous conside a ion o di e en objec i es in an u ban planning
p oblem, hus acili a ing he decision on whe e o unde ake cos -e ec i e
ac ions in public spaces.
The second pa o he Thesis examines modeling p oblems ha ocus
on he es ima ion o hea s ess in ci ies due o clima e change. In his
second esea ch con ibu ion, he use ulness o image segmen a ion ech-
niques o co ela ing spa ial and me eo ological da a wi h s ee -le el ai
empe a u e is explo ed. The abili y o es ima e empe a u es wi h a high
deg ee o accu acy allows o he iden i ica ion o he highes p io i y a eas
in ci ies whe e u ban imp o emen s need o be made o educe he mal
discom o . The ai empe a u e a s ee le el is es ima ed bo h spa-
ially and empo ally o a speci ic use case, and compa ed wi h exis ing,
well-es ablished nume ical models. Based on he ob ained esul s, neu al
ne wo ks a e pos ula ed as a as e and less compu a ionally expensi e
al e na i e o nume ical modeling coun e pa s.
O e all, he ou comes de i ed om his Thesis con ibu e o he de el-
opmen o an emb yonic esea ch a ea –AI-based u ban op imiza ion and
modeling– ha in he u u e will enjoy ele an impo ance. The con ibu-
ions he ein epo ed expose wi h e idence he sui abili y o AI echniques
o e ec i ely sol ing u ban planning p oblems ha unde ly he design o
sa e and iendlie ci y ecosys ems.
ii
Acknowledgemen s
When I inished my physics s udies a he UPV/EHU, back in 2017,
I ne e imagined ha I would end up w i ing a Thesis on A i icial In-
elligence, a ield ha was o ally unknown o me a hose days. Thanks
i s o all o TECNALIA o us ing me, and in pa icula , o my wo
supe iso s who ha e accompanied me du ing hese 4 yea s, P o . Ja i Del
Se and D . Pa icia Molina. Thanks also o hei con inuous suppo ,
pa ience, and guidance h oughou my esea ch. Thei insigh ul eedback
and encou agemen ha e been in aluable.
Special hanks also o all my wo k colleagues I ha e had in hese ou
yea s o Thesis. Fi s o all, o my colleagues a JRL-A2I, who helped me
in he i s wo yea s o my Thesis when Co id s ill made social in e ac ion
di icul . La e , when I mo ed o TECNALIA I ound a wonde ul wo king
g oup. Thanks o all he Co e IA colleagues o p o iding suppo , collab-
o a ion and help when I needed i . You cama ade ie has made achie ing
his challenge easie and mo e enjoyable.
Finally, hanks o all my amily. Thanks o my g andpa en s, specially
o my g andpa en s ai i e y amama o aking ca e o me when I was
li le as i hey we e my pa en s. Thanks o my aun s and uncles. Thanks
especially o my pa en s Elisa and Sal ado and o my sis e Ane. Thanks
o my pa en s o ins illing in me he alues o e o , sac i ice and wo k,
wi hou which I could ne e ha e eached whe e I am. Thanks o my
sis e o always aking ca e o me, pu ing up wi h me and suppo ing
me. Thanks also o all my iends, which a e pa o my amily oo. And
las , bu no leas , hanks o my li e pa ne Ga azi which has always
been he e in good and bad momen s and has being a sou ce o ene gy
and happiness o me. Lo e you all.
ix
Con en s
Abs ac
Acknowledgemen s ii
1 In oduc ion 1
1.1 Mo i a ion ........................... 3
1.2 Objec i es ............................ 4
1.2.1 Ou line and Con ibu ions o he Thesis ....... 5
Chap e 2: Theo e ical F amewo k and Li e a u e
Re iew .................... 5
Chap e 3: Imp o ing U ban Accessibili y and En-
i onmen al Condi ions using G aph Mod-
eling and Mul i-objec i e Op imiza ion . . 5
Chap e 4: E icien Es ima ion o G ound-Le el Ai
Tempe a u e in U ban A eas using Machine
Lea ning ................... 6
Chap e 5: Concluding Rema ks and Fu u e Resea ch 6
1.2.2 Reading his Thesis .................. 6
2 Theo e ical F amewo k and Li e a u e Re iew 9
2.1 Theo e ical F amewo k ..................... 9
2.1.1 Op imiza ion P oblems ................ 11
Mul i-Objec i e Op imiza ion ............. 13
Non-domina ed So ing Gene ic Algo i hm II . . . . 14
Non-domina ed So ing Gene ic Algo i hm III . . . . 17
S eng h Pa e o E olu iona y Algo i hm 2 ...... 18
2.1.2 Modeling P oblems ................... 20
Con olu ional Neu al Ne wo ks ............ 20
Encode -Decode A chi ec u e ............ 22
2.1.3 Simula ion P oblems .................. 24
2.2 Li e a u e Re iew ....................... 25
3 Imp o ing U ban Accessibili y and En i onmen al Con-
di ions using G aph Modeling and Mul i-objec i e Op i-
miza ion 31
3.1 Rela ed Wo k .......................... 33
3.1.1 Accessibili y and En i onmen al Quali y o Age-
iendly Ci ies ...................... 33
3.1.2 Mul i-c i e ia Design and Mul i-objec i e Op imiza-
ion ........................... 34
x i
4.6 Time se ies o Ta o di e en LWT days, o all he es lo-
ca ions. The alues measu ed by he me eo ological s a ions
a e ep esen ed by he black line. The Ta alues ob ained by
he U bClim model a e in ed, whe eas he ones es ima ed
by U-Ne model in g een. ................... 88
4.7 Tempo al p og ession o 24 hou s o he ela i e empe a-
u e (T𝑟𝑒𝑙
𝑎) o he a ea o Galindo ( op). Loca ion o he
ho spo s in he a ea (bo om). ................ 90
4.8 Tempo al p og ession o 24 hou s o he ela i e empe a-
u e (T𝑟𝑒𝑙
𝑎) o he a ea o A igo iaga ( op). Loca ion o
he ho spo s in he a ea (bo om). .............. 91

x ii
Lis o Tables
3.1 Values o he pa ame e s con igu ed o he expe imen s. . . 53
3.2 S a is ics (mean±s anda d de ia ion) o he quali y indica-
o s ob ained by e e y algo i hm o e use cases 𝐴1and 𝐴2.57
3.3 𝑝- alues esul ing om a wo- ailed Wilcoxon ank sum es
pe o med o e he quali y indica o alues ob ained by e -
e y pai o algo i hms o e use cases 𝐴1and 𝐴2........ 57
4.1 Reg ession me ics compu ed be ween he wo models (U b-
Clim and U-Ne ) o each wea he s a ion du ing he pe iod
o s udy. ............................. 82
4.2 Reg ession me ics compu ed o he wo models (U bClim
and U-Ne ) wi h espec o he eal empe a u e da a col-
lec ed by each wea he s a ion du ing he pe iod o s udy.
Bes esul s o e e y s a ion and sco e a e highligh ed in
ligh blue. ............................ 82
xix
Lis o Abb e ia ions
Gene al Te ms
UN Uni ed Na ions
DSS Decission Suppo Sys em
PSS Planning Suppo Sys em
AI A i icial In elligence
GIS Geog aphic In o ma ion Sys em
IGN Ins i u o Geog a ico Nacional
DTM Digi al Te ain Model
Algo i hmic app oaches
ML Machine Lea ning
DL Deep Lea ning
ANN A i icial Neu al Ne wo k
CNN Con olu ional Neu al Ne wo k
DNN Deep Neu al Ne wo k
EC E olu iona y Compu a ion
EA E olu iona y Algo i hm
GA Gene ic Algo i hm
MOEA Mul i-Objec i e E olu iona y Algo i hm
NSGA Non-domina ed So ing Gene ic Algo i hm
SPEA S eng h Pa e o E olu iona y Algo i hm
MOCELL Mul i-Objec i e Cellula gene ic algo i hm
ReLU Rec i ied Linea Uni
Pe o mance Me ics
GD+ Gene a ional Dis ance plus
IGD+ In e ed Gene a ional Dis ance plus
HV Hype olume indica o
EP Epsilon indica o
PCC Pea son Co ela ion Coe icien
MAE Mean Absolu e E o
MAPE Mean Absolu e Pe cen age E o
RMSE Roo Mean Squa ed E o
Chap e 3
ABMS Agen -Based Modeling and Simula ion
ACO An Colony Op imiza ion
xx
GIV G een Index View
MCA Mul i-C i e ia Analysis
Chap e 4
TaS ee -Le el ai empe a u e
UHI U ban Hea Island
LST Land Su ace Tempe a u e
U bClim U ban Clima e
LWT Local Wea he Types
NDVI No malized Di e ence Vege a ion Index
1
Chap e 1
In oduc ion
Since he beginning o he indus ial e olu ion, he p opo ion o he
popula ion li ing in u ban a eas has g own s eadily. In he las 50 yea s,
his g ow h has e en accele a ed due o u al exodus, pa icula ly on he
Asian and A ican con inen s [1], [2]. Acco ding o he Uni ed Na ions
(UN), 55% o he wo ld’s popula ion li ed in ci ies in 2019, while a sha e o
70% is expec ed by 2050 [3]. This end o u ban popula ion g ow h b ings
wi h i a ious challenges, such as en i onmen al issues and p oblems wi h
mobili y and accessibili y due o poo ly planned u ban a eas. Fu he mo e,
ci ies no only in luence hei immedia e su oundings bu also ha e a -
eaching global e ec s, such as hei con ibu ion o he cu en clima e
c isis [4].
Faced wi h his s a egic global ole and he impo ance ha ci ies a e
adop ing, in ecen decades u ban planne s ha e come up wi h p oposals
o new models o ci ies. A clea case o his is he so-called age- iendly
ci y concep [5], which e e s o u ban a eas ha a e designed aking in o
accoun he needs o olde people, imp o ing hei quali y o li e and inde-
pendence. The key ac o is o keep olde people as ac i e as possible in
hei daily li es. To his end, hese ci ies p io i ise imp o ing he heal h,
sa e y and pa icipa ion o olde people. Ano he concep o ci y ha
has gained ele ance in he XXI cen u y is he concep o esilien ci y
[6]. A esilien ci y is one ha possesses he capaci y o abso b, eco e ,
and p epa e o u u e shocks, whe he hey be economic, en i onmen al,
social, o ins i u ional [7].
Following his end, and wi h a iew o he u u e, he UN has in o-
duced sus ainable ci ies as one o he goals o be achie ed unde he UN
Sus ainable De elopmen Goals [8]. Sus ainable ci ies a e de ined by he
UN as ci ies ha a e esilien , sa e and sus ainable. This sus ainabili y has
h ee main backbones, commonly called “ he h ee pilla s o sus ainabili y"
which a e: social equi y, economic iabili y, and en i onmen al p o ec ion
[9]. These h ee cha ac e is ics a e no independen and a e in e wined
wi h each o he , as depic ed in Figu e 1.1.
This Thesis add esses he sus ainabili y o ci ies om he en i onmen-
al and social side, wi h he challenge o educing he en i onmen al oo -
p in o ci ies, as well as imp o ing hei esilience while ensu ing hei
social inclusi i y. Se e al c i ical ac o s play a pi o al ole in de e mining

2Chap e 1. In oduc ion
Sus ainable
ci ies
En i onmen Economy
Social
Figu e 1.1: The h ee pilla s o sus ainabili y.
he sus ainabili y o ci ies. These ac o s in e sec and collec i ely con-
ibu e o c ea ing esilien , li able u ban en i onmen s. Some o hese
key aspec s a e now p esen ed:
•Ai and noise pollu ion: Ci ies a e acing en i onmen al challenges
de i ed om he ossil uel dependen economy, as well as om expanded
u ban models ha equi e high le els o commu ing o daily ac i i ies.
Designing ci ies o he use o p i a e ehicles means ha public spaces
a e looded wi h la ge numbe s o mo o ehicles, leading o acciden s,
noise and ai pollu ion. Ci ies cu en ly emi 36% o o al annual CO2
emissions [4]. Fu he along his line, besides CO2, ehicles also emi
ca cinogenic subs ances such as ni ogen oxides [10], leading o poo ai
quali y in ci ies and p ejudicial impac on human heal h.
•Accessibili y and walkabili y: Fo a ious easons (e.g., need o
building wi hin sho pe iods o ime, lack o esou ces) he apid and
s eady g ow h o u ban a eas has gi en ise o poo ly planned u ban
a eas ha oday leads o accessibili y p oblems in ci ies. Sa e ci ies o
pedes ians inc ease accessibili y in ci ies by encou aging walking o he
de imen o mo o ehicles. In addi ion, ha ing an accessible u ban
a ea o all ci izens p omo es social equi y among hem.
•The mal com o : Ci ies ace se e al challenges in hei way o be a
mo e li able spaces. Among hose obs acles a ises he U ban Hea Island
(UHI) phenomenon ha oge he wi h global wa ming has inc eased
sha ply he ai empe a u e and he annual hea wa e days o ou ci ies,
becoming less li able [11]. The UHI can be de ined as he di e ence
o empe a u e be ween an u ban a ea and he su ounding u al a ea,
building on he ac ha u ban a eas a e usually wa me han u al
1.1. Mo i a ion 3
en i onmen s, especially a nigh [12]. This leads o ci ies no cooling
down du ing he nigh , esul ing in highe he mal s ess du ing he day.
This s ess hinde s he he mal com o o hei inhabi an s.
These ac o s a e no always independen o each o he , and some imes
solu ions mus go h ough mul i-c i e ia planning p ocesses, o cing u -
ban planne s o adop a mo e holis ic pe spec i e when making decisions.
Wi h he ad en o compu a ional ools, planne s now ha e access o a -
ious digi al sys ems ha suppo hei decision-making p ocesses, bo h in
he sho and he long e m. Among hese ools, Decision Suppo Sys-
ems (DSS) and Planning Suppo Sys ems (PSS) ha e eme ged as aluable
asse s. They con ibu e o he planning and managemen p ocess, enhanc-
ing he quali y o ou comes, e en hough hei adop ion emains ela i ely
low. Dynamic da a om di e se sou ces should in o m hese decisions, and
hei impac should be con inuously e alua ed o ec i y any unin ended
consequences.
O e he pas ew decades, dis up i e ools o da a p ocessing and
acquisi ion—such as senso s, he In e ne o Things (IoT) and o he sma
echnologies—ha e eshaped he u ban landscape. The Sma Ci y con-
cep , as o ged by Gi inge in 2007 [13], has gained p ominence. As ci ies
become inc easingly digi ized, he e is ample oppo uni y o he adop ion
o ad anced echnologies like AI and DSS ools. In pa icula , and ocusing
on AI, he ad an ages o his echnology a e ye o be exploi ed bo h as
long- e m (PSS) and sho - e m (DSS) planning ools. Al hough he e a e
some sca ce examples o hei use in DSS (e.g.,[14]–[18]), and PSS ([19],
[20]) he exploi a ion o hese echnologies o he a o emen ioned pu pose
is s ill in an emb yonic s age. The e o e, g ea po en ial emains un apped
in he use o new echnologies o suppo DSS and PSS and ela ed p o-
cesses in public policies [21]–[23], including u ban planning. In ecen
decades, bo h p i a e companies and public en i ies a e al eady using DSS
and PSS amewo ks endowed wi h a an -ga de echnological unc ionali-
ies like AI. Ne e heless, he e is s ill a long way o go owa ds imp o ing
u he AI applica ion in u ban planning and managemen p ocesses.
1.1 Mo i a ion
AI has p og essi ely become an essen ial pa o ou li es o e he las
decades. Indeed, AI p e ails as a echnological d i e ha has ueled he
e olu ion and p og ess o many applica ion domains, yielding mani old
bene i s o bo h he economic p ospe i y and he socie y as a whole. Fo
ins ance, in medicine and heal hca e he la ency and accu acy o diagnos ic
p ocesses ha e imp o ed [24], [25]. Physical sciences ha e also ha nessed
ad ances in AI: in he ield o quan um physics, AI has educed calcula-
ion imes and has o e ed adically new app oaches o sol ing complex
many-body sys ems [26]. Simila ly, adi ional p oduc i e sec o s such
as indus ial manu ac u ing ha e obse ed dis up i e changes in hei in-
spec ion and main enance p ocesses as a esul o da a-based p ognosis,
4Chap e 1. In oduc ion
adop ing Machine Lea ning (ML) models a i s co e [27]. The echnologi-
cal momen um o AI is e lec ed by he s a egic echnological de elopmen
o impo an companies such as Google, Apple o Mic oso , all o which
emb ace AI as a nuclea compe ence o hei p ojec s and alue o e . Cu -
en p ospec s es ima e a global G oss Domes ic P oduc (GDP) inc ease
om $75 illion in 2016 o app oxima ely $114 illion by 2030: his es-
ima e is expec ed o be 14% highe as a esul o he p e alence o AI in
indus ial manu ac u ing and logis ics [28].
In con as o he acknowledged ma u i y o AI-based solu ions in p o-
duc i e p ocesses, he adop ion o his key echnology in u ban planning
and managemen has gone a a signi ican ly lowe pace. The po en ial
o AI as an in o ming ool o expe s du ing u ban decision-making p o-
cesses is ye o be apped. None heless, u ban planne s a e s a ing o use
AI-suppo ed DSS [21], [29] and PSS [30], [31] in o de o ind he mos
sui able solu ions o he complex p oblems unde gone by ci ies in mod-
e n imes. Howe e , despi e he p og essi e inco po a ion o AI in u ban
planning, he e emains a no able gap in me hodological app oaches ha
apply AI o u ban planning while conside ing c i e ia such as accessibili y,
pollu ion, and he mal com o . I is p ecisely his niche ha his Thesis
aims o ill, by add essing he objec i es se ou in he ollowing sec ion.
1.2 Objec i es
AI models o echniques can be used in di e en a eas o u ban planning
ha , ca ied ou wi h so wa e planning ools lacking in elligen unc ion-
ali ies, can be unp oduc i e, ine icien , slow o cos ly. Consequen ly, he
gene al objec i e o his Thesis is o explo e he di e en capabili ies o
AI in sol ing u ban ela ed issues. AI echniques will be explo ed and
exploi ed o imp o e u ban com o in i s di e en b anches. Speci ically,
he ocus will be on enhancing wo c ucial elemen s ha signi ican ly con-
ibu e o he sus ainabili y o ci ies: pedes ian accessibili y and he mal
com o . Di e en echniques will be add essed o each o he challenges.
As miles ones o be comple ed, he ollowing speci ic objec i es ha e been
es ablished:
•Imp o ing walkabili y h ough in elligen op imiza ion ech-
niques: To his end, a new amewo k will be p oposed o imp o e
bo h he pedes ian accessibili y o ci ies and hei pe cep ion o he
en i onmen al quali y. Special emphasis will be placed on imp o ing
accessibili y o he elde ly and people wi h educed mobili y, aking
in o accoun hei demands. The aim is also o make a con ibu ion o
he sca ce li e a u e on he imp o emen o walking ou es wi h AI.
•Explo ing AI-based modeling echniques o he mal com o :
Modeling is a e y ea ly ield in his a ea and has a lo o po en ial o
de elopmen o u ban applica ions. In his scena io, he emphasis is
on image- o-image modeling o es ima e ai empe a u e in ci ies. Deep
Neu al Ne wo ks (DNNs) can be a ime-sa ing way o de ec hea s ess
in di e en ci ies.
1.2. Objec i es 5
The wo speci ic objec i es o his hesis a e echnologically indepen-
den . Howe e , om an u ban planning pe spec i e, hese ac o s can-
no be conside ed in isola ion, as hey in ol e dis inc decision-making
elemen s ha may con lic wi h one ano he , ul ima ely in luencing he
o e all decision-making p ocess.
1.2.1 Ou line and Con ibu ions o he Thesis
This Thesis begins by p o iding some p elimina ies on he echnology
adop ed du ing he cou se o esea ch, ocusing on he speci ic echniques
and sub- asks ha will be used in he wo echnical con ibu ions o he
Thesis: me a-heu is ic op imisa ion and deep lea ning (DL). Then, a b ie
e iew o he gene al s a e-o - he-a o he ield will be made, ocusing
i s co e age on he applica ion o AI-based app oaches o esilien ci ies.
He ea e , he wo main echnical con ibu ions o his hesis will be p e-
sen ed, del ing in o he me hodology p oposed in each o hem and he
esul s alida ing hei applicabili y o eal-wo ld u ban p oblems. The
las chap e will ou line and conclude he Thesis wi h a summa y o he
main conclusions d awn om he Thesis. A sho desc ip ion o each o
he chap e s o he Thesis is gi en nex :
Chap e 2: Theo e ical F amewo k and Li e a u e Re iew
This chap e p o ides he essen ial con ex o a p ope unde s anding o
he echnical con ibu ions p esen ed in he ollowing chap e s. Fi s , an
in oduc ion o he me hods used in his wo k and an explana ion o he
basic concep s in ol ed will be done. The discussion will e ol e a ound AI
and he h ee main p oblems co e ed by his echnological pa adigm: op i-
miza ion, modeling and simula ion. Addi ionally, a s a e-o - he-a sec ion
is included, showcasing di e en wo ks ha display he applicabili y o AI
in he ield o esilien ci ies.
Chap e 3: Imp o ing U ban Accessibili y and En i onmen al
Condi ions using G aph Modeling and Mul i-objec i e Op imiza-
ion
This chap e p esen s a no el amewo k based on geo e e enced g aph
modeling combined wi h e olu iona y op imiza ion o he imp o emen
o accessibili y and pa h condi ions o olde people. An u ban a ea is
ep esen ed by a mul i-weigh ed g aph and a ached o a mul i-objec i e
op imiza ion algo i hm. This mul i-c i e ia p oblem is ackled wi h mul i-
objec i e e olu iona y algo i hms (MOEAs), esul ing in a se o Pa e o-
op imal se o solu ions ha e e o he se o possible u ban in e en ions.
This chap e mani es s he use ulness o hese ounda ions as a DSS o
u ban planne s, who usually ha e o decide on an u ban in e en ion on
he basis o di e en ac o s such as in es men and impac .
12 Chap e 2. Theo e ical F amewo k and Li e a u e Re iew
he cons ain s a e all linea unc ions. Tha is, hey can be exp essed in
i s scala o m. A linea objec i e unc ion can be exp essed as:
𝑓(x)=A·x+𝑏, (2.1)
whe e A∈R1×𝑁and 𝑏∈R.
On he o he hand, non-linea op imiza ion p oblems a e a se o p ob-
lems whe e ei he he objec i e unc ion o he cons ain s o bo h con ain
nonlinea pa s. This means ha he p oblems ha e mul iple a iables and
he ela ionship be ween hese a iables is no linea .
Fo sol ing non-linea p oblems he e a e se e al echniques. Fo ex-
ample, when he p oblem in ol es a con ex objec i e unc ion, echniques
such as g adien descen can be applied. G adien descen is an i e a i e
op imiza ion algo i hm o inding he minimum o a unc ion. I wo ks
by aking s eps p opo ional o he nega i e o he local g adien (o ap-
p oxima e g adien ), i.e. he g adien o he unc ion a he cu en poin .
The algo i hm is said ha i con e ges when i eaches a minimum. The
same echnique can be applied o a conca e unc ion, using he g adien
ascen me hod ins ead.
Howe e , no all op imiza ion p oblems ha e hese p ope ies. Some
p oblems a e non-linea , non-con ex, o in ol e complexi ies whe e no
ma hema ical exp ession can adequa ely ep esen he sys em o model o
be op imized, making i a black box p oblem ha classical me hods s ug-
gle o sol e. Fo hese cases in elligen op imiza ion, a b anch o AI ha
c ea es algo i hms capable o lea ning by hemsel es o sol e op imiza ion
p oblems, is applied. In elligen op imiza ion is also gene ally called me a-
heu is ics [32]. Me a-heu is ics, a e m coined by F ed Glo e in [33], a e
mode n na u e-inspi ed algo i hms ha ou pe o m basic sea ch heu is ics
(e.g. Tabu sea ch algo i hm, G eedy algo i hm o A* algo i hm). Unlike
heu is ics, me a-heu is ics a e no p oblem-speci ic. Me a-heu is ics a e in
gene al s ochas ic algo i hms ha le e age a balance be ween explo a i e
and exploi a i e (local) sea ch, also known as di e si ica ion and in en-
si ica ion. Finding a good balance be ween hese wo sea ch beha io s
h ough he design and con igu a ion o me a-heu is ic sol e s is impe a-
i e o explo e he solu ion space o an op imiza ion p oblem e icien ly,
ind p omising a eas ea ly du ing he sea ch, and ocus he sea ch in such
a eas [34]. These algo i hms can ind high-quali y solu ions o challenging
op imiza ion p oblems in a easonable compu ing ime, bu hey do no
gua an ee op imal solu ions. They a e expec ed o wo k mos o he ime
and a e well-sui ed o global op imiza ion. The selec ion o he bes solu-
ions ensu es con e gence o he op imum, while di e si ica ion inc eases
he di e si y o solu ions and allows escape om local op ima.
Among he mos enowned me a-heu is ic algo i hms a e he amily
o E olu iona y Compu a ion echniques (EC), al hough o he s such as
Simula ed Annealing [35] can also be ound. Inside he amily o EC he
mos p ominen b anch is ha o E olu iona y Algo i hms (EAs), ha ing
Gene ic Algo i hms (GAs) [36] as he mos equen ly encoun e ed ype
o EA. GAs a e inspi ed by he biological e olu ion o selec ion, ep o-
duc ion and mu a ion. The gene al algo i hmic low o a GA s a s wi h

2.1. Theo e ical F amewo k 13
a popula ion o candida e solu ions ( he decision a iables) and e ol e i
o e gene a ions. Each solu ion has p ope ies ha can be mu a ed and
al e ed. The objec i e unc ion o each solu ion, which is also known as
i ness in he con ex o EC, is e alua ed and he i es ones a e selec ed
o he nex gene a ion, whe e hei decision a iables (genes) a e again
modi ied ia he applica ion o sea ch ope a o s (namely, c osso e and
mu a ion) ha imp in changes o hei nume ical alues. The p ocess
con inues un il a p ede ined s opping c i e ion is me (e.g. a maximum
numbe o gene a ion o a lack o imp o emen o he bes solu ion o e
gene a ions).
So a , op imiza ion wi h a single objec i e unc ion has been desc ibed.
In he nex subsec ion, p oblems d i en by se e al objec i es will be de-
sc ibed, along wi h a b ie p esen a ion o he EC-based sol e s capable o
ackling hese p oblems ha ha e been conside ed in he Thesis.
Mul i-Objec i e Op imiza ion
No mally, in di e en a eas o eal li e se e al objec i es mus be aken
in o accoun , which can be opposing each o he . Fo ins ance, isk e sus
p o i abili y in be ing games o cos e sus eliabili y in indus ial main-
enance. This is he case also in u ban planning, whe e decision make s
o en ace he dilemma o accomplishing u ban enewal while educing he
cos o u ban ac ions. This is why echniques o add ess mul i-objec i e
p oblems a e o g ea in e es . Building upon he example p o ided in he
sec ion on single-objec i e op imiza ion (TSP), he de ini ion is ex ended
o illus a e wha a mul i-objec i e op imiza ion p oblem s ands o .
In he single-objec i e TSP, he objec i e is o minimize he o al dis-
ance 𝐿co e ing all ci ies once. Howe e , le us now assume ha he
anspo a ion company no only wan s o minimize he dis ance bu also
maximize he p o i s. This p oblem is known as he T a eling Salesman
P oblem wi h P o i s (TSPP) [37]. As an illus a i e example, an open
a ian o he TSPP, e e ed o as he Open TSPP, is selec ed. In his
open e sion, he a ele is no equi ed o isi all ci ies bu can choose
a subse o op imize he o e all cos . Speci ically, wo key condi ions a e
assumed: (1) he p oblem is open, allowing lexibili y in ci y selec ion,
and (2) he e is a dis inc cos associa ed wi h isi ing each ci y, mean-
ing ha he bene i s o he jou ney depend no only on he numbe o
ci ies isi ed bu also on he speci ic ci ies chosen. This se up leads o
a Pa e o on ha eme ges om he in ui ion ha he cos is in e sely
ela ed o he dis ance a eled and he pa icula subse o ci ies selec ed
o he ou e. Hence, in his case, he p oblem becomes om ha ing a
single objec i e unc ion o minimize, he o al dis ance 𝑓(x)=𝐿 o ha -
ing wo objec i e unc ions: he o al dis ance 𝑓1(x)and he p o i 𝑓2(x).
The goal is o ind he se o Pa e o-op imal solu ions ha minimize 𝑓1(x)
(dis ance) and maximize 𝑓2(x)(p o i s). In mul i-objec i e op imiza ion,
a Pa e o op imal solu ion (also known as a non-domina ed solu ion) is a
solu ion o which i is impossible o imp o e one objec i e wi hou wo s-
ening a leas one o he objec i e. The se o non-domina ed solu ions
composes he sough Pa e o on , ep esen ed in Figu e 2.4 o a min-min
14 Chap e 2. Theo e ical F amewo k and Li e a u e Re iew
mul i-objec i e op imiza ion p oblem. The challenge o mul i-objec i e
op imiza ion sol e s is o e icien ly compu e his Pa e o on o a gi en
se o objec i es, o a leas p o ide a se o solu ions ha app oach his
Pa e o op imal se , bo h in e ms o con e gence and di e si y.
Op imal Pa e o on
1(x)
2(x)
Domina ed solu ion
Non-domina ed solu ion
Figu e 2.4: Pa e o dominance in a mul i-objec i e op imiza ion p oblem
comp ising wo objec i es 𝑓1(x)and 𝑓2(x) ha mus be minimized. Poin s x
a e ca ego ized as Domina ed solu ions and Non-domina ed solu ions. The
Pa e o on cu e ep esen s he op imal ade-o be ween objec i es 𝑓1(x)
and 𝑓2(x). The goal is o p oduce a se o solu ions ha bes app oxima e
his op imal Pa e o on .
App oxima ing he op imal Pa e o on o a mul i-objec i e op imiza-
ion p oblem can become a complica ed ask when he dimensionali y o
he p oblem makes an exhaus i e e alua ion o solu ions compu a ionally
una o dable, and/o he objec i es do no possess a ac able ma hema -
ical o mula ion. In such cases, me a-heu is ic op imiza ion sol e s can be
ex ended o e ain solu ions based on di e en c i e ia ela ed o Pa e o
op imali y. This gi es ise o he amily o MOEAs. Once again, as in
single-objec i e p oblems, GAs a e he mos widely used class. Wi hin
GAs a wide ange o algo i hms can be ound ha a e conduci e o mul i-
objec i e op imiza ion. In wha ollows we ocus on h ee o he mos
exploi ed GAs in pa icula and explain hei mos di e en ial ea u es:
Non-domina ed So ing Gene ic Algo i hm II (NSGA-II), Non-domina ed
So ing Gene ic Algo i hm III (NSGA-III) and S eng h Pa e o E olu ion-
a y Algo i hm 2 (SPEA2).
Non-domina ed So ing Gene ic Algo i hm II
The NSGA-II is an imp o ed e sion o he NSGA algo i hm, p oposed
by Deb e al. [38]. I is conside ed as he EA o e e ence when applying
2.1. Theo e ical F amewo k 15
mul i-objec i e op imiza ion due o i s widely p o en obus ness and e -
iciency. NSGA-II has a compu a ional complexi y o O(𝑀𝑁2), whe e M
is he numbe o objec i es and N is he popula ion size, gi ing i lowe
compu a ional imes compa e o o he MOEAs. Be o e explo ing he algo-
i hmic low, some key ea u es will be p esen ed. The e a e h ee aspec s
ha make he NSGA-II di e en om o he mul i-objec i e algo i hms:
•Non-domina ed so ing: Indi iduals a e anked based on Pa e o domi-
nance. An indi idual is said o domina e ano he i i is no wo se in all
objec i es and be e in a leas one objec i e. The popula ion is so ed
in o di e en on s (F1,F2,F3, ...), which a e shown in Figu e 2.5a. The
i s on (F1) consis s o non-domina ed indi iduals, he second on
(F2) is domina ed only by indi iduals in he i s on , and so o h.
•Di e si y p ese ing mechanism: I is based on he concep o c owding
dis ance (CD). The CD, illus a ed in Figu e 2.5b, quan i ies how close
a solu ion is o i s neighbo ing solu ions, ep esen ed by he pe ime-
e o he cuboid su ounding he indi idual in he objec i e space.
To main ain di e si y wi hin he popula ion, solu ions loca ed in less
c owded egions— hose wi h la ge c owding dis ances—a e p io i ized
o selec ion. In he e en ha he popula ion limi is exceeded, and
mul iple solu ions sha e he same ank (e.g., 7 indi iduals in ank 1 and
8 in ank 2), he algo i hm e ains all he ank 1 indi iduals (i.e. non-
domina ed so ing) and selec s addi ional solu ions om ank 2 based
on hei c owding dis ance, a o ing hose wi h he highes alues. This
ensu es a well-dis ibu ed se o solu ions ac oss he Pa e o on while
main aining he popula ion size.
•Eli ism: In NSGA-II, eli ism e e s o he p ese a ion o he bes solu-
ions om a popula ion, speci ically he non-domina ed solu ions, in o
he nex gene a ion. To implemen eli ism, i is essen ial o e alua e
he quali y o a solu ion in e ms o i s p oximi y o Pa e o op imali y.
This is achie ed using a dual c i e ion: non-domina ed so ing, which
anks solu ions based on hei dominance ela ionships, and CD, which
ensu es di e si y by measu ing he densi y o solu ions in he objec i e
space.
Once he basic inhe en concep s o NSGA-II ha e been explained, now
he algo i hmic p ocedu e will be p esen ed. The Figu e 2.6 p esen s he
schema ic o he p ocedu e. The i s s ep is o c ea e a ini ial popula ion
P , whe e 𝑡 ep esen s he gene a ion numbe , wi h 𝑡=0co esponding o
he i s gene a ion. A e applying he usual gene ic ope a o s o c osso e
and mu a ion ope a ions, an o sp ing popula ion Q is c ea ed. By using
bo h popula ions oge he , a new one o size 2N is o med, say R . Then,
he non-domina ed so ing is applied, placing he indi iduals in he di e -
en on s acco ding o hei non-dominance. The bes solu ions will be
placed a he 𝐹1 on . Once he i s on is ull, he solu ions will be
placed in he second on 𝐹2and so o h. The R size is 2N while he
new popula ion R +1 size mus be N. Hence, he mos domina ed on s
a e ejec ed. When e alua ing he inal accep ed on (𝐹3in Figu e 2.6),
16 Chap e 2. Theo e ical F amewo k and Li e a u e Re iew
i migh con ain mo e poin s han he a ailable spaces in he new popula-
ion. Ra he han andomly elimina ing some elemen s om he las on ,
he poin s ha maximize he di e si y o he chosen poin s a e selec ed,
hanks o he c owding dis ance. A e all he p ocess, he new popula ion
P +1 is c ea ed and he p ocess is epea ed again, un il an op imal solu ion
is eached.
Pa e o on Rank 1
Rank 2
Rank 3
1(x)
2(x)
(a)
Cuboid
i-1
i
i+1
1(x)
2(x)
(b)
Figu e 2.5: The ollowing igu es illus a e wo key aspec s o he NSGA-II
algo i hm: (A) The non-domina ing so ing p ocedu e. The cu e closes o
he o igin is he Pa e o on , ep esen ing he op imal se o solu ions. (B)
The c owding dis ance calcula ion. i-1, i and i+1 a e di e en solu ions o
he Pa e o on .
Non-domina ed
so ing
F1
F2
F3
Rejec ed
P +1
P
Q
R
Figu e 2.6: NSGA-II p ocedu al scheme o a popula ion 𝑃𝑡and a o sp ing
popula ion 𝑄𝑡, which oge he o m he new popula ion 𝑅𝑡. Each o he
𝐹1, 𝐹2, 𝐹3a e he di e en on s con aining solu ions o 𝑅𝑡so ed by applying
he non-domina ed so ing. The ho izon al dashed line limi s he numbe o
N solu ions allowed o he nex 𝑃𝑡+1popula ion.
2.1. Theo e ical F amewo k 17
Non-domina ed So ing Gene ic Algo i hm III
The NSGA-III [39] can be seen as an ex ension o he NSGA-II algo i hm,
de eloped o imp o e he e iciency in op imiza ion p oblems comp ising
mo e han wo objec i es. NSGA-III sha es simila i ies wi h i s p ede-
cesso s (NSGA and NSGA-II) since he popula ion is so ed in di e en
on s acco ding o he non-dominance o he indi iduals, in he simila
way i is done in NSGA-II. Howe e , he di e si y p ese ing mechanism
is di e en . In NSGA-III he c owding dis ance in oduced in NSGA-II is
eplaced by a se o e e ence poin s o di ec ions in he objec i e space
ha guide he sea ch owa ds di e en egions o he Pa e o on . The
se o e e ence poin s is de e mined in he beginning o he sea ch and
is widely dis ibu ed in he objec i e space and nea he Pa e o on o
ensu e di e si y. Fo each se o solu ions, he pe pendicula dis ances be-
ween each solu ion and he di e en e e ence di ec ions a e calcula ed,
keeping he sho es dis ance o one o he e e ence di ec ions. Fo all
solu ions ha all wi hin his egion, i.e. o all solu ions sha ing he same
e e ence di ec ion, niche p ese a ion is applied. This means ha only
he solu ion wi h he smalles dis ance is sa ed o he nex popula ion.
This p ocedu e is applied in all he di e en egions ha a e di ided by he
di e en e e ence di ec ions. This ensu es di e se and uni o m solu ions
a e kep wi hin he a ge space in he nex popula ion. Figu e 2.7 shows
he concep s o e e ence poin and di ec ions and he niche p ese a ion.
1
2
1
2
3
4
5
6
Re e ence
line
Re . poin
No malized
hype plane
Figu e 2.7: NSGA-III algo i hm wo king p ocedu e. Poin s numbe ed
om 1 o 6 a e plo ed, showing po en ial solu ions. Re e ence lines and he
No malized hype plane in e sec a a Re e ence poin s, indica ing e e ence
di ec ions o he op imiza ion p ocess. In 2D o cla i y.
NSGA-III has yielded consis en esul s on mul i-objec i e p oblems
o be ween h ee and 15 objec i es [39]. The p oblems conside ed in he
ex ensi e benchma k pe o med in his la e wo k a e he e ogeneous and

18 Chap e 2. Theo e ical F amewo k and Li e a u e Re iew
comp ise di e en objec i e unc ions anging om con ex, conca e, dis-
join ed and mul i-modal ones, in ol ing mul iple local on s. In all o
hem, NSGA-III managed o ou pe o m o he mul i-objec i e sol e s wi h
supe io Pa e o on app oxima ions in e ms o con e gence and di e -
si y.
S eng h Pa e o E olu iona y Algo i hm 2
The SPEA2 [40] is an imp o ed e sion o he SPEA algo i hm. The up-
g ades hinge on h ee main pilla s: i) imp o ed i ness assignmen s a egy,
ii) a densi y es ima ion echnique and iii) a new a chi e unca ion me hod.
In he ollowing a desc ip ion o he h ee mos ele an ea u es o SPEA2
is p esen ed:
•Ex e nal a chi e: In bo h SPEA and SPEA2, he e is a ex e nal a chi e
whe e non-domina ed solu ions encoun e ed du ing he sea ch p ocess
a e s o ed. The a chi e is upda ed e e y gene a ion o main ain he
bes solu ions so a . Howe e , while o SPEA he size a ies in each
gene a ion, SPEA2 ope a es wi h a cons an size o bo h he popula ion
and he ex e nal a chi e.
•Fi ness assignmen : The i ness o an indi idual x𝑝in he popula ion is
composed by he sum o i s aw i ness 𝑅(𝑝)and densi y 𝐷(𝑝). S a -
ing om a popula ion 𝑃𝑡 ha se es as he exis ing pool o po en ial
solu ions and an ex e nal a chi e ¯
𝑃𝑡, le assign o each indi idual x𝑝
inside ha popula ion and a chi e wi h a s eng h alue 𝑆(𝑝), which
ep esen s he numbe o solu ions i domina es:
𝑆(𝑝)=|{𝑝′|𝑝∈𝑃𝑡+¯
𝑃𝑡∧𝑝≻𝑝′}|,(2.2)
whe e | · | deno es he ca dinali y o a se , + s ands o mul ise union
and he symbol ≻co esponds o he Pa e o dominance ela ion. The
aw i ness 𝑅(𝑝)o an indi idual x𝑝is calcula ed based on he 𝑆(𝑝):
𝑅(𝑝)=∑︁
𝑝′∈𝑃𝑡+¯
𝑃𝑡, 𝑝≻𝑝′
𝑆(𝑝′).(2.3)
In o he wo ds, he aw i ness is calcula ed based on he s eng hs o
hose ha domina e i in bo h he a chi e and he popula ion, unlike
in SPEA whe e only membe s o he a chi e a e aken in o accoun in
his ega d. The goal he e is o minimize i ness, meaning ha 𝑅(𝑝)=0
signi ies a non-domina ed indi idual, while a high 𝑅(𝑝) alue indica es
ha he indi idual 𝑝is domina ed by a la ge numbe o indi iduals, as
depic ed in Figu e 2.8. Finally, he densi y is compu ed using he k- h
nea es neighbo [41] me hod:
𝐷(𝑝)=1
𝜎𝑘
𝑝+2,(2.4)
2.1. Theo e ical F amewo k 19
1
2
0
0
0
Non-domina ed
Domina ed
2
5
9
19 14
Figu e 2.8: The aw SPEA 2 i ness alues o a maximiza ion p oblem.
whe e 𝜎𝑘
𝑝is he euclidean dis ance be ween he indi idual 𝑝and i s k- h
nea es neighbo in objec i e space. The inal i ness alue o he x𝑝
indi idual is:
𝑄(𝑝)=𝑅(𝑝) + 𝐷(𝑝).(2.5)
•En i onmen al selec ion: ensu es bo h he di e si y and con e gence
o he popula ion. The en i onmen al selec ion p ocess copies all he
non-domina ed solu ions (𝑄(𝑝)<1) o he nex gene a ion a chi e. I
he size o he non-domina ed solu ions i s he p e-de ined size o he
a chi e, deno ed as ¯
𝑁, he en i onmen al selec ion is ul illed. I he e
a e less non-domina ed solu ion han ¯
𝑁, he bes domina ed solu ions
a e added un il illing all he a chi e. Howe e , i he non-domina ed
solu ions exceed he size ¯
𝑁, an a chi e unca ion is done in an i e a i e
emo al p ocess whe e he indi idual which has he minimum euclidean
dis ance o ano he indi idual is emo ed.
The algo i hmic low is p esen ed as in [40]. Fi s an ini ial popula ion
𝑃0is gene a ed unde an ini ializa ion c i e ion, while he a chi e ¯
𝑃𝑡=∅
is ini ially emp y. Then he i ness compu a ion p ocedu e is pe o med
o all he indi iduals in 𝑃𝑡and ¯
𝑃𝑡( o he i s a chi e no i ness com-
pu a ion is done, since i is emp y). Nex , he en i onmen al selec ion is
applied. A ha poin , i he maximum numbe o gene a ions has been
eached o ano he s opping c i e ia is sa is ied he ¯
𝑃𝑡+1is se as he ou pu
non-domina ed se . Then a bina y ou namen selec ion wi h eplacemen
on ¯
𝑃𝑡+1is pe o med on P in o de o ill he ma ing pool. Nex , ecom-
bina ion and mu a ion ope a o s a e applied o he ma ing pool and he
esul ing popula ion is assigned o ¯
𝑃𝑡+1. Finally, he gene a ion coun e is
inc emen ed (𝑡=𝑡+1) and he i ness assignmen is ecalcula ed again. Like
NSGA-II, SPEA2 is a widely used algo i hm in he ield o op imiza ion.
20 Chap e 2. Theo e ical F amewo k and Li e a u e Re iew
2.1.2 Modeling P oblems
Wi hin modeling p oblems, se e al ypes can be encoun e ed depending
on he da a o ma o be p ocessed. Gi en he la ge modeling capabili y
o mode n ML echniques (e.g. DNN), mos modeling p oblems ackled
ia AI a e o mula ed on complex da a modali ies, including ex , audio
and image da a. Among hem, image modeling is a guably one o he
mos widely conside ed da a modali ies ackled ia AI-based echniques.
In image modeling p oblems, se e al modeling asks can be dis inguished,
including:
•Image Classi ica ion: Image classi ica ion modeling consis on classi ying
images wi h a unique label. A ypical example o image classi ica ion
modeling is he dog-ca classi ie , whe e he model is ained wi h se e al
dog and ca images and lea ns o classi y hem in o hei espec i e
ca ego ies.
•Image Segmen a ion: This ype o modeling can be conside ed as a s ep
o wa d in image classi ica ion modeling. Image segmen a ion models
wo k a pixel-le el and lea n o label pixels acco ding o hei class, di-
iding he image in di e en segmen s. The e a e wo ypes o segmen-
a ion: seman ic segmen a ion and ins ance segmen a ion. In seman ic
segmen a ion he pixels a e labeled and g ouped based on he ca ego ies
hey belong o (e.g. ees, pe sons, ca s...). In ins ance segmen a ion,
his pa i ioning is mo e exac and he di e en ca ego ies a e also seg-
men ed depending on he indi idual objec s (e.g. ee one, ee wo,
human one, human wo...).
•Image Reg ession: Unlike he wo p e ious models, image eg ession
wo ks wi h con inuous alues ins ead o labels. The model is ained
o ansla e he pixels o an image, o he whole image, o con inuous
nume ical alues. The model lea ns o map he ea u es ex ac ed om
he image o he a ge eg ession alues. Image eg ession models ha e
a wide ange o applica ions, including age es ima ion [42], [43], acial
landma k de ec ion [44], [45], and bounded box eg ession in objec
de ec ion [46].
In ecen yea s, modeling p oblems o mula ed on image da a as hose
exempli ied abo e a e mos ly sol ed by using DNNs. In o de o accoun
o he spa ial ela ionships wi hin image da a ha a e inhe en o such
p oblems, DNNs used in image classi ica ion a e buil on op o con olu-
ional neu al ne wo ks (CNNs), which we e ini ially p oposed by Lecun
[47] inspi ed by ea ly wo k on he neocogni on by Fukushima [48]. In
wha ollows we b ie ly e isi he undamen als o his pa icula la o o
DNNs.
Con olu ional Neu al Ne wo ks
CNNs a e a ype o eed- o wa d neu al ne wo ks p ima ily used o im-
age p ocessing and objec ecogni ion [48]. The eason ha CNNs a e so
2.1. Theo e ical F amewo k 21
ex ensi ely used in his a ea is ha hey a e designed o au onomously
and dynamically lea n spa ial posi ions o ea u es om he inpu da a.
In CNNs he spa ial in o ma ion is p ese ed, which p o ides he con-
ex and ela ionship be ween di e en pa s o he image. This can be
achie ed hanks o he con olu ion ope a ion ha his neu al ne wo k ap-
ply o he inpu da a, which downsamples he inpu da a while main aining
he spa ial ela ionships be ween he pixels o he image. The supe icial
a chi ec u e o CNNs does no di e om hose o o he ypes o ANNs
and consis s o h ee la ge building blocks, as can be seen in Figu e 2.9 .
Fi s , an inpu laye holds he inpu pixel ma ix o he image. Then,
se e al hidden laye s can be ound. The hidden laye s a e he co e pa
o he CNNs and he building block ha di e en ia es hem om o he
ypes o ANNs. Finally, an ou pu laye which is he inal laye ha
p oduces he p edic ions o classi ica ions based on he lea ned ea u es
om he p e ious laye s. The main building block -hidden laye s- de-
pic ed in Figu e 2.9 , can be b oken down in o h ee main ypes o laye s:
con olu ional laye s, pooling laye s and ully connec ed laye s.
•Con olu ional laye s: The con olu ional laye s apply he con olu ion
ope a o o he inpu image. A con olu ion is a linea ope a ion ha
in ol es he scala p oduc be ween he pixels o he image and he di -
e en weigh s o he ke nels. A ke nel (o il e ) is a ma ix o weigh s.
These il e s a e “con ol ed” o e he inpu ma ix o compu e a ea u e
map. In o he wo ds, he il e is slid o e he en i e inpu image, le o
igh , op o bo om. The con olu ion ope a ion exhibi s se e al p ope -
ies ha a e pa icula ly ad an ageous o compu e ision asks. Fi s ,
he pa e ns lea ned by con olu ional laye s a e ansla ion in a ian ,
meaning ha a pa e n ecognized a one loca ion can be de ec ed e-
ga dless o i s spa ial posi ion wi hin he image. Second, con olu ional
laye s cap u e he spa ial hie a chies o pa e ns. This hie a chical lea n-
ing allows ea ly laye s o iden i y undamen al, small-scale s uc u es
such as edges, while deepe laye s de ec inc easingly complex pa e ns,
which a e essen ially combina ions o ea u es ex ac ed by p eceding
laye s. As a esul , a CNN is able o p og essi ely comp ehend mo e
in ica e and abs ac isual pa e ns h ough i s laye ed a chi ec u e.
•Pooling laye s: Pooling laye s a e applied immedia ely a e a con olu-
ional laye o p ocess he esul ing ea u e maps. The p ima y unc ion
o a pooling laye is o pe o m downsampling, he eby educing he di-
mensionali y and he numbe o pa ame e s in he model. A commonly
employed pooling echnique is Max Pooling, which educes he size o
he ea u e maps by selec ing he maximum alue wi hin each de ined
window. This ope a ion no only dec eases compu a ional complexi y
bu also helps o e ain he mos salien ea u es o he inpu .
•Fully Connec ed laye s (FC): FC laye s, which ypically ollow he con o-
lu ional and pooling laye s, es ablish connec ions be ween e e y neu on
in he cu en laye and e e y neu on in he p eceding laye . Th ough
his dense connec i i y, he FC laye s lea n pa e ns and ela ionships
28 Chap e 2. Theo e ical F amewo k and Li e a u e Re iew
and hus calcula e he G een Index View (GIV). Ki e al. [66] go a s ep u -
he and also applies image segmen a ion o Google S ee View o ob ain
he GIV and hen, hey co ela e his GIV wi h he walking ime spen
on hose s ee s. The wo k also shows ha o he ac o s such as income
and ca use in a gi en neighbo hood also a ec bo h GIV and walking
ime. Again, as in he wo k o Naga a e al. [64], he impac o his wo k
o e s de elopmen o public policies ha encou age pedes ian- iendly
en i onmen s. In he ield o accessibili y, Google S ee View images ha e
also been used as a da abase. Fo ins ance, Ha a e al. [67] p opose a
new sys em called Tohme ha le e ages c owd-sou cing, compu e ision,
and machine lea ning o de ec cu b amps in Google S ee View scenes.
Tohme can e ec i ely analyze and iden i y cu b amps in Google S ee
View scenes by using image segmen a ion, con ibu ing o he sys em’s
o e all success in de ec ing accessibili y issues. The sys em’s inno a i e
app oach signi ican ly educes he human ime cos associa ed wi h man-
ual audi s while main aining high quali y in esul s, making i a aluable
ool o u ban planning and imp o ing accessibili y in ci ies wo ldwide.
So a , he e iew o he s a e o he a has ocused on segmen a ion
o Google S ee View images, a ield wi h a wide scope. Howe e , o he
ypes o examples can also be ound wi hin modeling. Fo example, Wang
e al. [68] in oduce a geome ic consis ency enhanced deep con olu ional
encode -decode amewo k o u ban seismic damage assessmen . The
encode -decode a chi ec u e used is a U-Ne wi h sligh modi ica ions.
They apply image segmen a ion o ecognize unde di e se wea he condi-
ions ( ain, og, da kness) ea hquake-damaged buildings. They ain he
model using unmanned ae ial ehicle images om he coun y o Beichuan
achie ing eliable, accu a e and obus esul s, inc easing he esilience
o he u ban a ea. The e a e also some wo ks whe e he wo echniques
me ge and he op imiza ion p ocess is made o e he pa ame e s o he
DL models chosen wi h he aim o de eloping mo e p ecise modeling ech-
niques. This is he case o Pan e al. [69], who p opose a di e en ial
e olu ion (DE) back p opaga ion neu al ne wo k a ic p edic ion model
o sma ci ies. The model accu a ely p edic s ne wo k a ic ends by
mapping he impac ac o o ne wo k a ic o he ac ual a ic olume.
The model u ilizes DE o global sea ch o op imize he connec ion weigh s
be ween laye s, enhancing he accu acy o p edic ions. Th ough aining
wi h pas a ic da a, he model achie es accu a e p edic ion o ne wo k
a ic ends in sma ci ies. The p edic ed a ic olume aligns wi h he
ac ual a ic olume end wi hin an small e o . I has also p o en o
be sui able o long- e m p edic ions in complex and he e ogeneous sma
ci y ne wo k en i onmen s. Thanks o quali y p edic ions, u ban planne s
a e assis ed in managing he a ic ne wo k as well as in imp o ing in as-
uc u e. Simila ly, Alghamdi in his wo k [70] employs a combina ion o
EAs and DL models, Recu en Neu al Ne wo ks (RNNs) speci ically, o
p edic he u ban e olu ion o u u e sma ci ies. RNNs a e ained wi h
da ase s coming om Ci yPulse amewo k and as in he p e ious case,
he EA is applied o he op imiza ion o he weigh s o he neu al ne -
wo k laye s. Once he model is ained, hey p opose wo use cases. The

2.2. Li e a u e Re iew 29
i s use case is o c ea ing a new ci y om sc a ch, wi h some cons ains
in ene gy, numbe o inhabi an s and ehicles applied. The amewo k
ends o buil ci ies wi h he maximiza ion o co e age, connec i i y, and
localiza ion. The second use case p edic s he po en ial de elopmen s in
he u ban g ow h o a cu en ci y. This scheme could help o guide he
expansion o a ci y.
The examples p o ided ep esen jus a ew s okes on he as can as o
modeling o u ban p oblems. The e exis s a mul i ude o o he examples
ha u he illus a e he dep h and b ead h o his ield. Fo ha , a
mo e comp ehensi e explo a ion o he s a e-o - he-a is p esen ed in he
echnical con ibu ion o Chap e 4.
This p elimina y ela ed wo k e eals ha he ield o AI applied o u -
ban planning and i s challenges is an incipien ield o signi ican p ac ical
in e es . Howe e , as his ini ial e iew e inces, mos wo ks concen a e
on a pu ely heo e ical app oach o he p oblems. Municipali ies and go -
e nmen s, on he o he hand, equi e s udies o immedia e and p ac ical
applica ion. As will be demons a ed la e in he chap e s o he echnical
con ibu ions, he e is no s a e-o - he-a o he ield o hose con ibu-
ions p oposed in his Thesis. These con ibu ions ocus no only on he
heo e ical amewo k, bu also on he applicabili y o he amewo k by
he ins i u ions.
31
Chap e 3
Imp o ing U ban
Accessibili y and
En i onmen al Condi ions
using G aph Modeling and
Mul i-objec i e
Op imiza ion
As commen ed in he in oduc o y chap e o his Thesis, u ban a eas a e
acing mul iple challenges ha may hinde hei social p og ess in he u-
u e. Among he p oblems ha ci ies su e om nowadays, accessibili y
is in he op ie . Accessibili y can be de ined as he oppo uni y o ac-
cess/ each communi y esou ces, se ices and ec ea ional acili ies [71].
Some a eas in ci ies howe e su e om lack o accessibili y. Mo eo e ,
o he conce ning issue is he age o he popula ion: he numbe o peo-
ple o e 65 yea s old will double in he nex h ee decades, eaching 16%
o he wo ld’s popula ion and exceeding 1500 million people in 2050 [72].
This las popula ion subs a e su e s he mos om he consequences o
he poo accessibili y o u ban a eas due o hei physical limi a ions. Ul-
ima ely, hei educed mobili y causes se ious social p oblems, including
isola ion, dependence, and ma ginaliza ion. O he challenges ha mod-
e n ci ies ace a e he ones ela ed wi h he en i onmen al aspec s o hei
s ee s. Today, all majo ci ies su e om ai pollu ion and high le els
o noise in public u ban spaces. Apa om he ob ious damaging con-
sequences o human heal h (gi ing ise o millions o dea hs yea ly [73]),
pollu ion also b ings a signi ican ly nega i e impac on he economy o
coun ies wo ldwide. Fo example, in 2018 alone he cos s o ai pollu ion
o he Na ional Heal h Se ice (NHS) and social ca e in England we e es-
ima ed o be £157 million, which could each an es ima e o £18.6 billion
by 2035 i no ac ions a e aken [74]. Fu he along his line, noise pollu ion
ep esen s o Eu opeans he i h mos signi ican p oblem (72%) wi hin
hei ci ies, jus second a e en i onmen al pollu ion (81%) [75]. Noise
32 Chap e 3. Imp o ing U ban Accessibili y and En i onmen al
Condi ions using G aph Modeling and Mul i-objec i e Op imiza ion
pollu ion, apa om i s inhe en annoyance, can cause ansi o y acu e
e ec s ( empo a y hea ing loss) and long- e m ch onic diseases such as de-
p ession, anxie y, and loss o sleep. In he medium-long e m, such diseases
can en ail se ious isks o heal h, such as o al hea ing loss o ca dio as-
cula diseases [76], [77]. Fu he mo e, o hese heal h- ela ed isks i mus
be added ha neighbo s o noisy u ban esiden ial a eas may su e also
om socio-beha io al modi ica ions [77]. These e ec s, in u n, a e mo e
likely o appea in ulne able g oups, such as people wi h heal h p oblems
and olde people.
Bea ing in mind his scena io, u ban planne s con on he challenge o
making ci ies iendlie , especially o he olde popula ions, who a e mos
a ec ed by he wo p oblems exposed abo e. Hence, u ban a eas should
be adap ed o accoun o he needs o olde people, making hem sa e ,
heal hie , and mo e inhabi able places o his g oup, i.e., hey should
become age- iendly [5], [78]. Wi h his objec i e in mind, in line wi h
he g owing in eg a ion o compu a ional ools in u ban planning, u ban
planne s and designe s ha e begun explo ing he po en ial o AI o enhance
decision-making p ocesses and de elop u ban plans ocused on imp o ing
accessibili y. This e lec s a b oade shi owa d he use o ad anced
echnologies, such as DSS and PSS, which, as men ioned ea lie in Chap e
1, con ibu e o mo e in o med and holis ic planning ou comes.
Howe e , gi en he gene ally es ic ed municipal public budge s, a
widely acknowledged decision d i e in u ban design decision-making is he
cos e ec i eness o any p oposed decision. Fo ins ance, when he p ocess
in ol es he modi ica ion o addi ion o u ban elemen s (e.g., mechanical
amps, escala o s, li s), decisions a e o en made based on es ima ions o
he imp o emen in e ms o accessibili y de i ed om he in e en ion.
Se e al p ac ical p oblems a ise along he decision-making p ocess: A)
in o ma ion sou ces needed o making an in o med decision a e di e se,
he e ogeneous and no consolida ed in o a single poin o in o ma ion; B)
he e a e no echnological means o quan i a i ely gauge he es ima ed
accessibili y and/o quali y imp o emen s o he u ban en i onmen co -
esponding o a ce ain decision, no any gua an ee ha he decision i -
sel is he bes one in e ms o cos e ec i eness; C) he consequences o
he in e play be ween objec i es ela ed o accessibili y and u ban qual-
i y h ough he cos o he in e en ion, by which he u ban planne is
unce ain whe he he addi ion o an elemen (e.g., a mechanical amp)
is he bes op ion ha can be made wi hou impac ing nega i ely on he
o he objec i es. As a esul , he p ocess is o en app oached by eso ing
o common sense, s a is ical s udies, audi s, o he expe knowledge o
expe ienced u ban planne s [79]–[81].
This chap e add esses his scena io depa ing om he obse a ion
ha he simul aneous conside a ion o di e en design objec i es in an
u ban planning p oblem can be ega ded as a mul i-c i e ia op imiza ion
p oblem. As such, he goal o his kind o p oblems is o disco e he se o
easible solu ions ha bes app oxima es he Pa e o ade-o be ween he
conside ed objec i es. Wi hin he con ex o his in es iga ion, a Pa e o-
op imal se o solu ions e e s o he se o possible u ban in e en ions
3.1. Rela ed Wo k 33
ha bes balance he economic cos o he in e en ion and nume ical
measu emen s o i s impac on he accessibili y, en i onmen al noise, and
ai pollu ion o he u ban a ea unde s udy.
The echnical con ibu ion o his chap e builds upon hypo hesizing
ha he use o a mul i-weigh ed g aph model o he u ban a ea, oge he
wi h mul i-objec i e e olu iona y op imiza ion, can be used o sol e he
issues A, B and C desc ibed p e iously. All hese ing edien s embody a
no el amewo k ha showcases he po en ial o e olu iona y compu a ion
o decision-making p ocesses o e u ban elemen s, aimed a imp o ing ac-
cessibili y and pa h condi ions o olde people. The amewo k elies on
a geo- e e enced g aph model o collec and cen alize all in o ma ion o
in e es o e a gi en u ban a ea. EAs o mul i-objec i e op imiza ion a e
u ilized o compu e an es ima ion o he Pa e o on d awn by he cos
o ins alla ion o di e en u ban elemen s (i.e., mechanical amps, escala-
o s, li s, acous ic panels), and hei impac in e ms o a el ime, ai
pollu ion educ ion and en i onmen al noise mi iga ion. Such objec i es
a e compu ed o e key (o igin,des ina ion) pai s ha ep esen poin s o
in e es (hospi al, day-ca e cen e s, o d ug s o es), also conside ing e-
alis ic models o hese ou objec i es. Expe imen al esul s in wo use
cases loca ed in he ci y o Ba celona (Spain) a e discussed o alida e he
applicabili y o he p oposed amewo k in eal-wo ld se ings.
3.1 Rela ed Wo k
Be o e p oceeding wi h he desc ip ion o he amewo k and he unde -
lying op imiza ion p oblem, his sec ion o e iews baseline li e a u e and
echnical ad ances epo ed a he c oss oads o u ban accessibili y and en-
i onmen al quali y (Subsec ion 3.1.1) and mul i-c i e ia decision-making
(Subsec ion 3.1.2). The sec ion ends by highligh ing he con ibu ion done
o he s a e-o - he-a (Subsec ion 3.1.3).
3.1.1 Accessibili y and En i onmen al Quali y o Age-
iendly Ci ies
Ci ies a e complex ecosys ems whe e mul iple dimensions coexis and in e -
ac [82], anging om physical elemen s such as buildings, in as uc u es,
and public spaces o socioeconomic ela ionships, cul u al exp essions, en-
i onmen al (ai , noise) and na u al elemen s (blue and g een a eas, o
auna, o men ion a ew). U ban quali y elies on mul iple aspec s: one
o hem is he accessibili y o places and se ices necessa y o daily li e
in a easonable ime, which has been la ely amed unde he “15-minu e
ci y” concep [83]. This is especially ue o ulne able popula ions, such
as olde people, which ha e g ea e mobili y di icul ies. The e o e, acces-
sibili y is a key aspec o age- iendly ci ies [5], [84].
Fu he mo e, i is impo an no only o be able o ge o places, bu
also o gua an ee he quali y o he i ine a ies. One o he goals o age-
iendly ci ies is o os e olde people o s ay ac i e o as long in li e
as possible. The e o e, i is key o o e i ine a ies ha a e a ac i e and

34 Chap e 3. Imp o ing U ban Accessibili y and En i onmen al
Condi ions using G aph Modeling and Mul i-objec i e Op imiza ion
sa e, o induce people o walk a ound e e y day. The capaci y o a place o
imp o e people’s well-being ela es o he concep o a es o a i e en i on-
men , i.e., en i onmen s enhancing o acili a ing psychological es o a ion,
and hus con ibu ing o human heal h and well-being [85]. Among many
o he ac o s, acous ic com o has been p o en o g ea ly con ibu e o
en i onmen al quali y [86]. The e o e, i is a dimension ha should be
conside ed when designing age- iendly i ine a ies. Ai quali y is ano he
key aspec , g ea ly con ibu ing o u ban heal h indica o s [87], wi h olde
people being especially sensi i e o he e ec s o b ea hing pollu ed ai .
3.1.2 Mul i-c i e ia Design and Mul i-objec i e Op i-
miza ion
Mul i-c i e ia analysis (MCA) has been a widely used ool as a suppo
o decision-making when ca ying ou an in e en ion in u ban con ex s
[21], [88]–[90]. MCA is used o make a compa a i e disc imina ion be-
ween he e ogeneous p ojec s, measu emen s o ake and/o decisions o
be made. To his end, solu ions ha can be conside ed a e compa ed and
e alua ed sys ema ically acco ding o he case. The MCA app oach mus
be nou ished by a la ge amoun o da a, as well as by he p e e ences o
he decision-make s, which e en ually de e mine he speci ic weigh gi en
o one o ano he c i e ion [91]. In some cases, MCA can be combined
wi h simula ion engines o a highe deg ee o obus ness in he esul s
[92]. Howe e , his app oach is usually e y ime-consuming, since di e -
en possibili ies need o be e alua ed o di e se scena ios [91], [93], and
he app oach does no ensu e he op imali y bu a he he easibili y o
he plan [94].
Acco ding o he e iew in Chap e 2, he mul i-c i e ia design o u ban
egene a ion plans can be sol ed wi h MOEAs which allow e icien ly ex-
plo ing he se o possible solu ions (u ban in e en ions), and e en ually
app oxima ing he Pa e o on . As a esul , decision-make s a e in o med
wi h a highe numbe o possible ac ions o choose om [94]. MOEAs a e
used in di e en ields o esea ch, including chemis y [95], bio-in o ma ics
[96], business analy ics [97] o en i onmen al sus ainabili y [98], o men-
ion a ew. In u ban planning se e al ac o s gene ally come in o play,
which ul ima ely leads o an op imiza ion p oblem comp ising di e en
objec i es. Se e al examples o he use o MOEAs o make an op imal
use o land can be ound in he li e a u e. Fo example, he wo k in [99]
eso s o gene ic algo i hms o es ima e he u u e use o land in a ci y,
conside ing h ee objec i es: 1) he minimiza ion o he cos o he in e -
en ion; 2) he minimiza ion o a ic conges ion; and 3) he minimiza ion
o he numbe o changes o be made o e al eady buil -up a eas o he
ci y. Masoumi e al. seek in [100] a di e en use o he land, namely, o
c ea e an indus ial complex o e he ou ski s o a ci y. Fo ha pu pose,
hey o mula e di e en c i e ia han hose in [99], such as he dis ance
o he nea es ci y, he oad accessibili y, he a ailabili y o esou ces and
he dis ance o en i onmen ally p o ec ed a eas and o es s. They also
3.1. Rela ed Wo k 35
u ilize gene ic algo i hms o e icien ly deal wi h he op imiza ion p ob-
lem s emming om hei s udy. In a di e en ein, Ma ias Pé es e al.
[101] p opose o use MOEAs o educe a ic conges ion in Mon e ideo
(U uguay), op imizing he cycles o a ic ligh s in he ci y, leading o a
clea educ ion o he ai pollu ion in he ci y. In he nex subsec ion, we
ocus ou e iew on mul i-objec i e op imiza ion applied o he imp o e-
men o u ban accessibili y o olde people, add essing he h ee issues
discussed in he in oduc ion: noise pollu ion, ai pollu ion and pedes ian
accessibili y.
3.1.3 Mul i-objec i e Op imiza ion o U ban Acces-
sibili y
E en hough noise, ai pollu ion and accessibili y a e majo issues ha
mode n ci ies encoun e nowadays, e y ew examples can be ound whe e
op imiza ion, – nei he mul i-c i e ia no single-objec i e – has been em-
ployed o make u ban decisions o deal wi h hese p oblems. When consid-
e ing he speci ics o age- iendliness (e.g., educed mobili y), he li e a u e
is e en mo e sca ce.
To ackle he p oblem o noise pollu ion, one o he mos used ech-
niques is o analyze he p opaga ion o noise h ough he di e en s ee s,
h ough simula ions ha conside he geome y o he s ee s and di e en
elemen s ha impac he pe cep ion o noise, such as acades, he dis ance
o oads o he p esence o ege a ion [102], [103]. When i comes o he
adop ion o mul i-objec i e op imiza ion o deal wi h noise pollu ion, Ham-
mad e al. [104] add ess a mul i-objec i e acili y p oblem, whe e he aim
is o simul aneously minimize noise pollu ion and imp o e a el imes.
To do his, hey disc imina e be ween wo ypes o buildings: hose ha
gene a e noise (indus ies, shopping malls), and hose ha a e sensi i e o
noise pollu ion (hospi als, schools). They use g aph heo y o apply hei
op imiza ion, whe e he nodes a e he buildings, and he edges a e he
oads ha connec hem. They seek o ind he op imal con igu a ion ha
minimizes noise pollu ion and a el imes be ween nodes. Likewise, Ning
e al. [105] apply a hyb id gene ic algo i hm –ACO algo i hm– o educe
noise pollu ion in cons uc ion si es. The s udy conside s h ee di e en
objec i es o op imize: noise pollu ion o di e en cons uc ion asks, he
o e all anspo a ion cos , and po en ial isk ela ed o in e ac ion low
be ween he acili ies.
Rega ding ai pollu ion in u ban a eas, he end is o exploi he po en-
ial o da a-based modeling o o ecas pollu ion, so p edic i e es ima ions
can in o m subsequen decision-making p ocesses [106]–[108]. Howe e ,
s udies ha e been also epo ed using op imiza ion o deal wi h ai pol-
lu ion. An example is he wo k by Wang e al. in [109], which p oposes
a mul i-objec i e op imiza ion algo i hm o build an ai pollu ion ea ly
wa ning sys em which o ecas s pollu ion le els and calcula es ai quali y
based on ha o ecas . O he au ho s ha e ocused on p o iding di e -
en solu ions o challenge he g ow h o ai de e io a ion in u ban a eas.
The a o emen ioned wo k published in [101] aims o imp o e a ic lows
36 Chap e 3. Imp o ing U ban Accessibili y and En i onmen al
Condi ions using G aph Modeling and Mul i-objec i e Op imiza ion
o e ci ies, which is one o he majo causes o he inc ease in pollu-
ion. Di e en ly, he wo k in [110] ocuses a he on u ban spaces, using
mul i-objec i e e olu iona y op imiza ion o minimizing u ban ecosys em
se ices loss and maximizing he compac ness in planning new ci y de el-
opmen s, using Singapo e as a case s udy.
U ban accessibili y enjoys a g ea e bibliog aphical collec ion, wi h se -
e al examples o op imiza ion p oblems o mula ed o deal wi h his im-
po an d i e o age- iendly ci ies. Unlike in esea ch on noise and ai
pollu ion, in u ban accessibili y issues he mos ulne able g oups such as
olde people o people wi h educed mobili y mus be conside ed.
A widely used echnique no only in accessibili y p oblems bu also in
o he u ban- ela ed ques ions is he use o geo e e enced g aphs. G aphs
allow ep esen ing he u ban a ea unde s udy and embedding map in-
o ma ion and da a in a geo e e enced manne (pollu ion, noise, slope o
he s ee s, e c.). Fo ins ance, D´O so e al. [111] p opose a ool ha
helps o decide which u ban in e en ions a e o g ea e impo ance o
pedes ian walkabili y. To do his, hey ep esen he pedes ian ne wo k
wi h a ches. A quali y index is a ibu ed o each o he a ches associa ed
wi h i s walkabili y. To decide which a cs should be p io i ized o hei
imp o emen , bo h he a o emen ioned index and he usage equency o
he a ches a e conside ed. Based on hese ac o s, sho es ou es be ween
o igin nodes ( ain s a ions in his s udy) and a ac o s o des ina ion
nodes (places wi h high pedes ian in lux) a e compu ed.
A well-es ablished me hodology o use mul i-objec i e op imiza ion o
u ban accessibili y is o combine MOEAs wi h geo- e e enced g aphs. This
combina ion is p esen , o ins ance, in he wo k o Blecic e al. [112],
which concei es a suppo ool o u ban planne s based on g aph model
combined wi h he NSGA-II algo i hm. The ool compu es a Pa e o on
app oxima ion be ween he walkabili y imp o emen and he implemen-
a ion e o . They compu e he walkabili y cos o all he edges o he
g aph, which ep esen s ee segmen s, based on se e al pa ame e s ha
impac on walkabili y. They aim o educe cos s de i ed om he applica-
ion o imp o emen s o such s ee segmen s. Hence, he decision make
is o e ed di e en al e na i es. In he same way, Salcedo-Sanz e al. [113]
combine di ec ed g aphs ha ep esen he oad ne wo k o a Spanish
ci y wi h a MOEA o imp o e a ic low when he e a e punc ual a ic
cu s in ce ain zones o he ci y. They ob ain a Pa e o on whe e he
di e en solu ions a e he possible ou es om he inpu o he ou pu
nodes. When i e e s o he accessibili y o people wi h educed mobili y
(such as wheelchai use s o olde people), se e al s udies ha e adop ed
his g aph-based amewo k. Fo ins ance, Bol en e al. [114] u ilize g aph
heo y o ep esen and cha ac e ize he s ee s o an u ban a ea in e ms
o wheelchai accessibili y. Following his end, Ba czyszyn e al.[115]
ha e c ea ed a collabo a i e ou e planning se ice o wheelchai use s,
based on sidewalk-based model. Fo ha , hey also use a g aph-model o
desc ibe an a ea whe e he sidewalks and c osswalks a e ep esen ed by he
edges and he junc ions on he sidewalks by he nodes o he g aph. Those
3.1. Rela ed Wo k 37
edges a e weigh ed acco ding o some ea u es such as he slope, he exis-
ence o cu b amps on c osswalks and he main enance o he sidewalks.
Based on ha g aph, he ou planne algo i hm o e s al e na i e acces-
sible ou es o wheelchai use s. Simila ly, Rahaman e al. [116] de ise a
Con ou -based Accessible Pa h Rou ing Algo i hm (CAPRA) ha hinges
on he well-known A∗heu is ic o compu e se e al al e na i e ou es ha
enhance he accessibili y o olde people. Ano he wo k in his di ec ion is
he one published by Sasaki e al. in [117]. Speci ically, hey add ess he
disco e y o ou es o a walk conside ing measu es o he sa e y, ameni y
and walkabili y o a use a e sing hem. Algo i hmically he s udy com-
bines he A∗algo i hm and a GA o yield sub ou es ha maximize such
measu es. Al e na i e ypes o u ban in e en ions a e also o g ea im-
po ance o c ea e accessible ci ies, mainly o olde people. Fo example,
Zhang e al. [118] p opose MOEAs o decide he loca ions o heal hca e
cen e s ha maximize accessibili y o e e yone, co e ing as many people
as possible wi hin an accep able dis ance adius om hem, and also min-
imizing he cos s o he p ocess. Mo e ecen ly, he wo k in [119] explo es
he use o MOEAs o ind cos -e ec i e u ban in e en ions ha balance
accessibili y and cos . Howe e , no o he objec i es a e conside ed in his
s udy, no does i analyze quali a i ely whe he he op imized in e en-
ions co espond o decisions ha ac ually make sense om he p ac ical
poin o iew.
Con ibu ion. In ligh o he abo e li e a u e e iew, o he bes o
ou knowledge no p e ious wo k has so a p oposed he imp o emen
o u ban accessibili y and quali y conside ing h ee majo p oblems ha
me opolises su e oday. In gene al, noise, pollu ion and accessibili y
p oblems a e ea ed independen ly. Howe e , o olde people, hese h ee
ac o s a e o i al impo ance when choosing one ou e o ano he . The e-
o e, hey should no be ackled exclusi ely, bu join ly. Mo eo e , he
o e all cos o he u ban in e en ion also induces an in e play be ween
such objec i es, making i unclea whe he he decision make should in-
es in he ins alla ion o an u ban elemen o ano he . Fo his eason, we
ex end he amewo k in [119] o accoun o he h ee ac o s men ioned
p e iously, owa ds inding u ban in e en ions ha bes balance (in he
Pa e o sense) be ween he noise and ai pollu ion pe cei ed by a use going
h ough hem, his/he accessibili y gi en he opog aphical cha ac e is ics
o he pa hs, and hei economic cos . This amewo k, which is desc ibed
in he nex sec ion, combines geo e e enced g aphs wi h MOEAs o de e -
mine whe e o ins all u ban elemen s aimed o imp o e noise, ai pollu ion
and accessibili y (e.g., escala o s, ele a o s, g een o acous ic panels), in
o de o imp o e u ban accessibili y and en i onmen al quali y and ul i-
ma ely, gi e ise o a mo e age- iendly ci y.
44 Chap e 3. Imp o ing U ban Accessibili y and En i onmen al
Condi ions using G aph Modeling and Mul i-objec i e Op imiza ion
yielding:
𝑓𝑛𝑜𝑖𝑠𝑒 (𝛀|V𝑜,V𝑑)
=1
|V𝑜||V𝑑|∑︁
𝑣𝑜∈V𝑜∑︁
𝑣𝑑∈V𝑑∑︁
(𝑣,𝑣′)∈E𝑜,𝑑
𝑓𝑛𝑜𝑖𝑠𝑒 (𝑣, 𝑣′, 𝑡(𝑣, 𝑣′), 𝜓(𝑣, 𝑣′)),(3.19)
𝑓𝑎𝑐𝑐 (𝛀|V𝑜,V𝑑)
=1
|V𝑜||V𝑑|∑︁
𝑣𝑜∈V𝑜∑︁
𝑣𝑑∈V𝑑∑︁
(𝑣,𝑣′)∈E𝑜,𝑑
𝑓𝑎𝑐𝑐 (𝑣, 𝑣′, 𝑡(𝑣, 𝑣′), 𝜓(𝑣, 𝑣′)).(3.20)
Finally, he cos objec i e ollows s aigh o wa d om he sum o he
cos s associa ed wi h each asse o he in e en ion:
𝑓𝑐𝑜𝑠𝑡 (𝛀)=
𝐾
∑︁
𝑘=1
𝑓𝑐𝑜𝑠𝑡 (𝑣𝑘, 𝑣′
𝑘, 𝑡(𝑣𝑘, 𝑣′
𝑘), 𝜓(𝑣𝑘, 𝑣′
𝑘)),(3.21)
whe e, clea ly, 𝑓𝑐𝑜𝑠𝑡 (𝑣𝑘, 𝑣′
𝑘, 𝑡(𝑣𝑘, 𝑣′
𝑘), 𝜓(𝑣𝑘, 𝑣′
𝑘)) =0i 𝑡(𝑣𝑘, 𝑣′
𝑘)=∅. Based
on hese de ini ions, he mul iobjec i e op imiza ion p oblem o be sol ed
seeks a se o solu ions (in e en ions) ha bes balance hese 4 objec i es.
Since se e al objec i es a e conside ed, his se o solu ions mus be e al-
ua ed in e ms o hei Pa e o op imali y: solu ions o be p oduced by he
amewo k mus delinea e he ade-o be ween he imp o emen in e ms
o accessibili y and en i onmen al condi ions enabled by he in e en ions
(gi en by he h ee i s objec i es) and he economic cos o be in es ed
o implemen he in e en ion o e he u ban a ea unde s udy ( ou h ob-
jec i e). Such objec i es a e con lic ing wi h each o he : he mo e asse s
a e o be deployed o enhance he accessibili y and enjoyabili y o olde
pedes ians in hei ou es, he highe he economic cos o he in e en ion
will be. Solu ions o be disco e ed by he amewo k mus ensu e ha an
imp o emen in any o he objec i es would equi e a deg ada ion o he
o he s.
The e o e, he op imiza ion p oblem can be o mula ed as he disco -
e y o di e se in e en ions ha app oxima e he Pa e o-op imal balance
be ween he ou conside ed objec i es. The p oblem is augmen ed by in-
se ing echnical limi a ions associa ed wi h each asse , so ha he leng h
o amps, escala o s, sound and pollu ion panels a e kep below ce ain
limi s (𝑙𝑅
𝑚𝑎𝑥,𝑙𝐸
𝑚𝑎𝑥,𝑙𝑆𝑃
𝑚𝑎𝑥, and 𝑙𝑃𝑃
𝑚𝑎𝑥, espec i ely). Ma hema ically:
min
𝐾𝑝,{𝛀𝑝}𝐾𝑝
𝑘=1
𝑓𝑎𝑐𝑐 (𝛀𝑝|V𝑜,V𝑑), 𝑓𝑎𝑖𝑟 (𝛀𝑝|V𝑜,V𝑑), 𝑓𝑛𝑜𝑖𝑠𝑒 (𝛀𝑝|V𝑜,V𝑑), 𝑓𝑐𝑜𝑠𝑡 (𝛀𝑝),(3.22)
subjec o:
V𝑜(se o o igin nodes), V𝑑(se o des ina ion nodes), (3.23)
𝜆𝑎𝑐𝑐 +𝜆𝑎𝑖𝑟 +𝜆𝑛𝑜𝑖𝑠𝑒 =1,(3.24)
Ψ(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘) · 𝑙(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘)< 𝑙𝑡(𝑣𝑝
𝑘,𝑣′𝑝
𝑘)
𝑚𝑎𝑥 ,∀𝑝∈ {1, . . . , 𝑃}
and ∀𝑘∈ {1, . . . , 𝐾𝑝}:𝑡(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘) ∈ {𝑅, 𝐸, 𝑆𝑃, 𝑃𝑃},(3.25)

3.3. Sea ch Me hodology 45
whe e index 𝑝in 𝛀𝑝,𝑣𝑝
𝑘and 𝑣′𝑝
𝑘 e e s o he 𝑝- h in e en ion o he
Pa e o-app oxima ing se o in e en ions sough in he abo e p oblem
o mula ion. Decision a iables o be disco e ed include, he e o e, he
numbe o asse s 𝐾𝑝in ol ed in e e y in e en ion 𝛀𝑝o he Pa e o-
app oxima ing se , he edges (𝑣𝑝
𝑘, 𝑣′𝑝
𝑘)in which he 𝑘- h asse in 𝛀𝑝is in-
s alled, i s ype 𝑇(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘)and cha ac e is ics ( ela i e leng h) Ψ(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘).
The mix u e o eal- and in ege - alued decision a iables in ol ed in he
abo e o mula ion mo i a es he use o mul i-objec i e e olu iona y algo-
i hms o e icien ly explo e he space o possible solu ions. De ails on
how his sea ch o Pa e o-op imal in e en ions is accomplished by he
amewo k a e gi en in he nex sec ion.
3.3 Sea ch Me hodology
To e icien ly explo e he sea ch space spanned by he decision a iables
o he p oblem o mula ed p e iously, he amewo k eso s o MOEAs.
As discussed in Chap e 2 MOEAs comp ise a b anch o me a-heu is ic
op imiza ion esea ch ha ocuses on he exploi a ion o concep s om
he heo y o e olu ion, including b eeding, mu a ion, and su i al o he
i es [121], [122]. The emula ion o hese na u al p ocesses as sea ch op-
e a o s wi hin an i e a i e compu e p og am gi es ise o app oxima e
g adien - ee sol e s ha do no equi e any speci ic ma hema ical knowl-
edge o he objec i es o be minimized. In doing so, he sea ch algo i hm
ea u ing e olu iona y ope a o s i e a i ely ailo s and checks o he i -
ness o candida e solu ions (indi iduals) o he p oblem a hand, e aining
he mos p omising ones in a popula ion. By epea edly applying ope a-
o s o he popula ion o indi iduals, mo e e ined solu ions a e p oduced
o e i e a ions (gene a ions). Once a s op c i e ion is me (e.g., a ixed
numbe o gene a ions), he bes solu ion in he popula ion is e u ned as
he solu ion o he p oblem.
As in any o he mul i-objec i e p oblem ackled wi h MOEAs, c i ical
design choices o he algo i hm a e he solu ion encoding (namely, how so-
lu ions o he p oblem a e nume ically ep esen ed wi hin he e olu iona y
sea ch) and he sea ch ope a o s (i.e., how indi iduals a e e ined and e-
ained in he popula ion du ing he e olu iona y sea ch). These elemen s
o he MOEAs conside ed in he p oposed amewo k a e de ailed below.
3.3.1 Solu ion Encoding
The solu ion encoding (geno ype) is designed in close esemblance wi h he
pheno ype imposed by he p oblem. I is ecalled ha e e y indi idual o
be e ined du ing he e olu iona y sea ch mus ep esen an in e en ion
𝛀𝑝, gi en by:
𝛀𝑝=[𝑣𝑝
𝑘, 𝑣′𝑝
𝑘, 𝑡(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘), 𝜓(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘)]𝐾𝑝
𝑘=1,(3.26)
namely, a s uc u e di ided in 𝐾𝑝asse s, each indica ing he edge (𝑣𝑝
𝑘, 𝑣′𝑝
𝑘)
on which he 𝑘- h asse is ins alled, i s ype and cha ac e is ics. To ease
46 Chap e 3. Imp o ing U ban Accessibili y and En i onmen al
Condi ions using G aph Modeling and Mul i-objec i e Op imiza ion
he applica ion o sea ch ope a o s and e ec i ely a oid a iable-leng h
geno ypes, a maximum numbe o asse s pe in e en ion 𝐾𝑚𝑎𝑥 is assumed
and an addi ional ype o asse (i.e., 𝑡(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘) ∈ {𝑅, 𝑆, 𝐸, 𝑆𝑃, 𝑃𝑃, ∅}) is
inse ed o accoun o he case whe e no asse is ins alled. In his way,
he e ec i e numbe o asse s o a gi en in e en ion can be compu ed as:
𝐾𝑝=
𝐾𝑚𝑎𝑥
∑︁
𝑘=1
I(𝑡(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘)≠∅),(3.27)
whe e I(·) is an auxilia y indica o unc ion aking alue 1 i i s a gumen
is ue (and 0 o he wise). I is impo an o no e ha his solu ion en-
coding s a egy pe mi s o de ine se e al di e en asse s o e a gi en edge
(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘), which should be allowed only when he ype o asse s o be de-
ployed on he edge a e compa ible wi h each o he . Fo ins ance, sound
panels (𝑇(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘)=𝑆𝑃) and amps (𝑇(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘)=𝑅) can be simul ane-
ously ins alled on he same edge. Howe e , a amp and mechanical s ai s
(𝑇(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘)=𝑆) on he same gi en edge make no sense om he p ac ical
poin o iew. Fo his eason, each newly p oduced indi idual is checked
o hese incompa ibili ies, e aining (when needed) he asse implying a
lowe ins alla ion cos .
3.3.2 Sea ch Ope a o s
Following he algo i hmic p inciples o MOEAs, he i e a i e e inemen o
he solu ions by he p oposed amewo k equi es he de ini ion o se e al
ope a o s o 1) selec indi iduals om he popula ion o b eeding (pa -
en s); 2) mix hei encoded ep esen a ions o yield new o sp ing solu ions
(c osso e ); 3) mu a e he o sp ing o en o ce geno ypical di e si y wi h
espec o hei pa en indi iduals (mu a ion); and 4) e ain in he popula-
ion hose indi iduals ha a e be e as pe he i ness o he op imiza ion
p oblem a hand (su i o selec ion). Since dealing wi h a mul i-objec i e
p oblem, he i ness o indi iduals mus be assessed in e ms o Pa e o
op imali y. The de ini ion o he c i e ion by which one indi idual is be -
e han ano he in a space comp ising se e al objec i es has spu ed a
lu y o MOEA p oposals in he ela ed li e a u e, ho oughly e iewed
in se e al su eys on he ma e .
This being said, he amewo k conside s ou di e en MOEAs ha
ha e been widely p o en o adap and pe o m compe i i ely in mul i-
objec i e op imiza ion p oblems wi h mixed- ype decision a iables: NSGA-
II [38], NSGA-III [39], SPEA2 [40], and Mul iObjec i e Cellula gene ic
algo i hm (MOCELL [123]). Fo he sake o eadabili y, de ails on hei
di e en ial cha ac e is ics a e b ie ly e isi ed he e, al hough hey we e
de ailed in Chap e 2. Howe e , MOCELL is a new in oduc ion in his
con ex .
•NSGA-II e alua es he Pa e o op imali y o newly p oduced solu ions by
so ing hem oge he wi h indi iduals al eady con ained in he popula-
ion as pe hei Pa e o dominance (non-dominance ank) and di e si y
3.3. Sea ch Me hodology 47
inside e e y ank. In doing so, old and new solu ions a e so ed acco d-
ing o an ascending le el o non-domina ion in he space o objec i es. I
he e is no oom in he popula ion o include all indi iduals wi hin he
las u ilized ank, an es ima o o he densi y o o he solu ions a ound
e e y solu ion wi hin his ank is used o selec hose indi iduals ha
a e loca ed in less dense a eas o he objec i e space (hence, hey a e
mo e di e se). The seminal e sion o he NSGA-II emb aces he so-
called c owding dis ance unc ion as a densi y es ima o : howe e , many
o he de ini ions o his es ima o can be adop ed ins ead.
•NSGA-III can be ega ded as an ex ension o he NSGA-II algo i hm
ha is adap ed o p o ide a be e pe o mance in p oblems comp ising
mo e han wo objec i es. The main di e ence lies on he popula ion
di e si y, which is gua an eed by means o e e ence di ec ions, i.e., a
se o poin s used o guide he op imiza ion p ocess owa ds a speci ic
egion o he objec i e space. Re e ence di ec ions di ide he space o
objec i es in a numbe o sub- egions (di isions), so ha each di ision
is associa ed wi h a speci ic e e ence di ec ion. These di isions a e used
by NSGA-III o d i e i s sea ch owa ds egions o he objec i e space
ha a e mos p omising. Hence, c owding dis ance is no used as a den-
si y es ima o o e ain indi iduals in he popula ion. Ins ead, a dual
c i e ion based on non-domina ed so ing and en i onmen al selec ion is
used, he la e aimed o main ain di e si y among he solu ions main-
ained in he popula ion. Since solu ions e enly dis ibu ed along he
Pa e o-op imal on should be p e e ed, NSGA-III p io i izes solu ions
ha a e close o he bounda y be ween sub- egions o di isions, as hese
a e mo e ep esen a i e o he di e en pa s o he Pa e o-op imal on
as pe he con igu ed e e ence di ec ions.
•SPEA2 wo ks in he same line as NSGA-II bu associa es a new measu e
o i ness (s eng h) wi h e e y indi idual 𝛀𝑝in he popula ion. The
s eng h indica es he numbe o o he indi iduals in he popula ion
ha a e domina ed by he indi idual a hand conside ing hei alues
o he de ined objec i es, no malized by he size o he popula ion 𝑃.
This de ined s eng h is hen used o compu e and assign a aw i ness
o e e y solu ion based on he s eng h o solu ions ha domina e i . To
disc imina e be ween indi iduals ea u ing he same aw i ness alue, a
measu e o local densi y is calcula ed in e ms o he dis ance o e e y
indi idual o hei neighbo s. Fu he mo e, SPEA2 u ilizes an ex e nal
eposi o y o solu ions (a chi e) whe e non-domina ed indi iduals ound
du ing he sea ch p ocess a e s o ed. The a chi e size is kep ixed
h oughou he sea ch by i ue o a unca ion ope a o .
•MOCELL de ines he popula ion as a egula g id ha d i es he p o-
cess o selec ing indi iduals o b eeding: only solu ions ha belong o
he same neighbo hood in he g id can be ecombined o yield a new
o sp ing. Di e en opologies can be es ablished o he neighbo hood,
among which he o iginal MOCELL e sion p oposed in [123] used he
Moo e opology ( he neighbo hood o a solu ion loca ed in a cell inside
he g id is i sel and he solu ions o he eigh cells su ounding i in
48 Chap e 3. Imp o ing U ban Accessibili y and En i onmen al
Condi ions using G aph Modeling and Mul i-objec i e Op imiza ion
a wo-dimensional g id). As in SPEA2, MOCELL employs an ex e nal
a chi e o s o e non-domina ed solu ions disco e ed du ing he sea ch
p ocess. Howe e , such a chi ed indi iduals a e ed back o he popu-
la ion, eplacing o he solu ions selec ed a andom. C owding dis ance
is used as he densi y es ima o o bo h he popula ion ( o selec ion)
and he a chi e ( o unca ion).
As can be in e ed om he abo e desc ip ion, he algo i hmic di e -
ences be ween such MOEAs do no a ec he e olu iona y ope a o s used
o ecombine and mu a e newly imp o ised solu ions along he sea ch.
Ins ead, such di e ences concen a e mainly on he c i e ia used o se-
lec indi iduals o b eeding and o e ain good solu ions o e he sea ch.
The e o e, he same p ocedu es o ecombina ion and mu a ion can be
adop ed o he h ee o hem, a o ing a ai algo i hmic compa ison be-
ween hem. An adap ed e sion o he Simula ed Bina y C osso e (SBX)
ope a o [124] wi h p obabili y 𝑃𝑐and dis ibu ion index 𝐼𝐷a e used. Fo
in ege a iables (𝑣𝑝
𝑘,𝑣′𝑝
𝑘,𝑡(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘)) a andom mu a ion wi h p obabili y
𝑃𝑖𝑛𝑡
𝑚is u ilized, such ha he alue o e e y a iable is d awn a an-
dom om i s co esponding alphabe (ei he he exis ing edges (𝑣𝑝
𝑘, 𝑣′𝑝
𝑘)
in he g aph G ep esen ing he u ban a ea unde s udy, o he ypes o
asse s ha can be deployed in he edge a hand). Fo eal- alued a iables
(Ψ(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘)) a polynomial mu a ion wi h p obabili y 𝑃𝑟𝑒𝑎𝑙
𝑚applied o e -
e y pa o he solu ion. Finally, all MOEAs u ilize a bina y ou namen
selec ion me hod o selec which indi iduals o b eed om he popula-
ion (NSGA-II, NSGA-III, SPEA2) o om he neighbo hood o he gi en
cell in he cellula g id (MOCELL). The amewo k can seamlessly accom-
moda e o he e olu iona y sea ch ope a o s sui ed o he mixed ype o
decision a iables de ined in he p oblem s a emen .
3.3.3 O e all Algo i hmic Flow o he F amewo k
Once he sea ch p ocedu es ha e been de ined, now all he s eps needed
o he applica ion o he p oposed amewo k o he imp o emen o he
accessibili y and en i onmen al condi ions o olde pedes ians in a new
a ge a ea 𝐴can be de ined. Such s eps a e desc ibed in Algo i hm 1.
To begin wi h, he s a ing poin is he de ini ion o he a ea unde
s udy, which can be speci ied in se e al ways by he use o he ame-
wo k (e.g., as a bounding box delimi ed by geog aphical coo dina es). I
is necessa y ha ai and noise pollu ion measu emen s o e he a ea unde
s udy a e a ailable so as o ha e a ough es ima ion o he pollu ion p esen
on e e y edge o i s g aph. The use should also speci y se e al poin s o
in e es (bo h o igin and des ina ion) o olde pedes ians in 𝐴, so ha
u ban ou es connec ing such poin s o in e es a e u ilized o he com-
pu a ion o he objec i es o be op imized. O he equi emen s o be me
be o ehand also include he echnical speci ica ions o he asse s ha can
be ins alled in he a ea (including a enua ion o he noise/ai pollu ion
in ensi y achie ed by panels, nominal speeds o mechanical accessibili y
asse s, hei maximum ins alla ion leng h and economical cos models).
Such equi emen s can be que ied o echnological supplie s o he asse s.
3.3. Sea ch Me hodology 49
The use o he amewo k mus also indica e he ela i e impo ance o
accessibili y (𝜆𝑎𝑐𝑐) and he en i onmen al condi ions (𝜆𝑎𝑖𝑟 and 𝜆𝑛𝑜𝑖𝑠𝑒) in
he disco e y o op imal ou es be ween he es ablished o igin and des ina-
ion poin s o in e es . This can be easily in oduced by he use by using a
iangula sys em o ba ycen ic coo dina es, ensu ing au oma ically ha
any poin inside he iangle sa is ies 𝜆𝑎𝑐𝑐 +𝜆𝑎𝑖𝑟 +𝜆𝑛𝑜𝑖𝑠𝑒 =1.
Once hese equi emen s ha e been ul illed, he amewo k s a s by
compu ing he g aph G={V,E} co esponding o he a ea 𝐴(line 1
in Algo i hm 1). The simples e sion o his g aph can be cons uc ed
om GIS da abases by de ising e ices o nodes Vas s ee in e sec-
ions/junc ions, and by ep esen ing s ee s hemsel es as edges E. The
use o GIS da abases can allow o he s aigh o wa d anno a ion o UTM
and ele a ion da a o e e y e ex in V. Mo e de ailed e sions o his
g aph can be achie ed by using spa ially collec ed ine-g ained geog aph-
ical da a om o he in o ma ion sou ces, such as LiDAR (Ligh De ec ion
and Ranging) measu emen s. Once g aph Ghas been cons uc ed, he use
inse s he o igin and des ina ion nodes o he ou es o be la e op imized
(line 2), and es ablishes he nominal noise and ai pollu ion le els o e e y
edge in G(line 3). Fo his la e pu pose, measu emen s aken by pub-
lic anspo se ices and/o ixed ai /sound quali y moni o ing s a ions
can be used, oge he wi h me hods o ex apola e spa ially he pollu ion
le els eco ds o non-moni o ed egions o he u ban a ea. As a esul o
his ex apola ion, nominal pollu ion le els o e e y edge in he g aph a e
compu ed.
A e ini ializing a andom a 𝑃-sized popula ion o in e en ions (line
4), he amewo k e alua es he objec i e alues associa ed wi h hese
ini ial guesses o he Pa e o-op imal in e en ions. To his end, asse s
wi hin each in e en ion in he popula ion a e deployed on g aph G(line
6), so ha weigh s e lec ing he accessibili y and he en i onmen al con-
di ions o e e y edge esul ing om he ins alla ion o such asse s can be
e eshed (line 7). The g aph wi h upda ed edge weigh s can be used o
compu e he op imal ou e be ween e e y pai o o igin-des ina ion nodes
(𝑣𝑜, 𝑣𝑑) ∈ V𝑜× V𝑑 ia he Dijks a algo i hm [125] (line 9), and e alua e
he alue o he accessibili y, ai and noise pollu ion associa ed o e e y
ou e (line 10). By a e aging such objec i es o e all ou es (line 11),
each in e en ion 𝛀𝑝is e alua ed in e ms o he objec i es de ined o
he p oblem. The ou h objec i e (economic cos o he in e en ion) is
compu ed as in Exp ession (3.21) (line 12).
The es o he s eps in Algo i hm 1(lines 13 o 20) comp ise he
e inemen o he ini ial se o in e en ions by he chosen MOEA algo-
i hm. This e olu iona y sea ch consis s o he i e a i e applica ion o he
a o emen ioned selec ion (line 15), c osso e and mu a ion ope a o s (line
16) on he indi iduals in he popula ion o e 𝑚𝑎𝑥𝐺𝑒𝑛 gene a ions. Be-
o e il e ing and e aining good indi iduals in he popula ion (line 18),
newly p oduced in e en ions a e e alua ed (line 17) based on he de ined
se o o igin and des ina ion poin s o in e es , ollowing he p ocedu e
explained p e iously (lines 6-12). Once 𝑚𝑎𝑥𝐺𝑒𝑛 gene a ions o his e o-
lu iona y sea ch p ocess ha e been comple ed, he amewo k e u ns he

50 Chap e 3. Imp o ing U ban Accessibili y and En i onmen al
Condi ions using G aph Modeling and Mul i-objec i e Op imiza ion
Algo i hm 1: Algo i hmic low o he p oposed amewo k
Inpu : U ban a ea unde s udy 𝐴,[𝜆𝑎𝑐𝑐, 𝜆𝑛𝑜𝑖𝑠𝑒, 𝜆𝑎𝑖𝑟 ](weigh s
indica ing he ela i e impo ance o accessibili y and
en i onmen al condi ions when ou ing pedes ians),
[𝐶𝑓 𝑖𝑥 (𝑡), 𝐶𝑣𝑎𝑟 (𝑡)] ∀𝑡∈ {𝑅, 𝑆, 𝐸 , 𝑆𝑃, 𝑃𝑃}(cos model o
e e y asse ype), 𝐾𝑚𝑎𝑥 (maximum numbe o asse s pe
in e en ion), 𝑃𝑐,𝑃𝑖𝑛𝑡
𝑚,𝑃𝑟𝑒𝑎𝑙
𝑚,𝑃(MOEA pa ame e s),
𝑚𝑎𝑥𝐺𝑒𝑛 (maximum numbe o sea ch gene a ions),
[𝜏𝑛𝑜𝑖𝑠𝑒, 𝜏𝑎𝑖𝑟 ]( educ ion o noise and pollu ion pe uni
leng h when panels ins alled), [𝑆𝑅
𝑎𝑐𝑐, 𝑆𝐸
𝑎𝑐𝑐](nominal speed
o a amp/escala o ), 𝑙𝑡
𝑚𝑎𝑥 ∀𝑡∈ {𝑅, 𝐸, 𝑆𝑃, 𝑃𝑃}(maximum
leng h o amp/escala o /sound panel/pollu ion panel)
Ou pu : Se o Pa e o-op imal in e en ions balancing
accessibili y, ai pollu ion, noise pollu ion and cos
1Ex ac g aph o he a ge a ea: 𝐴↦→ G // F om ex e nal
da abases
2De ine poin s o in e es : o igin (V𝑜) and des ina ion nodes (V𝑔)
3Re ie e nominal noise (𝑓𝑛𝑜𝑚
𝑛𝑜𝑖𝑠𝑒 (·)) and ai pollu ion le els ( 𝑓𝑛𝑜𝑚
𝑎𝑖𝑟 (·))
o all edges Ein G
4Ini ialize a andom a popula ion o in e en ions: {𝛀𝑝}𝑃
𝑝=1
5 o each 𝛀𝑝in he popula ion do
6Rese Gand deploy asse s [𝑣𝑝
𝑘, 𝑣′𝑝
𝑘, 𝑡(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘),Ψ(𝑣𝑝
𝑘, 𝑣′𝑝
𝑘)]𝐾𝑝
𝑘=1
7Upda e weigh s 𝑓𝑎𝑐𝑐 (·),𝑓𝑎𝑐𝑐 (·) and 𝑓𝑎𝑐𝑐 (·) o edges in G
8 o each (𝑣𝑜, 𝑣𝑑) ∈ V𝑜× V𝑑do
9Compu e min-cos ou e E𝑜,𝑑 om 𝑣𝑜 o 𝑣𝑑 h ough G
based on 𝑓𝑇(·) gi en in Exp ession (3.1)// Dijks a
10 E alua e 𝑓𝑎𝑐𝑐 (·),𝑓𝑛𝑜𝑖𝑠𝑒 (·) and 𝑓𝑎𝑖𝑟 (·) o ou e E𝑜,𝑑
11 A e age he objec i e alues o e all ou es (Exp .
(3.18)-(3.20)), yielding 𝑓𝑎𝑐𝑐 (𝛀𝑝|V𝑜,V𝑑),𝑓𝑎𝑖𝑟 (𝛀𝑝|V𝑜,V𝑑)and
𝑓𝑛𝑜𝑖𝑠𝑒 (𝛀𝑝|V𝑜,V𝑑)
12 Compu e cos 𝑓𝑐𝑜𝑠𝑡 (𝛀)o 𝛀𝑝as pe Exp ession (3.21)
13 Ini ialize gene a ion coun e : 𝑔𝑒𝑛 =0
14 while 𝑔𝑒𝑛 < 𝑚𝑎𝑥𝐺𝑒𝑛 do
15 Selec pa en indi iduals om he popula ion // Bina y
ou namen
16 C osso e and mu a ion o p oduce o sp ing // SBX, andom
mu a ion
17 E alua e objec i e alues o o sp ing solu ions (lines 6 o 12)
18 Re ain i es indi iduals in popula ion // As pe he
selec ed MOEA
19 Upda e numbe o gene a ions: 𝑔𝑒𝑛 =𝑔𝑒𝑛 +1
20 Re u n he se o non-domina ed indi iduals in he popula ion
subse o in e en ions in he popula ion ha a e no domina ed in he
space o objec i es, decla ing i o be i s bes app oxima ion o he Pa e o
3.4. Expe imen al Se up 51
on be ween he de ined objec i es (line 20).
3.4 Expe imen al Se up
In o de o e i y he p ac ical applicabili y o he de ised amewo k in
eal-wo ld scena ios, he pe o mance will be e alua ed o e wo use cases
se led on he ci y o Ba celona (Spain). The choice o his ci y o he
de ini ion o he use cases inds i s a ionale in he public a ailabili y o
da a-se s wi h eal ai and noise pollu ion measu emen s, which is a e-
qui emen o be inpu o he amewo k (Figu e 3.1). The wo use cases
a e a eas wi h a high pe cen age o olde people in hei popula ion, a hilly
opog aphy and di e en ypes o walkable s ee s. In pa icula :
•The i s use case (he ea e deno ed as 𝐴1) is loca ed in he neighbo -
hood o Can Ba ó and he no h o Baix Gui na dó, comp ising an a ea
bounded by he la i ude and longi ude coo dina es [41.4191◦,41.4105◦]
and [2.1734◦,2.1510◦], espec i ely. This a ea is known o i s s eep
slopes, especially in he hill o Can Ba ó, which pose a g ea obs acle
o olde people and ci izens wi h educed mobili y. As o he no h-
e n a ea o Baix Gui na dó, i is ull o noisy a enues unde going hea y
a ic all day long.
•The second use case (co espondingly, 𝐴2) co e s he a ea bounded
by [41.4439◦,41.4342◦](la i ude) and [2.1748◦,2.1621◦](longi udes),
known as La Guineue a and Can Pegue a. I is a less hilly a ea wi h
lowe a ic in ensi y han he p e ious use case. Howe e , his neigh-
bo hood su e s om popula ion aging, being among he ones wi h high-
es pe cen age o people abo e 65 yea s wi h espec o he popula ion o
he en i e neighbo hood (as o 2018, be ween 26.79% and 30.50% [126]).
Once hese a eas we e selec ed o expe imen a ion, ele a ion da a was
e ie ed om he Na ional Geog aphic Ins i u e (IGN) o he Spanish
Go e nmen [127]. This eposi o y p o ides public access o Digi al Te -
ain Model (DTM) iles ea u ing a esolu ion equal o 2 me e s. Tiles
co esponding o he use cases we e e ie ed, yielding an accu a e 3D
model o he a ea co e ed by he use cases. The g aph Go each use
case was c ea ed by using he OSMnx lib a y [128], which au oma es
he p ocess o que ying and downloading da a om he well-known Open
S ee Maps mapping se ice [129]. Ele a ion da a 𝑧(𝑣) o e e y node
𝑣∈ V was compu ed based on he in o ma ion in he e ie ed DTM
iles, assigning i as he ele a ion alue o he DTM poin geog aph-
ically closes o he coo dina es (𝑥(𝑣), 𝑦(𝑣)) o he node. Figu es 3.2
(le ) and 3.2 ( igh ) depic he inally composed g aphs o he use cases
𝐴1and 𝐴2( espec i ely), o e laid wi h a simpli ied map o he ci y o
Ba celona. Ai pollu ion and acous ic noise da a we e collec ed om he
public da a eposi o y made a ailable by he ci y council o Ba celona a
h ps://openda a-ajun amen .ba celona.ca . The ai pollu an
conside ed o he wo use cases was NO2, as i was he subs ance wi h
highes concen a ion alues measu ed among he h ee ypes o pollu an s
52 Chap e 3. Imp o ing U ban Accessibili y and En i onmen al
Condi ions using G aph Modeling and Mul i-objec i e Op imiza ion
da a (PM2.5, PM10 and NO2) a ailable in he eposi o y. Ne e heless, ai
quali y indices combining di e en pollu an s can be also p oduced and
used wi hin he p oposed amewo k.
Use Case A2
Use Case A1
Ba celona
Spain
Eu ope
2.155 2.160 2.165 2.170
Longi ude
41.411
41.412
41.413
41.414
41.415
41.416
41.417
41.418
41.419
La i ude
2.162 2.164 2.166 2.168 2.170 2.172 2.174
Longi ude
41.436
41.438
41.440
41.442
41.444
La i ude
75
100
125
150
175
200
225
250
Heigh
Figu e 3.2: (Top) O e all geog aphical loca ion o he conside ed use cases;
(bo om) geo- e e enced g aphs o (le ) use case 𝐴1; ( igh ) use case 𝐴2. Fo
bo h cases, nodes ha e been colo ed as pe he heigh o hei loca ions in
he eal wo ld. In he case o 𝐴1a wide ange o colo s can be obse ed due
o he s eep hill o Can Ba ó.
Following he p ocedu e desc ibed in Algo i hm 1, o igin and des i-
na ion nodes a e selec ed by inspec ing he u ban a eas unde s udy and
de ec ing poin s o in e es o e he a ea. Fo example, when a ge ing ac-
cessibili y o olde people, poin s o in e es may include he add esses o
nu sing homes o senio cen e s, en ances o a pa k, d ug s o es o an am-
bula o y. O igin and des ina ion nodes a e assigned o he e ices 𝑣∈ V
whose coo dina es (𝑥(𝑣), 𝑦(𝑣)) a e close o he selec ed poin s o in e es .
Fo bo h cases o s udy |V𝑜|=5o igin nodes and |V𝑑|=6des ina ion
3.4. Expe imen al Se up 53
nodes we e inally es ablished, gi ing ise o 30 ou es o e which o com-
pu e he objec i e alues o he in e en ions op imized by he amewo k.
The numbe o nodes, and hence o ou es, was es ablished acco dingly
o u ban planning expe s, conside ing popula ion s a is ics and how he
olde popula ion is dis ibu ed spa ially in he neighbo hoods and nea he
poin s o in e es . Such poin s o in e es include:
•Use case 𝐴1:Residencia Vil La Salu (a home o he elde ly), Pa que
Guina do (a pa k), Residència Ac i a Pa c de les Aigües (ano he nu s-
ing home), Hospi al San Pau,Fa màcia Busque s Balsells /Hospi al
Hes ia G acia and Hospi al de la Espe anza / Fa macia ( he las wo
poin s o in e es include a d ugs o e in addi ion o a hospi al).
•Use case 𝐴2:Fa macia Xa ie Solani /El alle de cos u a (a d ug-
s o e), Clinica Pha (a den is ) / Fa macia Izquie do Rido sa (a second
d ugs o e), O icina de A encion Ciudadana del Dis i o Nou Ba is (a
ci izen se ice o ice), Pa que Canyelles (a pa k), Me cado de Canyelles
(a g oce y ma ke ) and Pa c Cen al (ano he pa k).
Table 3.1 summa izes he alues o he pa ame e s ha a e equi ed
o he applica ion o he amewo k. Wi hou loss o gene ali y, he same
alues ha e been used o he wo di e en cases o s udy, which ollows
om he ac ha bo h a e o mula ed o e he same ci y. I should be
no ed, howe e , ha such alues could a y when applying he amewo k
in o he ci ies/coun ies, wi h supply companies imposing o he cos mod-
els and o e ing di e en echnological asse s han he ones conside ed in
his expe imen a ion. Ne e heless, he design lexibili y o he p oposed
amewo k can accommoda e such a ia ions.
Pa ame e Value Desc ip ion
𝛽𝑎𝑐𝑐 1.1m/s a slope −2◦Maximum speed o pedes ian
𝜃𝑚𝑎𝑥 8◦Maximum walkable slope o an olde pedes ian
𝜃𝑅
𝑚𝑎𝑥 12◦Maximum slope o he ins alla ion o a amp
𝜃𝐸
𝑚𝑎𝑥 20◦Maximum slope o he ins alla ion o an escala o
𝑆𝑅
𝑎𝑐𝑐 0.4m/s Nominal speed o a amp
𝑆𝐸
𝑎𝑐𝑐 0.5m/s Nominal speed o a escala o
𝑆𝐿
𝑎𝑐𝑐 1m/s Nominal e ical speed o a li
𝜏𝑛𝑜𝑖𝑠𝑒, 𝜏𝑎𝑖𝑟 ,0.7A enua ion ac o o noise and ai pollu ion
𝑙𝑡
𝑚𝑎𝑥
100 m (𝑡∈ {𝑅, 𝐸}), 500 m
(𝑡∈ {𝑆𝑃, 𝑃𝑃})Maximum ins allable leng h o asse
𝐶𝑓 𝑖𝑥 (𝑡)25ke(𝑡=𝑅), 40ke(𝑡=𝐸),
100ke(𝑡=𝐿), 100e(𝑡∈ {𝑆𝑃, 𝑃𝑃})Fixed cos model o asse s
𝐶𝑣𝑎𝑟 (𝑡)6.8ke/m (𝑡=𝑅), 7.5ke/m (𝑡=𝐸),
10ke/m (𝑡=𝐿), 40e/m (𝑡∈ {𝑆𝑃, 𝑃𝑃})Va iable cos model o asse s
Table 3.1: Values o he pa ame e s con igu ed o he expe imen s.
Rega ding he sea ch p ocess, he ou di e en MOEA desc ibed in
Sec ion 3.3, namely, NSGA-II, NSGA-III, SPEA2 and MOCELL ha e been
chosen. Implemen a ions a ailable in he jMe alPy lib a y [130] ha e been
used. The popula ion size is se equal o 𝑃=100, whe eas he i e a i e
sea ch p ocess is s opped a e 𝑚𝑎𝑥𝐺𝑒𝑛 =2·104gene a ions. The max-
imum numbe o asse s pe in e en ion is se o 𝐾𝑚𝑎𝑥 =60. C osso e
60 Chap e 3. Imp o ing U ban Accessibili y and En i onmen al
Condi ions using G aph Modeling and Mul i-objec i e Op imiza ion
be ween he selec ed o igin and des ina ion nodes. Toge he wi h o he
sa u a ed s ee s (e.g., Ca e de les Camélies o Ca e de Tene i e), his
s ee a i es a he Plaça de la Fon Cas ellana, a oundabou wi h ypi-
cally high in ensi y o a ic all day long. The e o e, he selec ion o hese
wo edges o he deploymen o ai pollu ion panels can be hough o
ma ch sui ably he a ic in ensi y pa e ns o he oads a e sing hese
s ee s.
Figu e 3.7: Quali a i e inspec ion o he in e en ion wi h highes ai pol-
lu ion imp o emen among hose op imized by SPEA2 o e use case 𝐴1. The
in e p e a ion o he plo s included in his igu e can be done as in Figu e
3.6. Hea maps o he noise in ensi y le els o he a ea in which he asse s
a e sugges ed o be deployed a e included in each case.
Now he scope o he discussion is shi ed owa ds he solu ion bes
imp o ing he noise pollu ion o pa hs be ween he selec ed o igin and
des ina ion nodes o use case 𝑆1. This solu ion is shown in Figu e 3.8, in
he same o ma and in e p e a ion p o ocol as in he p e ious igu es.The
solu ion is obse ed o comp ise 3 di e en edges (s ee s) o he a ea unde
s udy o he deploymen o noise a enua ing panels: i) Ca e de F ancesc
Aleg e, which con e ges o he a o emen ioned Ma e de Déu de Monse a
a enue and has a pe ol s a ion (shown in he igu e) as a clea sou ce o
ambien noise; ii) Ca e de Ca agena, which, in addi ion o i s high a ic
in ensi y, su e s om a high le el o ambien noise due o he p esence o
a axi s op and he main en ance o he Fundaci’o Puig e (a uni e si y
hospi al), which ensu es a egula ly massi e a luence o pedes ians; and
iii) Ca e e a del Ca mel wi h Ca e d’Ana M. Ma u e Ausejo, an a ea
whe e a pa king zone o mo o bikes is alloca ed. These loca ions p oposed
by he amewo k a e ac ually in a eas wi h conside able acous ic s ess
(>70 dB), hence alida ing he amewo k also o his objec i e. I is
also wo hwhile no ing ha he ins alla ion o hese asse s does no a ec
hea ily in he cos , yielding one o he cheapes solu ions gi en by he
amewo k o his use case. This is due o he la ge di e ences in e ms o

3.5. Resul s and Discussion 61
ins alla ion cos s assumed o noise/ai pollu ion panels and accessibili y
in as uc u es, such as ele a o s.
Figu e 3.8: Quali a i e inspec ion o he in e en ion wi h highes ambien
noise pollu ion imp o emen among hose op imized by SPEA2 o e use case
𝐴1.
Quali a i e Assessmen o In e en ions: Use Case 𝐴2
A simila line o easoning can be ollowed when inspec ing in dep h he
in e en ions e ol ed by SPEA2 o e he second use case unde s udy in he
expe imen al se up. As such, Figu e 3.9 depic s he in e en ion leading o
he la ges accessibili y imp o emen among hose composing he Pa e o
on app oxima ion ound by his sol e . This in e en ion comp ises 4
asse s concen a ed in wo di e en pa s o he scena io: he exi o a
public pa king space close o he me opoli an ain s a ion o Canyelles,
and he access o he s a ion i sel . I is s aigh o wa d o no e ha as in
he p e ious use case, his depic ed in e en ion e inces, on one hand, ha
hese loca ions in he a ea ac ually equi e a mechanical asse o o e come
e iden accessibili y issues. On one hand, he exi o he pa king space is
no u banized and is s eep o a pedes ian wi h mobili y cons ain s. On
he o he hand, he di e ence in heigh be ween he pedes ian access o
he me opoli an ain s a ion and he s ee le el is iden i ied as an edge
ha equi es an asse o imp o e i s accessibili y. Howe e , he pho o
co esponding o his loca ion e eals such an asse is al eady ins alled in
p ac ice (mechanical s ai s), he eby alida ing he in e en ions op imized
by he p oposed amewo k.
The discussion on he in e en ions op imized o he use case 𝐴2 ol-
lows in Figu e 3.10, which illus a es wo di e en solu ions o he es i-
ma ed Pa e o on . The map and anno a ed plo s on he le mos pa o
he igu e co espond o he in e en ion co esponding o he lowes (bes )
alue o he noise pollu ion objec i e 𝑓𝑛𝑜𝑖𝑠𝑒 (·). By con as , he igh mos
62 Chap e 3. Imp o ing U ban Accessibili y and En i onmen al
Condi ions using G aph Modeling and Mul i-objec i e Op imiza ion
Figu e 3.9: Quali a i e inspec ion o he in e en ion wi h highes acces-
sibili y imp o emen among hose op imized by SPEA2 o e use case 𝐴2.
pa o he image depic s he map o sugges ed asse s o he solu ion ea-
u ing he bes alue o he ai pollu ion objec i e 𝑓𝑎𝑖𝑟 (·). Rega ding he
i s depic ed solu ion, i s geno ype su p isingly con ains no asse s o be
deployed, which makes his solu ion coincide wi h ha implying minimum
cos . The eason o his unexpec ed esul is ha he amewo k disco e s
ha when no inse ing any asse s, he sho es ou es be ween he selec ed
o igin and des ina ion nodes unde such ci cums ances al eady go h ough
u ban a eas wi h minimal ambien noise. To shed ligh on his s a emen ,
a plo is anno a ed in he map showing he op imal ou es be ween he
public pa king space (o ange) and he wo o he poin s o in e es (blue)
co esponding o he case wi h la ges accessibili y imp o emen ( he one
al eady discussed in Figu e 3.9). I can be seen ha i accessibili y as-
se s we e deployed in he exi o he pa king space and he access o he
me opoli an ain s a ion, sho es pa hs be ween hese selec ed nodes
would change, making hem low close o he Via Fa éncia and he Ronda
de Dal , wo o he oads in his a ea wi h he highes daily a ic p o iles.
I such asse s we e no ins alled, he sho es pa hs change o low h ough
esiden ial a eas, becoming longe o be a e sed in e ms o dis ance, bu
much be e in e ms o acous ic pollu ion. This unexpec ed esul sup-
po s he inhe en u ili y o he amewo k o disco e coun e -in ui i e
ye easonable in e en ions.
The discussion con inues in he same igu e, ocusing on he solu ion
analyzed on he igh (la ges imp o emen in e ms o ai pollu ion). In
his second case, he geno ype sugges s he ins alla ion o 4 pollu ion pan-
els in di e en pa s o he u ban a ea: i) he exi o he Plaça de Ka l
Ma x oundabou , which is ele an gi en i s cen ali y in he oad ne -
wo k o he a ea and i s closeness o one o he selec ed poin s o in e es ;
3.5. Resul s and Discussion 63
Figu e 3.10: Quali a i e inspec ion o he in e en ion wi h he highes
noise imp o emen (le ) and ai pollu ion imp o emen ( igh ) among hose
op imized by SPEA2 o e use case 𝐴2. I is impo an o no ice ha he
solu ion leading o he highes noise imp o emen does no in ol e ins alling
any asse , since he amewo k de e mines, h ough i s e olu iona y sea ch,
ha in e en ions o imp o e he o he objec i es imply a change in he
op imal ou es be ween o igin and des ina ion nodes, leading o a highe
exposu e o he pedes ian o ambien noise.
ii) he junc ion be ween Rambla del Caçado and Ca e de l’Isa d, a
esiden ial a ea close o a pa king lo and he Plaça de la República, a
oundabou wi h in ense a ic; and iii) a p i a e access o a pa king lo
whose exi lows in o Ca e de Góngo a, a s ee loca ed in a densely
popula ed neighbo hood o he a ea ha is close o he Ronda de Dal
main oad. I is c ucial o unde s and ha since he a ea co e ed in 𝐴2is
mo e esiden ial han 𝐴1, and ha no accessibili y asse s a e sugges ed in
he solu ion cu en ly unde discussion, sho es pa hs be ween o igin and
des ina ion nodes al eady a e se a eas wi h ela i ely low ai pollu ion.
Consequen ly, pollu ion panels sugges ed o be deployed by his solu ion
a e mos ly a ec ed by hei p oximi y o sou ces o pollu ion (e.g., pa king
lo s, oundabou s).
The quali a i e examina ion o he in e en ions inishes by analyz-
ing he in e en ion wi h he highes economic cos ound by he p oposed
amewo k o e use case 𝐴2, which is depic ed in Figu e 3.11. Fi e di e en
accessibili y asse s a e p oposed: besides he access o he me opoli an
ain s a ion and he access amp o he public pa king space (al eady
discussed in Figu e 3.9), h ee new loca ions a e pinpoin ed by he in e -
en ion: i) a long non-mechanical amp (Ca e de Can Esenya) loca ed
close o one o he poin s o in e es ; and ii) wo consecu i e long s ai s lo-
ca ed in a neighbo hood nea he public pa king lo and he me opoli an
ain s a ion. Besides u he bu essing ha he loca ions sugges ed by
he amewo k make p ac ical sense, he ac ha he solu ion in ol ing
maximum cos is domina ed by accessibili y asse s is a consequence o he
64 Chap e 3. Imp o ing U ban Accessibili y and En i onmen al
Condi ions using G aph Modeling and Mul i-objec i e Op imiza ion
Figu e 3.11: Quali a i e inspec ion o he in e en ion equi ing he highe
economical cos among hose op imized by SPEA2 o e use case 𝐴2. In his
las depic ed case, he highe ins alla ion cos s assumed o accessibili y-
ela ed asse s imply ha he mos expensi e in e en ion consis s o 5 asse s
o be deployed in physical loca ions cha ac e ized by accessibili y issues, as
he anno a ed pho og aphs clea ly expose.
highe ixed cos model ealis ically assumed o his ype o asse s. Fo
non-ex eme in es men cos s, howe e , i is no s aigh o wa d o de-
cide which ype o asse o ins all and whe e, no is i simple o asce ain
he consequences o such ins alla ions in ai pollu ion, acous ic noise and
accessibili y. The e lies he co e pu pose and inhe en u ili y o he p o-
posed amewo k: o au oma e and op imize he in e en ions accoun ing
o hese mul iple c i e ia, as clea ly shown by he quali a i e examina ion
o he in e en ions done in his sec ion.
3.5.3 Conside a ions o he Applica ion o he F ame-
wo k in New Case S udies
Se e al aspec s mus be aken in o accoun when deploying his amewo k
in u u e p ac ical case s udies.
Quali y and ep esen a i eness o a ailable da a. Fi s ly, i is c i -
ical o ho oughly examine he ele ance and quali y o he da a a ailable
o each case. Un o una ely, despi e he hype o Sma Ci y and Open
Da a Pla o ms in he pas decade, no many ci ies p o ide high-quali y
da ase s ha allow o he applica ion o amewo ks as he one p oposed
3.5. Resul s and Discussion 65
in his Chap e 3. In some cases, da a is no cap u ed wi h enough es-
olu ion (spa ial and/o empo al) o allow o ine-g ained analyses. In
addi ion, his o ical da ase s a e no p ope ly a chi ed o made publicly
a ailable. Fo ins ance, ai quali y and noise da a need o e lec ypical
si ua ions o he a eas unde s udy, he e o e excluding non- ypical da a
de i ing om pa icula e en s o ci cums ances (e.g., empo a y oad o
cons uc ion wo ks). I ai quali y and noise models we e a ailable o
u u e case s udies, he algo i hm could be ine- uned o be e accoun
o speci ic a enua ion ac o s, he e o e including he impac o ins alling
each asse in o he model.
O he ex e nal da a sou ces. Mo eo e , as men ioned be o e, o he
ele an da a sou ces should be ed o he amewo k when possible, such
as he loca ion and cha ac e is ics o u ban accessibili y in as uc u e, o
he public o p i a e na u e o ce ain pa hs and access poin s included in
he g aph modeling he u ban a ea unde s udy. O he wise, he op imal
ou es ou pu by he amewo k could e en ually a e se segmen s ha
a e ac ually close o he public. A simila obse a ion can be made in
ega d o he spa ial g anula i y o pollu ion and noise da a: depending
on he sensing equipmen used o he measu emen campaign, noise and
pollu ion maps could e en ually be accu a e in ce ain pa s o he u -
ban a ea, whe eas in o he a eas da a in e pola ion s a egies should be
en o ced. The low-like p opaga ion na u e o pollu ion and he blockage
o noise by buildings could make his in e pola ion compu a ionally e y
cos ly, equi ing complex simula ion s ages.
Compu a ional complexi y. In pa ial connec ion o he abo e is he
compu a ional e o equi ed o he execu ion o he amewo k. Expe -
imen s pe o med o yield he esul s epo ed in his chap e we e un
o e an implemen a ion o he amewo k ha is no op imized o nea -
immedia e decision making. The eason is ha he cha ac e is ics o he
asse s composing he in e en ion does no equi e a eal- ime ope a ion
o he amewo k op imizing hei loca ion. Howe e , i is possible o o -
mula e al e na i e scena ios ela ed o accessibili y and walkabili y ha
could equi e sho e unning imes o he amewo k, such as hose im-
plying decision a iables ha can be uned in an agile ashion (e.g., ehicles
o on-demand anspo se ices) o hose in ol ing inpu da a lows wi h
sho p e alence pe iods (namely, sho - e m wea he o ecas s). Ne e -
heless, he algo i hmic phases o he wo k low desc ibed in Algo i hm
1can be easily implemen ed in pa allel, om he disco e y and e alua-
ion o he ou es o he applica ion o he e olu iona y sea ch ope a o s
o he me aheu is ic sol e a hand. Fu he mo e, ce ain design choices
such as he numbe o o igin and des ina ion nodes |V𝑜|and |V𝑑|can
u he be uned o alle ia e he compu a ional bu den o he amewo k
and o make i complian wi h he la ency cons ain s imposed o decision
making. All in all, he compu a ional e iciency o he amewo k should
ma ch he la ency equi emen s imposed by he op imiza ion objec i es,

66 Chap e 3. Imp o ing U ban Accessibili y and En i onmen al
Condi ions using G aph Modeling and Mul i-objec i e Op imiza ion
he speed o implemen a ion o he decision a iables in ol ed in he p ob-
lem s a emen , and he p e alence o he ex e nal in o ma ion lowing in o
he sys em.
Op imiza ion o o he walkabili y/accessibili y scena ios. De-
spi e no men ioned h oughou he chap e , he amewo k he e desc ibed
is also capable o op imizing ou es o wheelchai use s and people wi h
o he mobili y issues. As i is a e sa ile and easily con igu able ool, in
he case o hese use s i would be enough o modi y he pa ame e s o
he op imize which a ec hem signi ican ly, such as he maximum slope
alue, acco ding o he c i e ia o he expe s, o he speed. The u ban
elemen s o be inse ed could also be adap ed o he use s, o example,
dis ega ding he op ion o escala o s o gi ing g ea e impo ance o he
ins alla ion o ele a o s. When conside ing such al e na i e scena ios, in-
co po a ing new ex e nal sou ces o da a would become a he a necessi y
han a possibili y, such as wea he condi ions o seasonal sola shadowing
a s ee le el.
On he gene aliza ion o he benchma k esul s o o he use
cases. Finally, i is impo an o bea in mind ha in o he p ac ical use
cases o he ela ionships migh exis be ween he conside ed objec i es.
Consequen ly, one should no expec ha he bes pe o ming MOEA
ound in his con ibu ion eplica es i s ou s anding pe o mance in o he
se ups. The e o e, a new algo i hmic benchma k like he one shown in
he expe imen s o his chap e should be ollowed. Fu he mo e, an in-
e ac i e in e ace sui able o he audience o he amewo k would be
con enien o acili a e he assessmen o he solu ions by use s ha a e
no necessa ily expe s in mul i-objec i e op imiza ion. This being said,
enabling a quali a i e inspec ion o he solu ions composing he app ox-
ima ed Pa e o on es ima ed by he algo i hm would ease his p ocess
and edound o he us wo hiness o he use s in he pla o m.
3.6 Summa y
Ci ies ace un esol ed challenges when pu suing o become iendly spaces
o hei ci izens, especially o olde people. The c ea ion o age- iendly
u ban a eas equi es c ucial ac ions o imp o e hei accessibili y and o
enhance he en i onmen al quali y (noise, pollu ion) pe cei ed by pedes-
ians when a e sing he u ban en i onmen . To echnologically suppo
decisions o be made o his pu pose, his Chap e 3 has p esen ed a
no el amewo k ha emb aces mul i-objec i e e olu iona y op imiza ion
and g aph modelling o op imally de e mine whe e o ins all u ban asse s
ha help ealize he a o emen ioned objec i es, also accoun ing o he
cos e iciency o hei ins alla ion. Such con lic ing objec i es delinea e a
Pa e o ade-o comp ising di e en asse s/in e en ions ha di e en ly
balance he goals sough in he o mula ed p oblem: minimum ai pollu-
ion, minimum ambien noise, bes accessibili y, and minimum cos . E o-
lu iona y me aheu is ics o mul i-objec i e op imiza ion a e hen adop ed
3.6. Summa y 67
o e icien ly explo e he mixed in ege - eal sea ch space o he p oblem,
de e mining which asse s o deploy, whe e hey mus be ins alled in he
a ea unde s udy, and hei speci ic unc ional ea u es. The ou pu Pa e o
on app oxima ion elici ed by his new amewo k can in o m expe s in
u ban planning du ing hei decision-making p ocesses ela ed o acces-
sibili y and en i onmen al quali y, educing he ime, and imp o ing he
quali y o decisions made in his ega d. Mo eo e , his ou pu can be
used o p o ide objec i e g ounds o planning and in es men decisions
made in he con ex o he long- e m Gene al U ban Plans, o ins ance
including demog aphic p ojec ions and planned in as uc u es as inpu s
o he model.
I has been expe imen ally showcased he applicabili y o p oposed
amewo k in wo eal-wo ld use cases loca ed in he ci y o Ba celona
(Spain), conside ing eal ai pollu ion, en i onmen al noise, and opo-
g aphical da a collec ed o e his ci y. Expe imen s ha e con i med ha :
•Clea pe o mance gaps exis be ween di e en me aheu is ic op imiza-
ion algo i hms when used wi hin ou amewo k, as can be deduced
om he epo ed s a is ics on ou quali y indica o s, and a pos e io
signi icance analysis o di e ences obse ed among hem.
•In e en ions e ol ed by he amewo k con o m o in ui ion and com-
mon sense o each o he use cases, p oposing he ins alla ion o asse s
o imp o e all he conside ed objec i es in loca ions wi h eal issues in
e ms o accessibili y, noise and ai pollu ion.
The design lexibili y o he amewo k and he p omising esul s he ein
discussed s imula e se e al esea ch di ec ions o be pu sued in he nea
u u e. To begin wi h, a majo e o will be in es ed owa ds o e coming
he p ac ical limi a ions iden i ied in Sec ion 3.5.3: speci ically, means o
a be e ex apola ion o ai pollu ion and en i onmen al noise da a a he
s ee le el will be s udied by gauging he impac o he u ban opology in
he p opaga ion o noise and ai low. Likewise, a wide po olio o da a
sou ces should be conside ed o en ich he g aph model cons uc ed o
he a ea a hand, including public/p i a e access, al eady ins alled asse s
o o he pa ame e s om whe e addi ional op imiza ion objec i es can be
o mula ed. Fo ins ance, by including an es ima ion o he sunligh expo-
su e a s ee le el o he a ea om a 3D dynamic model, he amewo k
could also choose pa hs aking shade in o accoun , o seek he loca ion
whe e o ins all clima ic shel e s o he sake o he he mal com o o he
pedes ian along his/he pa h. Fu he mo e, he inclusion o o he ex e -
nal da a sou ces will enable he ex apola ion o he amewo k o op imize
o he o ms o walkabili y/accessibili y, comp ising objec i es and decision
a iables ha may di e wi h espec o he asse s he ein conside ed (e.g.,
wheel chai use s o on-demand anspo se ices o olde people). Fi-
nally, use cases in o he egions and coun ies should be selec ed o alida e
he amewo k, conside ing local pa icula i ies (e.g., p e alen egula o y
cons ain s, cul u al di e ences, and wea he condi ions) ha may equi e
a e o mula ion o he p oblem and he adap a ion o he amewo k o
hei ci izens o ully bene i om i s applica ion.
69
Chap e 4
E icien Es ima ion o
G ound-Le el Ai
Tempe a u e in U ban
A eas using Machine
Lea ning
Since he ea ly 19 h cen u y, human ac i i ies ha e con ibu ed o a ma ked
in ensi ica ion o he g eenhouse e ec , a na u al phenomenon in which
a mosphe ic gases ap hea and main ain he Ea h’s su ace empe a u e
[138]. This human-induced ampli ica ion, e e ed o as he an h opogenic
g eenhouse e ec [139], has led o a p onounced accele a ion in global
wa ming, wi h signi ican implica ions o he Ea h’s clima e sys em o
da e. This clima e change is clea ly e lec ed no only by he ise o he
global a e age empe a u e, which has aised 1.2C◦in he las 200 yea s
[11], bu also in he inc ease in he amoun o annual hea wa es days
[11]. These wo sening wea he condi ions se e ely a ec human ac i i ies,
ecosys ems and heal h. Indi iduals ha a e exposed o high empe a u es
and hea wa es can ha e a ious de imen al e ec s on hei heal h. These
e ec s may include inc eased isks o hea s ess and hea s oke leading
o mo bidi y and mo ali y, as well as he deg ada ion o espi a o y and
ca dio ascula condi ions [140], [141]. The hea - ela ed mo ali y a e has
expe ienced a signi ican inc emen o 53.7% in he pas 20 yea s o hose
o e 65 yea s [11].
No only human heal h is a ec ed by he high empe a u es. The
economic impac o he hea wa es is e lec ed, o example, in he 302
billion wo king hou s los in 2019, which ep esen s an inc ease o 51.8%
compa ed o yea 2000 [11]. A majo p oblem di ec ly a ec ing sus ain-
abili y lies in he ising ene gy consump ion o housing and indus y. As
hea consump ion is closely ela ed o empe a u e a ia ions in luenced
by clima e change, his poses a p essing conce n o sus ainable p ac ices.
Fo example, 40% o all ene gy consumed in Eu ope is consumed by he
housing s ock, as well as accoun ing o 36% o all CO2emissions [4].
76 Chap e 4. E icien Es ima ion o G ound-Le el Ai Tempe a u e in
U ban A eas using Machine Lea ning
use o his ype o echnology gi es a new ool o es ima e su ace ai em-
pe a u e in a much as e way han classical nume ical app oaches and
mo e e o lessly, wi h po en ials o be mo e scalable o o he ci ies han
he amewo ks e iewed abo e.
4.2 Ma e ials and Me hods
The me hod p oposed in his chap e aims o es ima e he Tawi h a spa-
ial esolu ion o 100 me e s in an hou ly basis. Fo his pu pose a U-Ne
a chi ec u e is p oposed, which co ela es he di e en spa ial and empo-
al da ase s wi h he Tacompu ed by U bClim o e he g id on which he
p oblem is de ined. In he nex subsec ions, he s udy a ea (Subsec ion
4.2.1) and he di e en da ase s used (Subsec ion 4.2.2) will be desc ibed,
and he modelled a chi ec u e (Subsec ion 4.2.3) will be p esen ed.
4.2.1 S udy A ea
The s udy a ea, shown in Figu e 4.1, co e s he hole me opoli an a ea o
he ci y o Bilbao (Spain) and pa o he p o ince o Biscay wi h an o e all
a ea co e ed o 25 ×25 km2. The a ea alls wi hin he C b (i.e. empe a e,
no d y season, wa m summe ) Köppen-Geige class [200]. The ci y is
cha ac e ized by mode a e empe a u es in bo h summe and win e , due
o he in luence o he A lan ic Ocean. This implies a mean empe a u e
o 14-15 ◦C, eaching maximum empe a u es o 25-26 ◦C in July and
Augus [201]. The Ne ión i e lows h ough he middle o he ci y and
he alley, whe e he ai masses a e channeled p o iding he ci y wi h an
impo an en ila ion pa h. The b eezes de eloped by bo h he opog aphy
and p oximi y o he sea a e ele an as well.
4.2.2 Da ase s
In his subsec ion, we p o ide an o e iew o he da ase s used o ain he
p oposed model, which in eg a es bo h me eo ological and spa ial in o ma-
ion o p edic g ound-le el ai empe a u e. This subsec ion in oduces
he a ious da ase s employed, ocusing on hei s uc u e, sou ces, and
oles in he model aining p ocess. Fou di e en da ase s a e p esen ed o
comp ehensi ely desc ibe he da a pipeline: he Ta ge T aining Da ase ,
which de ails he U bClim-simula ed empe a u e da a used as he p ima y
a ge o he model; he Me eo ological Da ase , de i ed om ERA5 e-
analysis da a and used o p o ide me eo ological inpu s o aining; he
Spa ial Da ase , which comp ises a ious geog aphical and mo phological
ea u es ha desc ibe he u ban landscape o he s udy a ea; and inally,
he Valida ion Da ase , which con ains eal ai empe a u e measu emen s
collec ed om local wea he s a ions and is used o e alua e he accu acy
o he model’s es ima ions. These da ase s, aken oge he , o m he basis
o bo h aining and assessing he model’s capaci y o accu a ely es ima e
Taunde a ying me eo ological and spa ial condi ions.

4.2. Ma e ials and Me hods 77
Ta ge T aining Da ase
The a ge aining da ase o he model was based on he da ase called
“Clima e a iables o ci ies in Eu ope om 2008 o 2017” p o ided by he
Cope nicus Clima e Change Se ice [202]. This da ase o e s ai empe -
a u e a a ine-g ained esolu ion o 100 me e s in an hou ly basis om
2008 o 2017 o a se o 100 Eu opean ci ies. The empe a u e alues
we e ob ained h ough simula ions wi h U bClim model [143].
Me eo ological Da ase
Expe imen s la e p esen ed and discussed in his chap e u ilize me e-
o ological da a coming om ERA5 Reanalysis [203], which is he same
inpu da ase used by U bClim. The eanalysis da a was acili a ed by
he Eu opean Cen e o Medium-Range Wea he Fo ecas s h ough he
Cope nicus Clima e Da a S o e in Ne CDF ile o ma . ERA5 o e s global,
hou ly da a spanning om 1940 o he p esen a a spa ial esolu ion o
0.25° × 0.25°. Da a be ween 1981 and 2017 was collec ed o he ollowing
su ace a iables: 2m ai empe a u e, p ecipi a ion, speci ic humidi y and
wind componen s (uand a 10m). Conside ing he coa se esolu ion o
ERA5, g id cell alues o each a iable ha all inside he s udy a ea
whe e a e aged o ob ain he co esponding 1981-2017 hou ly ime se ies.
To educe he clima ic a iabili y o he aining da ase and ocus he
a ge o he model o hose days in which he hea s ess can be mo e
p esen , he Local Wea he Type (LWT) classi ica ion p oposed by [204]
is applied. The idea is ha simila synop ic condi ions lead o a simila
he mal beha iou o he ci y and, hence, he model can ocus on es i-
ma ing how he Tabeha es unde hose pa icula condi ions (i.e. unde
ha pa icula LWT). To ob ain he LWT se e al daily me ics a e ob-
ained om he abo e men ioned hou ly a iables o he baseline pe iod
1981-2010: he mal ampli ude (Ta,max - Ta,min), o al p ecipi a ion, a e -
age speci ic humidi y, and wind speed and di ec ion. Wind di ec ion was
u he p ocessed o con e i in o a ca ego ical a iable by classi ying he
alues based on he ollowing c i e ia: N ( om -45º o 45º), E ( om 45º
o 135º), S ( om 135º o 225º), and W ( om 225º o 315º). This classi i-
ca ion sligh ly di e s om he one applied by [204], because i ollows he
ecommenda ions o he Basque Me eo ological Agency o he iden i ica-
ion o wind componen s [205]. 12 di e en LWT ha a e ep esen a i e
o he di e en seasons a e ob ained. Among hem, he selec ed one was
he mos ep esen a i e o summe (June, July, Augus and Sep embe )
and he one ha had he po en ial o gene a e ad e se he mal si ua ions.
This clus e was iden i ied as a sunny day wi h a weak (1.72 m/s) eas e n
wind componen , high speci ic humidi y (a ound 11 g/kg). A e sea ch-
ing he summe days ha belong o ha speci ic LWT wi hin he pe iod
2008-2017, a o al o 164 days we e ob ained.
78 Chap e 4. E icien Es ima ion o G ound-Le el Ai Tempe a u e in
U ban A eas using Machine Lea ning
Spa ial Da ase
The spa ial da ase s consis o di e en a iables ha de ine he mo -
phological cha ac e is ics o he s udy a ea. These a iables ma ch he
ones used by U bClim o simula e he aining da ase [206]: impe ious-
ness and ele a ion o he e ain and land co e . The impe iousness and
he land co e we e ob ained om he Cope nicus Land Moni o ing Se -
ice po al [207], which p o ides ee Eu opean and global geog aphical
in o ma ion. The DTM was e ie ed om he IGN o he Spanish Go -
e nmen [127] o con enience. This public eposi o y dis ibu es a DTM
p oduc a 200m, which is simila o he esolu ion o he Global Mul i-
esolu ion Te ain Ele a ion Da a (GMTED) ha is o iginally used by
U bClim. Al hough he yea o he e sions do no ma ch (2012 IGN s.
2010 GMTED), he impac o he di e ences can be conside ed negligible.
Among he a iables used o iginally by U bClim, wo we e no con-
side ed: he No malized Di e ence Vege a ion Index (NDVI) and he an-
h opogenic hea lux. Focusing only on summe days, and conside ing he
NDVI’s seasonal beha io , all he aining days a e expec ed o ha e simi-
la alues. In he case o he an h opogenic hea lux, i s spa ial esolu ion
(5km) was conside ed oo coa se o ge he ine spa ial pa e ns de i ed
om he dis ibu ion o , o example, a ic and buildings.
Valida ion Agains Real TaMeasu emen s
Fo alida ion agains eal Ta alues, poin measu emen s we e also ob-
ained om di e en wea he s a ions ha Euskalme , he Basque me eo-
ological agency, ope a es o e he a ea unde s udy. Tameasu ed alues
om a o al o 7 wea he s a ions we e collec ed o alida ion in he ime
ange om 2008 o 2017. La A boleda, loca ed in a moun ain; Pun a
Galea, placed in a ligh house; Galindo, Deus o and Zo o za in u ban
a eas nea a i e ; and A igo iaga and De io in u banized u al a eas.
Figu e 4.1 shows he loca ions o he wea he s a ions, ma ked wi h a ed
do .
4.2.3 P oposed Model
T ain and Tes Samples
Be o e explaining in de ail he a chi ec u e used o co ela e he Tawi h
he spa ial and me eo ological a iables, he subse s o aining and es
da a a e desc ibed. The i s s ep applied o all he da ase s desc ibed in
Subsec ion 4.2.2 was he no maliza ion o he a iables in he ange [-1,1]
in o de o imp o e he in e p e abili y o he da a, using:
𝑥𝑛𝑜𝑟𝑚 =2𝑥−min 𝑥
max 𝑥−min 𝑥−1.(4.1)
Once da a was no malized, a subse o aining and es da a was c e-
a ed by di iding he use case a ea in o se e al pa ches, as can be seen in
4.2. Ma e ials and Me hods 79
3.1°W
3.1°W
3.05°W
3.05°W
3°W
3°W
2.95°W
2.95°W
2.9°W
2.9°W
2.85°W
2.85°W
2.8°W
2.8°W
43.2°N 43.2°N
43.25°N 43.25°N
43.3°N 43.3°N
43.35°N 43.35°N
43.4°N 43.4°N
Figu e 4.1: The s udy a ea di ided in pa ches. The pa ches shadowed in
yellow co espond o he alida ion ones, while he pu ple ones deno e es
pa ches. The me eo ological s a ions a e loca ed in he ed do s.
Figu e 4.1. A o al o 64 pa ches o 3.1 ×3.1 km we e c ea ed. To ensu e
he ep esen a i eness o all map a eas du ing model aining, 53 pa ches
we e dis ibu ed o aining, highligh ed in black in Figu e 4.1. F om hose
53 pa ches, 5 o hem we e ese ed o model alida ion, highligh ed in
yellow in Figu e 4.1. The loca ions o he es pa ches we e chosen in
acco dance wi h he loca ions o he me eo ological s a ions. The se en
di e en s a ions lie in o six di e en es pa ches, wi h he s a ions o
Deus o and Zo o za alling in o he same pa ch. The es pa ches a e
highligh ed in pu ple in Figu e 4.1. Hence, om he 24 hou s o he 164
days o LWT, 3936 hou s a e ob ained o ain he U-Ne model. Conse-
quen ly, each o he ain and es iles will ha e 3936 examples.
U-Ne Model
The model used o es ima e Ta, shown in Figu e 4.2, is a a ia ion o he
well-known U-Ne a chi ec u e, which consis s o an encode -decode wi h
skip connec ions. The inpu da ase s o he encode consis o he h ee
80 Chap e 4. E icien Es ima ion o G ound-Le el Ai Tempe a u e in
U ban A eas using Machine Lea ning
spa ial maps desc ibed abo e. The no malized inpu enso ep esen ing
hose a iables has he shape o 32 ×32 ×3 o each pa ch o he map. As
explained in Chap e 2 (Subsec ion 2.1.2) he U-Ne is di ided in o wo
pa s, he encode and he decode . The encode , composed o CNNs, as
well as pooling laye s and d opou laye s, educes he dimensionali y o he
inpu enso in aim o ind di e en hidden ela ionships and ex ac high
le el ea u es om i . Once he desi ed dimension is achie ed, he decode
econs uc s he image, applying an in e se p ocess. In his upsampling
p ocess, skip connec ions a e added be ween he encode and he decode
o help in a high-quali y econs uc ion o he image. Fo he case desc ibed
in his chap e , h ee downsample blocks a e used each consis ing o double
con olu ional laye (wi h ReLu as he ac i a ion unc ion), a pooling laye
and a d opou laye . A e down-sampling he inpu da a he lowes -
dimensional space is eached, also known as la en space. A his s age,
due o i s low spa ial esolu ion, he me eo ological da a is added. Bo h Ta
and he es o he me eo ological a iables a ime s ep a e condi ioned
by he me eo ological condi ions o he p e ious hou s. Tha is why, o
es ima e he Tao an ins an , he me eo ological da a o he p e ious 2
hou s is aken in o accoun . Hence, he me eo ological ec o s go om -2
o . The me eo ological da a o each ime s ep is in inse ed in a enso
o shape 5 ×1 o he econs uc ed image, esul ing in he inal ou pu
map o he es ima ed Ta. Once he me eo ological da a is coupled o he
spa ial da a, he upsampling s a s. The decoding s uc u e mi o s ha
o he encode , comp ising 3 up-sample blocks. Each block includes a pai
o con olu ional anspose laye s, a conca ena ion laye , and a d opou
laye and leads o a econs uc ed image o shape 32 ×32 ×1.
The U-Ne con ains di e en pa ame e s ha ha e o be uned o ob-
ain he op imal pe o mance. Fo he case exposed he e, a lea ning a e
o 10−5, a ba ch size o 32, and 30 epochs a e chosen. As he loss unc ion,
Mean Squa e E o (MSE) is selec ed:
𝑀𝑆𝐸 =1
𝑛
𝑛
∑︁
𝑖=1
(𝑌𝑖−ˆ
𝑌𝑖)2,(4.2)
whe e 𝑌𝑖is he obse ed alue and ˆ
𝑌𝑖 he p edic ed alue.
4.3 Resul s and Discussion
In his sec ion he esul s ob ained o he U-Ne model o e he scena io
de ined o e he ci y o Bilbao a e p esen ed and discussed. The discussed
esul s aim o in o m he esponses o he h ee RQs o mula ed a he
beginning o his chap e . Each o hese is add essed indi idually in he
ollowing subsec ions.
4.3. Resul s and Discussion 81
64
128
256
512 512
256
128
64
32 x 32 x 3
Inpu da a
32 x 32
16 x 16
4 x 4
4 x 4
8 x 8
32 x 32
16 x 16
Ou pu
32 X 32 X 1
2 x 2 2 x 2
1024 1024
+5 x 1 =
Tempo al me eo ological da a
-2
-1
8 x 8
Figu e 4.2: The U-Ne a chi ec u e p oposed o he es ima ion o Ta.
The sum in he la en space co esponds o a conca ena ion o wo la ened
ec o s.
4.3.1 RQ1: Can an AI-based Da a Model Achie e Ac-
cu a e TaEs ima ions ha a e Close o Those
Elici ed by a Nume ical Model?
In his i s RQ he aim is o es whe he he es ima ed alues o Taa e
spa ially consis en and wi h alues ha a e close o he ones elici ed by
he U bClim nume ical model. We ecall he eade ha , as desc ibed in
Sec ion 4.2.2, he U-Ne model is being ained o es ima e he Ta alues
p o ided by he Cope nicus U bClim simula ions da ase . Hence, a good
model will be ha educing he e o be ween he Ta alues ob ained om
i and hose om U bClim. To assess whe he his condi ion is me by
ou p oposed U-Ne a chi ec u e, ou eg ession me ics be ween he wo
models a e compu ed o each wea he s a ion a ea du ing he pe iod o
s udy. The me ics chosen a e he Pea son Co ela ion Coe icien (PCC),
he Roo Mean Squa e E o (RMSE), he Mean Absolu e E o (MAE)
and he Mean Absolu e Pe cen age E o (MAPE). The esul s can be
ound in Table 4.1.
As i can be seen in he abo e able, o all he se en loca ions, he
Pea son co ela ion is abo e 0.95, and o almos all o hem is nea o
abo e 0.98 alue. Conside ing ha he de ini ion o he Pea son co e-
la ion coe icien is a alue be ween -1 and 1 ha quan i ies he linea
co ela ion be ween wo da ase s, being 0 no co ela ion and -1 and 1 ull

82 Chap e 4. E icien Es ima ion o G ound-Le el Ai Tempe a u e in
U ban A eas using Machine Lea ning
U bClim - U-Ne
S a ion Pea son RMSE MAE MAPE
A boleda 0.98 2.71 2.13 18.83%
A igo iaga 0.99 2.07 1.58 14.37%
De io 0.98 1.75 1.35 9.0%
Deus o 0.98 2.24 1.74 8.95%
Galindo 0.98 2.3 1.77 8.58%
Pun a Galea 0.95 2.51 1.95 12.47%
Zo oza 0.98 2.29 1.77 9.0%
Table 4.1: Reg ession me ics compu ed be ween he wo models (U bClim
and U-Ne ) o each wea he s a ion du ing he pe iod o s udy.
co ela ion depending on he slope, i can be seen ha he co ela ion
be ween bo h is almos absolu e. In o he wo ds, when in he e e ence
model, Tainc eases, in he o he one also does i in he same o e y
simila way, and when in he e e ence model he empe a u e dec eases,
so does he p oposed model. The loca ion whe e he Pea son is lowe
(0.95, Pun a Galea) may no be coinciden al. Dealing wi h a cape in he
A lan ic Ocean, empe a u e changes a e mos ly a ec ed by he sea, mak-
ing i po en ially mo e challenging o he model o accu a ely model he
pa e ns o he a ge a iable Ta.
We now s ee ou ocus owa ds he o he me ics. In he same able,
we obse e ha bo h RMSE and MAE a e loa ing a ound 1.5◦C o 2.5◦C
o di e ence. As pe i s de ini ion, RMSE penalizes mo e he la ge e o s
made by he model. None heless, he highes RMSE is s ill 2.71◦C o
A boleda, an accep able alue. Fo he MAE, he alues a e lowe han
o he RMSE, a ound 1.7◦C, being he highes 2.13◦C also o A boleda,
and 1.35◦C he lowes , in De io.
Con inuing wi h he quan i a i e esul s, Table 4.2 shows a compa -
ison in e ms o he same pe o mance me ics be ween he U-Ne and
U bClim models, and he eal Tameasu emen s collec ed by he chosen
me eo ological s a ions.
U bClim - Real U-Ne - Real
S a ion Pea son RMSE MAE MAPE Pea son RMSE MAE MAPE
A boleda 0.87 7.14 5.99 33.21% 0.88 5.95 4.94 26.68%
A igo iaga 0.94 6.01 4.75 25.13% 0.95 5.25 4.28 23.22%
De io 0.90 4.97 3.80 19.60% 0.92 4.70 3.63 18.44%
Deus o 0.89 6.27 4.75 22.19% 0.91 5.56 4.37 20.98%
Galindo 0.89 6.77 5.10 23.26% 0.91 5.84 4.53 21.10%
Pun a Galea 0.87 3.95 3.05 16.40% 0.88 4.24 3.25 16.73%
Zo oza 0.89 5.75 4.20 20.13% 0.91 4.76 3.58 17.33%
Table 4.2: Reg ession me ics compu ed o he wo models (U bClim and
U-Ne ) wi h espec o he eal empe a u e da a collec ed by each wea he
s a ion du ing he pe iod o s udy. Bes esul s o e e y s a ion and sco e
a e highligh ed in ligh blue.
As can be no iced in he abo e able, esul s o he U-Ne model a e
4.3. Resul s and Discussion 83
sligh ly close o he eal measu emen s collec ed by all s a ions, excep
om Pun a Galea. None heless, he esul s a e simila o bo h models.
The U bClim simula ions exhibi spa ial a e ac s in hei Ta alues, pa -
icula ly in he o m o exagge a ed empe a u e g adien s. When using
he U-Ne model o empe a u e p edic ion, i s con olu ional p ocessing
inhe en ly smoo hs hese a e ac s. Al hough his smoo hing esul s in a
less accu a e app oxima ion o he U bClim ou pu , i has an unin ended
posi i e consequence. By blu ing he a i icial sha p g adien s p esen
in he U bClim da a, he U-Ne model p oduces es ima es ha , despi e
being a poo e i o he U bClim simula ions, align mo e closely wi h he
ac ual empe a u e alues obse ed in eali y. This sugges s ha he U-
Ne ’s abili y o mi iga e spa ial a e ac s enhances i s capaci y o gene a e
p edic ions ha a e close o eal-wo ld empe a u e dis ibu ions, e en
hough i echnically unde pe o ms in eplica ing U bClim’s ou pu .
The Pea son co ela ion o he U-Ne model is s ill qui e high and
encompasses alues om 0.88 o 0.95, being mos o hem a ound 0.9.
The model co ela es he Ta alues easonably well, wi h MAE o U-Ne
model wi h espec o he eal Ta eco ded by he wea he s a ions os-
cilla ing be ween 3.25◦C and 4.94◦C. I should be no ed ha he model
es ima es Tawi h a esolu ion o 100 me e s, while he Ta alues gi en by
he me eo ological s a ion a e single-poin measu emen s, some imes col-
lec ed in u ban a eas, which may no be as ep esen a i e as he hole a ea
co e ed by he es ima ed pixel o Ta. This can be ega ded as an in insic
sou ce o e o o he e alua ion. The e is no signi ican a ia ions in he
e o be ween di e en loca ions. Howe e , he highes e o is loca ed in
la A boleda, which is a moun ain o 329 me e s abo e sea le el, while he
lowes e o is in Pun a Galea, a me eo ological s a ion su ounded by he
ocean.
Mo ing owa ds quali a i e esul s, Figu e 4.3 shows he agg ega ed
spa ial dis ibu ion o he p edic ed Ta o 05:00 and 14:00 hou s. We
begin ou discussion on he quali a i e esul s wi h he map co esponding
o 5:00, whe e a spa ial consis ency o empe a u es can be obse ed in
he iles and hei ansi ions be ween hem. The i s hing ha can
be obse ed in his plo is he clea in luence o he sea as a la ge body
o wa e ha plays he ole o empe a u e ese oi . The a eas close
o he sea and he es ua y main ain milde empe a u es (ligh e blue).
The e ec o he sea is clea ly and co ec ly ep esen ed. I can also be
obse ed ha he mos densely popula ed a eas and whe e he la ges oad
ne wo ks a e loca ed a e he ones ha e ain mo e hea . This is due o
he land uses o hese a eas. The ma e ials ha co e hese a eas a e
gene ally asphal and conc e e, which accumula e mo e hea du ing he
day. Also, he impe iouness o he land plays an impo an ole. This
di e ence o es ima ed Taas a unc ion o land co e and impe iousness is
he expec ed om he model, since i has been ained aking in o accoun
hose da ase s. Finally, i can be seen ha he a eas wi h mo e ege a ion
and heigh such as he moun ains a e hose ha e lec lowe empe a u es
(da k blue). Speci ically, he Ganekogo a moun ain ange (1000 me e s) is
he a ea whe e he lowes empe a u es a e es ima ed, acco ding o wha
84 Chap e 4. E icien Es ima ion o G ound-Le el Ai Tempe a u e in
U ban A eas using Machine Lea ning
occu s in p ac ice. Hence, we con i m ha he U-Ne model lea ns he
ela ionship be ween he di e en spa ial da ase s ed in he inpu o he
model, such as heigh , and Ta.
Following ou quali a i e analysis, we now ocus on he map o he
14:00, which bes e lec s he he mal con as s be ween di e en zones.
The b igh colo s ep esen a highe empe a u e, and as in he ea ly mo n-
ing map, hea e en ion in u ban a eas is much highe han in g een a eas
and moun ains. In addi ion, he geog aphical idiosinc asy o Bilbao and
i s su oundings, nes led in a na ow alley, makes he mal di e ences be
e y la ge wi hin sho dis ances. Once again, he Ganekogo a moun ain
ange s ands ou as he a ea wi h he lowes empe a u es. In he same
way, bo h he sea and he moun ains a e e y well delimi ed by he model.
A he cape o Pun a Galea, in he uppe igh pa , he e is a clea change
o es ima ed empe a u e be ween he sea a ea and he land a ea.
3.1°W
3.1°W
3.05°W
3.05°W
3°W
3°W
2.95°W
2.95°W
2.9°W
2.9°W
2.85°W
2.85°W
2.8°W
2.8°W
43.2°N 43.2°N
43.25°N 43.25°N
43.3°N 43.3°N
43.35°N 43.35°N
43.4°N 43.4°N
3.1°W
3.1°W
3.05°W
3.05°W
3°W
3°W
2.95°W
2.95°W
2.9°W
2.9°W
2.85°W
2.85°W
2.8°W
2.8°W
43.2°N 43.2°N
43.25°N 43.25°N
43.3°N 43.3°N
43.35°N 43.35°N
43.4°N 43.4°N
Figu e 4.3: Agg ega ed spa ial dis ibu ion U-Ne a 05:00 ( op) and 14:00
(bo om).
Now, we u n he ocus owa ds analyzing in de ail he esul s co e-
sponding o he es iles, compa ing hem wi h he es ima ions o he
U bClim nume ical model. This analysis is done in Figu e 4.4, whe e i is
wo h no icing he di e ence in he g adien s o Ta. The i s hing ha
comes o eye in all scena ios is he di e ence in he g adien s o Ta. The
esul s o he U-Ne model show smoo h ansi ions be ween empe a u e
di e ences, wi h mo e homogeneous alues and no ab up changes. This
does no occu in he nume ical model. The eason elies on he na u e
o he con olu ional ope a ions ha ake pa in he a chi ec u e o he
U-Ne . The con olu ion ope a ion, as explained in Chap e 2 (Subsec-
ion 2.1.2), by de ini ion, ends o educe noise and o smoo hs g adien s
spa ially.
In he 5:00 maps, he spa ial esul s ob ained by he U-Ne a e e y
simila o hose o he nume ical model aken as e e ence. In he Deus o-
Zo oza ile, i is able o disce n be ween he empe a u e o he es ua y,
i s in luence in he ma gins, and he su ounding heigh s in a e y simila
way o he nume ical model. In De io, he colde zones ha appea in
4.3. Resul s and Discussion 85
he nume ical model a e also localized by he U-Ne model e en hough
hey ha e a complex geog aphy. Howe e , U-Ne ensu es a smoo he spa-
ial g adien o he es ima ed empe a u e which, as a gued p e iously,
gua an ees a mo e plausible spa ial dis ibu ion o his a iable in he a ea
unde s udy. The same happens in he case o Galindo and A igo iaga.
A Pun a Galea, he U-Ne manages o es ima e di e en empe a u es
o he small a ea o land ha appea s on he igh side o he ile. In he
14:00 maps, he same end con inues as in he p e ious ones. The spa-
ial dis ibu ion o he Taa e e y simila in he wo models. The U-Ne
model is able o di e en ia e small empe a u e pa e ns. Fo example, in
he case o A igo iaga, in he cen e o he image can be seen a colde
spo which is he ep esen a ion o he Malmasin moun ain ha s ands
ou o e he whole a ea. A simila pa e n is obse ed in La A boleda,
whe e he spa ial dis ibu ion o lowe Ta alues aligns wi h a eas o highe
al i ude and g ea e ege a ion co e age. A Pun a Galea he p o uding
land a ea is s ill es ima ed qui e accu a ely in bo h models. In he cases
o Deus o-Zo oza and Galindo, howe e , due o he homogeniza ion and
smoo hing p ocess applied by he U-Ne , spa ial de ails a e los ha do
appea in he nume ical model, such as he i e pa ha appea s in a
ligh e o ange. None heless, i can no be con i med wi hou eal da a
measu emen , ha he sha p g adien in he nume ical mode is co ec in
bo h cases.
In conclusion, i is ai o s a e ha he p oposed U-Ne model is
capable o es ima ing he spa ial dis ibu ion o Ta, achie ing a plausible
and accu a e spa ial dis ibu ion o his a iable, and ecognizing singula
spa ial pa e ns.
4.3.2 RQ2: Can an AI-based Da a Model Consis en ly
Es ima e Tempe a u e O e Time?
One o he model’s aims is o be able o es ima e he Ta o a esolu ion
o 100 me e s and wi h a empo al equency o one hou o he 164 days
chosen be ween 2008 and 2017. To answe RQ2, an hou ly ime se ies was
ex ac ed om he es ima ed Ta alues o di e en LWT days. Then, he
es ima ed Ta alues a he same loca ions as he wea he s a ions we e
compa ed bo h wi h he wea he s a ion da a and wi h he da a p o ided
by he U bClim model. The ime se ies a e p esen ed in Figu e 4.6.
The ends o he day-nigh cycles closely esemble hose o eali y.
Howe e , in bo h models (U-Ne and U bClim) a clea o e es ima ion o
he empe a u e is obse ed du ing he day cycle. This o e es ima ion,
which in some speci ic cases such as Deus o o Galindo can each mo e
han 6◦C, gene ally emains a abou 3-4 ◦C, simila o he e o s shown
abo e in Table 4.1. Du ing he nigh cycle, es ima ions a y depending on
he wea he s a ion and bo h cases o o e es ima ion and unde es ima ion
can be ound.
As shown in hese plo s, he U-Ne model e ec i ely ep oduces he
empe a u e pa e ns o he nume ical model i was ained o app oxi-
ma e, demons a ing a s ong capaci y o mi o U bClim ou pu s on an
92 Chap e 4. E icien Es ima ion o G ound-Le el Ai Tempe a u e in
U ban A eas using Machine Lea ning
hese a eas do no dissipa e hea du ing he nigh and a e subjec o s ong
hea s ess. O he wo ho spo s (bo om and op- igh o he image, shad-
owed in yellow and o ange espec i ely) a e ela ed o hea y indus y and
indus ial pa ks. The las ho spo (cen e -le , shadowed in ed) is close o
he junc ion o wo impo an highways, which gene a es a la ge asphal ed
su ace su ounded by g een a eas, leading o a high g adien in Ta.
4.4 Summa y
In he nex decade, ci ies will ine i ably ha e o implemen measu es ha
will lead hem o become mo e sus ainable and esilien places. As seen
h oughou his chap e , Taplays a pi o al ole in his desi ed u ban
sus ainabili y. U ban planne s ace he challenge o implemen ing ac ions
ha p oduce angible impac s and bene i s, equi ing access o as many
esou ces as possible. In his con ex , ha ing access o ine-g ained, hou ly
Taes ima ions o a ci y ep esen s a aluable esou ce o hei decision-
making p ocesses. Fu he mo e, i u ban planne s ha e access o ools ha
p o ide ele an in o ma ion in a sho ime, he e iciency and speed o
decision-making p ocesses can be signi ican ly imp o ed.
In his con ex , his chap e has p esen ed a new model o es ima -
ing ai empe a u e a g ound le el, wi h an accu acy o 100 me e s and
wi h a empo al esolu ion o one hou . This model, based on he U-Ne
a chi ec u e, has p o ed o be e y e icien in es ima ing he Tain a ea-
sonable ime, and wi h a much smalle amoun o a iables and da a han
he nume ical models. The h ee RQs posed in he in oduc ion o his
chap e ha e been answe ed wi h sa is ac o y esul s. On he one hand,
he model has shown ha es ima es he Tawi h a p ope spa ial esolu-
ion and consis ency. The esul s show an accep able e o and a good
co espondence no only a he ampli ude le el, bu also in cap u ing he
in a-days pa e ns o he a iable o in e es . The empo al esul s ha e
e inced ha he p opsed model excels a es ima ing he empe a u e a i-
a ions occu ing in a pa ch. I can be said, ha he p oposed model lea ns
co ec ly o co ela e he empe a u e wi h he inpu da ase s, bo h spa ial
and me eo ological. Finally, a new o mula o de ec ing ho spo s in ci ies
has been p oposed and has been p o en o be e y use ul o de ec ing
zones wi h highe hea s ess wi hin u ban a eas.
The p omising esul s epo ed in his chap e should no conceal he
ac ha se e al aspec s o he p oposed model ha e s ill oom o im-
p o emen . To begin wi h, he U-Ne model he ein p oposed has been
p o en o wo k unde ce ain condi ions. The model has demons a ed i s
alidi y jus o one scena io bu i canno be asce ained ha i is scal-
able o o he s ci ies. Also, i has been ained jus o a ew speci ic LWT
days o summe . Fo hese days U-Ne has exposed a eliable and obus
pe o mance when elici ing spa ially and empo ally cohe en es ima ions
o Ta. Howe e , o a whole yea ha obus ness has ye o be e i ied.
Addi ionally, he esul s ob ained in compa ison o he nume ical model
may sugges he possibili y o en i ely eplacing he nume ical model wi h
a da a-based model. Howe e , nume ical models canno be subs i u ed

4.4. Summa y 93
a p esen , bo h o he easons ou lined ea lie and because he model
equi es aining on an ex ensi e empe a u e da abase. A p esen , he
only way o ob ain such a la ge da abase is h ough nume ical models,
since no ci y has a wide enough senso co e age o be able o es ima e Ta
wi h a high spa ial esolu ion. Howe e , he model can be use ul o a wha
i scena io, since hanks o he compu a ional speed, empe a u es can be
es ima ed in a ew minu es a e applying changes o he inpu da a. Fo
example, i could be used o see he e ec on Ta ha a speci ic u ban
ac ion would cause, be o e ac ually implemen ing he u ban in e en ion.
Ano he op ion would be o use he ool o compa e di e en u ban mod-
i ica ions and hei e ec s, hus allowing decision make s o choose he
mos app opia e in e en ion ega ding hei impac on he u ban hea
dis ibu ion o he ci y a hand.
Besides he a o emen ioned imp o emen s, he indings discussed he e
inspi e a ious a enues o esea ch o be explo ed in he upcoming u u e.
The ex ension o he empo al ange o o he seasons should be done in o -
de o see i he model main ains i s obus ness. Analyzing and b oadening
he scope o inpu da ase s ha a e ed in o he U-Ne model could be an-
o he in e es ing s ep o wa d in he esea ch. I he model is o be ained
in di e en seasons apa om summe , NVDI should be conside ed o
example, due o i s seasonal a ia ion ha could a ec i s es ima ions.
Ano he line o esea ch could be o deploy he model in o he ci ies o
he same clima ic zone o see i i main ains sa is ac o y esul s. I he
obus ness o he model is con i med in di e en loca ions and seasons o
he same clima ic zone, he logical p og ession would be o expand he
model o o he clima ic zones, e en ually allowing o a o al eplacemen
o he nume ical model o he ze o-sho es ima ion o he empe a u e in
new u ban scena ios.
95
Chap e 5
Concluding Rema ks and
Fu u e Resea ch
As i has been ou lined h oughou his Thesis, he u u e o ci ies e-
lies in some key aspec s like sus ainabili y, esiliency and age- iendliness.
Ci ies should be planned o be accessible, inclusi e, adap able o changes
and o wa d- hinking. Mo eo e , hey should also be p epa ed o e-
spond quickly o bo h en i onmen al and socio-economic changes lea ing
no s a a o socie y behind, especially he mos ulne able g oups. In sho
ci ies mus ha e he capaci y o adap , acco ding o he imes, o he needs
o hei inhabi an s.
In his con ex , he ole o decision-making p ocesses is e ol ing o
mee he apid adap abili y equi emen s o 21s cen u y ci ies. As a e-
sul o he new digi al echnologies and a mass p oduc ion o da a, a shi
owa ds da a-d i en, in elligen sys ems is occu ing, enabling he analysis
o complex u ban issues and p o iding be e insigh s o decision-making.
Among hese ools, AI has eme ged as one o he cen al ac o s. As ex-
plained in his Thesis, AI’s abili y o p ocess as amoun s o da a, and
hidden o complex ela ionships om i , allows o mo e in o med and
e ec i e decision-making p ocesses. Also, AI can be linked wi h o he
inno a i e o isual echnologies such as Digi al Twins o enhance i s ca-
pabili ies, and o simula e and p edic a ious scena ios, enabling ci y
planne s and decision-make s o es s a egies and make be e -in o med
decisions. Summing up, AI-based echniques can be o g ea use in im-
p o ing decision-making p ocesses, accomplishing he goal o sus ainable
and iendly ci ies easie o achie e.
To his end, his Thesis has in es iga ed he capabili ies o se e al AI
echniques o sol e op imiza ion and modeling p oblems ha u ban a eas
su e om nowadays, and ha , so a ha e been ea ed wi h classical
me hods. Speci ically, wo di e en p oblems ela ed o u ban en i on-
men s ha e been ackled by using AI-based me hods: 1) an op imiza ion
p oblem o imp o e age- iendliness o he ci ies and 2) a modeling p ob-
lem o de ec and es ima e he mal s ess in u ban a eas. In wha ollows
we p esen he main con ibu ions and indings o he Thesis wi h espec
o hese wo use cases:
96 Chap e 5. Concluding Rema ks and Fu u e Resea ch
•Chap e 3.Imp o ing U ban Accessibili y and En i onmen al Con-
di ions using G aph Modeling and Mul i-objec i e Op imiza ion. A no el
amewo k o op imiza ion o u ban accessibili y in as uc u e ins alla-
ion aking in o accoun o og aphy and en i onmen al condi ions has been
p oposed, based on MOEAs and g aph modeling. The usion o geo e -
e enced g aph modeling desc ibe he u ban scena io unde conside a ion
and MOEAs has been shown o be use ul o de e mine which u ban ac ua-
ions o pe o m in ega ds o accessibili y, balancing he ade-o be ween
cos and impac on he accessibili y and walkabili y o pedes ians in hei
way o poin s o in e es . The main conclusions d awn om he esea ch
p esen ed in his chap e can be summa ized as ollows:
•The amewo k is in ui i e and based on common sense. I sugges s
asse ins alla ions in a eas wi h signi ican accessibili y, noise and ai
pollu ion p oblems, hus p o iding aluable suppo o u ban planne s
in hei decision-making p ocesses. Mo eo e , he use o MOEAs in
ou amewo k pe mi s no only o ind op imal loca ions o u ban
accessibili y based on hei con ibu ion o accessibili y, bu also accoun
o he cos o he ins alla ion and he a oidance o noisy and pollu ed
a eas o e he ci y. U ban planne s o en ha e o ake decisions wi hin
he cons ain s o public and p i a e budge s. The Thesis is one o i s
kind when conside ing cos s o he asse s and he ins alla ion p ocess
oge he wi h o he ac o s ela ed o he accessibili y and walkabili y
o pedes ians.
•F om he epo ed s a is ics on ou quali y indica o s and subsequen
signi icance analysis o obse ed di e ences, he e a e dis inc pe o -
mance dispa i ies among a ious MOEAs when implemen ed wi hin he
amewo k. In ou benchma k he chosen algo i hm was he SPEA2.
•The high-quali y and ele ance o da a a e c ucial o he applica ion
o he p oposed amewo k. The da a a ailable is no ine-g ained and
his o ical da a is o en no p ope ly a chi ed o publicly accessible. The
a ailabili y o high p ecision da a o ai quali y and noise could enhance
he algo i hm’s p ecision in he u u e.
•Chap e 4.E icien Es ima ion o G ound-Le el Ai Tempe a u e in
U ban A eas using Machine Lea ning. As discussed in he s a e-o - he-a
o Chap e 4, he use o AI o g ound-le el ai empe a u e modeling is
s ill in i s in ancy, wi h e y sca ce con ibu ions epo ed o da e in his
esea ch a ea. The Thesis has buil upon his mo i a ing esea ch niche o
p opose an encode -decode DNN a chi ec u e o es ima e he spa ial dis-
ibu ion o ai empe a u e a g ound le el in u ban en i onmen s. The
p oposed model is ed wi h in o ma ion abou land co e , impe iousness,
e ain model, and me eo ological a iables om he use case a ea and
di ec ly ou pu s a map wi h he es ima ed empe a u e o he a ea unde
s udy. The goal has been wo old: 1) o yield an es ima ion o his a i-
able in a mo e e icien ashion han nume ical models used o he same
pu pose; and 2) o de ec u ban ho spo s and suppo decision making
5.1. Resea ch Ou comes 97
p ocesses ela ed o he mi iga ion o he UHI e ec in ci ies. Ou esul s
discussed in his chap e ha e e inced ha :
•DL models, in pa icula he U-Ne a chi ec u e, ha e p o en o be a
use ul s a egy o shaping g ound-le el ai empe a u es. The p oposed
model is capable o es ima ing Tabo h spa ially and empo ally wi h
a g ea simila i y o nume ical models. Addi ionally, i has ob ained
a success ul pe o mance in he de ec ion o ho spo s wi hin u ban a -
eas. This de ec ion o ho spo s has been alida ed by expe knowledge,
as he iden i ied a eas co espond o egions whe e speci ic elemen s
o buildings exhibi he mal beha io s conduci e o he o ma ion o
UHI. This alignmen be ween de ec ed ho spo s and eal-wo ld condi-
ions demons a es he eliabili y o he de ec ion me hod.
•Thanks o i s educed compu a ional ime compa ed o la ge nume ical
models, he p oposed model can be a co e pa o a new decision-making
ool o u ban planne s. Indeed, he es ima ion o he ai empe a u e
enabled by he model can be used o he inspec ion o wha -i scena ios,
eeding cus omized spa ial da ase ha e lec possible ac ua ions o e
he a ea unde s udy. The ou pu o he model co esponding o such
cus omized inpu s can help he decision make asce ain how ac ua ions
would impac on he es ima ed ai g ound empe a u e.
•The a ge a iable modeled by he p oposed neu al a chi ec u e is he
ou pu o a nume ical model. The use o CNN laye s wi hin he a chi-
ec u e allows o a spa ial smoo hing e ec in he es ima ed ai g ound
empe a u e maps, which in u n allows o a educ ion o he spa ial a -
i ac s in he ou pu p oduced by he nume ical model. This byp oduc
yields an es ima ion ha is close o eal measu es o his a iable han
ha o he nume ical model. In o he wo ds, ou p oposed a chi ec u e
p o ides ai g ound empe a u e es ima es ha a e close o eal alues
collec ed by wea he s a ions o e he ci y.
Ne e heless, while ecognizing he exis ing po en ial, i emains o be
es ed in o he scena ios, clima es and seasons.
In conclusion, his PhD has ad anced he ield o AI-d i en modeling
and op imiza ion by demons a ing he p ac ical alue ha hese algo-
i hms can b ing in add essing c i ical u ban challenges, pa icula ly in
he a eas o accessibili y and he mal com o . By igo ously alida ing
he e ec i eness o AI ools in complex u ban en i onmen s, he esul s
epo ed in he Thesis p o ide solid e idence ha AI can be a ans o -
ma i e asse in he decision-making p ocesses o u ban planne s, enabling
mo e in o med, da a-d i en s a egies o c ea ing sus ainable, accessible,
and esilien ci ies.
5.1 Resea ch Ou comes
The esea ch ca ied ou in he cou se o his PhD Thesis led o wo con i-
bu ions o in e na ional con e ences and wo jou nal a icle, one published

98 Chap e 5. Concluding Rema ks and Fu u e Resea ch
in a JCR-indexed jou nal and he o he one cu en ly unde e iew. In
addi ion, he amewo k p esen ed in Chap e 3 has been p o ec ed ia a
pa en applica ion. De ails o hese esea ch ou comes a e p o ided below:
•Jou nal publica ions:
– Iñigo Delgado-Enales, Pa icia Molina-Cos a, Ja ie Del Se ,
“A amewo k o imp o e u ban accessibili y and en i onmen al
condi ions in age- iendly ci ies using g aph modeling and mul i-
objec i e op imiza ion”, Compu e s En i onmen and U ban Sys-
ems, Volume 102, 101966, June 2023. JCR: 7.1 (Q1), 18/182,
En i onmen al S udies.
– Iñigo Delgado-Enales, Joshua Lizundia-Loiola, Pa icia Molina-
Cos a, Ja ie Del Se , “A Machine Lea ning App oach o he E -
icien Es ima ion o G ound-Le el Ai Tempe a u e in U ban A -
eas”, U ban Clima e, JCR: 6.0 (Q1), 12/110, Me eo ology &
A mosphe ic Sciences (unde e iew).
•Con e ence publica ions:
– Iñigo Delgado-Enales, Pa icia Molina-Cos a, Eneko Osaba, Sil-
ia U a, Ja ie Del Se , “Imp o ing he U ban Accessibili y o
Olde Pedes ians using Mul i-objec i e Op imiza ion”, IEEE Con-
g ess on E olu iona y Compu a ion (CEC), Padua, I aly, pp. 1-8,
2022.
– Iñigo Delgado-Enales, Pa icia Molina-Cos a, Ja ie Del Se ,
“Spa ial Es ima ion o G ound-Le el Tempe a u e o Clima e-Sen-
si i e U ban Mobili y using Image- o-Image Deep Neu al Ne wo ks”,
IEEE In e na ional Con e ence on In elligen T anspo a ion Sys-
ems (ITSC), Bilbao, Spain, pp. 6206-6212, 2023.
•Pa en :
–“Sys em and Me hod o De e mining Ins alla ion o U ban In as-
uc u es in U ban En i onmen s”, Eu opean Pa en Applica ion
No.: 22382426.9. In en o s: Iñigo Delgado Enales, Pa icia
Molina-Cos a, Ja ie Del Se .
5.2 Fu u e Resea ch Di ec ions
This Thesis concludes ou lining he po en ial pa hs o u he explo a ion
based on he indings o he cu en esea ch done in he PhD Thesis. Fo
bo h he op imiza ion and modeling domains p omising esea ch di ec ions
will be exposed.
U ban g aph da ase s wi h inc eased node/link ea u es. S a -
ing om in elligen op imiza ion o u ban accessibili y in as uc u e, as
highligh ed be o e, en iching he geo e e enced u ban g aph model wi h
be e and mo e p ecise da ase s will signi ican ly enhance he accu acy
5.2. Fu u e Resea ch Di ec ions 99
and eliabili y o u ban planning and de elopmen s a egies. By inco po-
a ing high- esolu ion da a such as he loca ions whe e u ban asse s ha e
al eady been ins alled, including u ban opology o aking in o accoun he
pi o al ole ha s ee canyons play in he noise and pollu ion dissipa ion,
a mo e comp ehensi e and ealis ic ep esen a ion o u ban landscapes can
be c ea ed. Wi h his en iched u ban g aph model, MOEAs would ely
on la ge in o ma ion ames, inc easing hei p ecision and acili a ing a
mo e in o med decision-making p ocess. Fo ha , he apid digi aliza ion
o u ban a eas is c ucial, in o de o ha e access o b oade and iche da a.
Exploi a ion o u ban digi al wins. Ano he esea ch di ec ion wo h
o be explo ed in he u u e ocuses on he ad en o digi al winning,
which has opened up new possibili ies o u ban planning. U ban digi al
wins enable new unc ionali ies such as in e ac ing wi h he i ual u ban
landscape in eal ime, manually adding o modi ying asse s and ins an ly
obse ing he e ec s in he model [208]. Fo example, i is possible o sim-
ula e sunligh exposu e using a dynamic 3D model [209]. The eal- ime
isualiza ion p o ides aluable in o ma ion on he en i onmen al impac o
p oposed changes o he u ban planne . The amewo k p oposed in Chap-
e 3 o his Thesis could be in eg a ed in o local digi al wins o planning
mo e accessible and age- iendly ci ies, like p oposed by [210]. In ha way,
i would be possible, in pseudo- eal ime, o ecalcula e and isualize how
he in e en ions p oposed by he MOEA eshape he u ban landscape. In
he same way, ano he op ion could allow o change he mode o anspo
o he use , e.g. om walking o wheelchai , e-con igu ing all he scena io
and ideally he MOEA o e ing o he al e na i e in e en ions. In sho ,
he idea would be ha he changes could be isualized a he momen ,
o e ing a use le el expe ience whe e he u ban planne could e alua e
di e en al e na i es. To his end, howe e , a la ge imp o emen o he
compu a ional imes and be e compu a ional esou ces would be needed.
Scalabili y o he model. When i comes o modeling ai empe a-
u es in u ban a eas (Chap e 4), a p omising esea ch pa h can be o
explo e whe he he es ima ion model lea ned om da a co esponding o
a gi en ci y can gene alize well o p edic he ai g ound empe a u e in
o he ci ies. This would in ol e in es iga ing whe he he e is a co ela-
ion o Tabe ween ci ies wi h simila u ban mo phology, clima ic zones
o o he a ibu es. Fo ins ance, i would be in e es ing o p o e i da a
om an speci ic ype o ci y could be used o es ima e Ta o ano he ci y
ha sha es he same cha ac e is ics. This line o esea ch could p o ide
aluable insigh s in o he applicabili y o u ban planning models ac oss
di e en ci ies.
Unce ain y es ima ion. Ano he u u e esea ch line could be he
es ima ion o unce ain y. This would in ol e de eloping a model ha ,
in addi ion o p edic ing o es ima ing he empe a u e, also p o ides an
100 Chap e 5. Concluding Rema ks and Fu u e Resea ch
assessmen o he con idence le el o i s p edic ions. The model would
indica e a eas whe e i is mo e ce ain o i s esul s and a eas whe e he
unce ain y is highe , allowing o mo e in o med decision-making p o-
cesses. To his end, echniques o model con idence es ima ion such as
Mon e Ca lo d opou [211], Bayesian ne wo ks [212] o e iden ial o mu-
la ions [213] e o he eg ession loss used o aining he encode -decode
a chi ec u e can be conside ed.
This Thesis has in es iga ed he po en ial o a ious AI Techniques
o suppo be e -in o med decision-making p ocesses in he ield o u -
ban planning and managemen . Op imiza ion o e s g ea po en ial o
he loca ion o u ban in as uc u e o any kind conside ing mul iple ac-
o s, including physical, en i onmen al, social and economic ones. Fu he
esea ch could include housing de elopmen decisions, public acili ies en-
su ing co e age o all ci y inhabi an s, EV-cha ge s in as uc u e deploy-
men , among o he s. Modeling u ban phenomena can help o p edic u-
u e scena ios and making decisions acco dingly. Apa om empe a u e,
o he en i onmen condi ions such as Ai Quali y can be modelled linked
o ac o s such as a ic and wea he o p edic scena ios and deploy solu-
ions and policies in ad ance (such as dynamic low emissions zones in ci y
cen es) [214], [215]. Howe e , he po en ial o applying AI echniques o
complex u ban p oblems is dependan on he a ailabili y o ele an and
high-quali y da a, and he e o e i is c ucial ha ci y go e nmen s ad ance
on hei da a s a egies and go e nance o be able o g asp he ull po en-
ial o AI as a decision-suppo sys em o enhanced u ban planning and
managemen .
101
Bibliog aphy
[1] S. V. Lall, H. Selod, and Z. Shalizi, Ru al-u ban mig a ion in de el-
oping coun ies : A su ey o heo e ical p edic ions and empi ical
indings. The Wo ld Bank, 2006.
[2] H. X. Wu and L. Zhou, “Ru al- o-u ban mig a ion in china,” Asian-
Paci ic Economic Li e a u e, ol. 10, no. 2, pp. 54–67, 1996.
[3] U. N. D. o Economic and S. A ai s, Wo ld u baniza ion p ospec s:
The 2018 e ision. Uni ed Na ions, 2018.
[4] A. M. R. Nishimwe and S. Rei e , “Using a i icial in elligence mod-
els and deg ee-days me hod o es ima e he hea consump ion e o-
lu ion o a building s ock un il 2050: A case s udy in a empe a e
clima e o he no he n pa o eu ope,” Cleane and Responsible
Consump ion, ol. 5, p. 100069, 2022.
[5] W. H. O ganiza ion, Global age- iendly ci ies: A guide. Wo ld Heal h
O ganiza ion, 2007.
[6] M. Albe i, J. M. Ma zlu , E. Shulenbe ge , G. B adley, C. Ryan,
and C. Zumb unnen, “In eg a ing humans in o ecology: Oppo u-
ni ies and challenges o s udying u ban ecosys ems,” BioScience,
ol. 53, no. 12, pp. 1169–1179, 2003.
[7] O ganisa ion o economic co-ope a ion and de elopmen (oecd),h
ps://www.oecd.o g/c e/ egionalde elopmen / esilien -c
i ies.h m (Da e accessed: 13.10.2024).
[8] U. N. D. o Economic and S. A ai s, The sus ainable de elopmen
goals epo 2023: Special edi ion. Uni ed Na ions, 2023.
[9] B. Pu is, Y. Mao, and D. Robinson, “Th ee pilla s o sus ainabil-
i y: In sea ch o concep ual o igins,” Sus ainabili y science, ol. 14,
pp. 681–695, 2019.
[10] J. Lew as, “Ai pollu ion combus ion emissions: Cha ac e iza ion
o causa i e agen s and mechanisms associa ed wi h cance , ep o-
duc i e, and ca dio ascula e ec s,” Mu a ion Resea ch/Re iews in
Mu a ion Resea ch, ol. 636, no. 1-3, pp. 95–133, 2007.
[11] N. Wa s e al., “The 2020 epo o he lance coun down on heal h
and clima e change: Responding o con e ging c ises,” The Lance ,
ol. 397, no. 10269, pp. 129–170, 2021.
[12] T. R. Oke, G. Mills, A. Ch is en, and J. A. Voog , U ban Clima es.
Camb idge Uni e si y P ess, 2017.
108 Bibliog aphy
[89] A. Alkha absheh, S. Moslem, L. Oubahman, and S. Duleba, “An in-
eg a ed app oach o mul i-c i e ia decision-making and g ey heo y
o e alua ing u ban public anspo a ion sys ems,” Sus ainabili y,
ol. 13, no. 5, p. 2740, 2021.
[90] R. Ca li, M. Do oli, and R. Pelleg ino, “Mul i-c i e ia decision-
making o sus ainable me opoli an ci ies assessmen ,” Jou nal o
En i onmen al Managemen , ol. 226, pp. 46–61, 2018.
[91] M. Sibilla and F. H. Abanda, “Mul i-c i e ia decision making op i-
misa ion amewo k o posi i e ene gy blocks o ci ies,” Sus ain-
abili y, ol. 14, no. 1, p. 446, 2022.
[92] I. Iš oka O ko ić, B. Ka leuša, A. Deluka-Tibljaš, S. Šu donja, and
M. Ma ušić, “Combining a ic mic osimula ion modeling and mul i-
c i e ia analysis o sus ainable spa ial- a ic planning,” Land, ol. 10,
no. 7, p. 666, 2021.
[93] M. Cieśla, A. Sobo a, and M. Jacyna, “Mul i-c i e ia decision mak-
ing p ocess in me opoli an anspo means selec ion based on he
sha ing mobili y idea,” Sus ainabili y, ol. 12, no. 17, p. 7231, 2020.
[94] W. Jiang-Ping and T. Qun, “U ban planning decision using mul i-
objec i e op imiza ion algo i hm,” in ISECS In e na ional Collo-
quium on Compu ing, Communica ion, Con ol, and Managemen ,
IEEE, ol. 4, 2009, pp. 392–394.
[95] A. Domenico, G. Nicola, T. Daniela, C. Ful io, A. Nicola, and N.
O azio, “De no o d ug design o a ge ed chemical lib a ies based on
a i icial in elligence and pai -based mul iobjec i e op imiza ion,”
Jou nal o Chemical In o ma ion and Modeling, ol. 60, no. 10,
pp. 4582–4593, 2020.
[96] S.-Y. Shin, I.-H. Lee, D. Kim, and B.-T. Zhang, “Mul iobjec i e
e olu iona y op imiza ion o DNA sequences o eliable DNA com-
pu ing,” T ansac ions on E olu iona y Compu a ion, ol. 9, no. 2,
pp. 143–158, 2005.
[97] A. S. Yalcin, H. S. Kilic, and D. Delen, “The use o mul i-c i e ia
decision-making me hods in business analy ics: A comp ehensi e
li e a u e e iew,” Technological Fo ecas ing and Social Change, ol. 174,
p. 121193, 2022.
[98] W. Zhang, C. Wang, Y. Li, P. Wang, Q. Wang, and D. Wang,
“Seeking sus ainabili y: Mul iobjec i e e olu iona y op imiza ion
o u ban was ewa e euse in China,” En i onmen al Science &
Technology, ol. 48, no. 2, pp. 1094–1102, 2014.
[99] R. J. Balling, J. T. Tabe , M. R. B own, and K. Day, “Mul iob-
jec i e u ban planning using gene ic algo i hm,” Jou nal o U ban
Planning and De elopmen , ol. 125, no. 2, pp. 86–99, 1999.
[100] Z. Masoumi, J. Maleki, M. S. Mesga i, and A. Mansou ian, “Using
an e olu iona y algo i hm in mul iobjec i e geog aphic analysis o
land use alloca ion and decision suppo ing,” Geog aphical Analysis,
ol. 49, no. 1, pp. 58–83, 2017.

Bibliog aphy 109
[101] M. Pé es, G. Ruiz, S. Nesmachnow, and A. C. Oli e a, “Mul iob-
jec i e e olu iona y op imiza ion o a ic low and pollu ion in
Mon e ideo, U uguay,” Applied So Compu ing, ol. 70, pp. 472–
485, 2018.
[102] A Mag ini and A Liso , “Noise educ ion in e en ions in he u ban
en i onmen as a o m o con ol o indoo noise le els,” Ene gy
P ocedia, ol. 78, pp. 1653–1658, 2015.
[103] M. R. Ismail, “Quie en i onmen : Acous ics o e ical g een wall
sys ems o he islamic u ban o m,” F on ie s o A chi ec u al Re-
sea ch, ol. 2, no. 2, pp. 162–177, 2013.
[104] A. W. A. Hammad, A. Akba nezhad, and D. Rey, “Sus ainable u -
ban acili y loca ion: Minimising noise pollu ion and ne wo k con-
ges ion,” T anspo a ion Resea ch Pa E: Logis ics and T anspo a-
ion Re iew, ol. 107, pp. 38–59, 2017.
[105] X. Ning, J. Qi, C. Wu, and W. Wang, “Reducing noise pollu ion by
planning cons uc ion si e layou ia a mul i-objec i e op imiza ion
model,” Jou nal o Cleane P oduc ion, ol. 222, pp. 218–230, 2019.
[106] D. Mish a, P. Goyal, and A. Upadhyay, “A i icial in elligence based
app oach o o ecas pm2.5 du ing haze episodes: A case s udy
o Delhi, India,” A mosphe ic En i onmen , ol. 102, pp. 239–248,
2015.
[107] H. Niska, T. Hil unen, A. Ka ppinen, J. Ruuskanen, and M. Kolehm-
ainen, “E ol ing he neu al ne wo k model o o ecas ing ai pollu-
ion ime se ies,” Enginee ing Applica ions o A i icial In elligence,
ol. 17, no. 2, pp. 159–167, 2004, In elligen Con ol and Signal P o-
cessing.
[108] A. Masood and K. Ahmad, “A e iew on eme ging a i icial in-
elligence (ai) echniques o ai pollu ion o ecas ing: Fundamen-
als, applica ion and pe o mance,” Jou nal o Cleane P oduc ion,
ol. 322, p. 129072, 2021.
[109] J. Wang, L. Bai, S. Wang, and C. Wang, “Resea ch and applica-
ion o he hyb id o ecas ing model based on seconda y denoising
and mul i-objec i e op imiza ion o ai pollu ion ea ly wa ning sys-
em,” Jou nal o Cleane P oduc ion, ol. 234, pp. 54–70, 2019.
[110] S. Wicki, J. Schwaab, J. Pe hac, and A. G ê -Regamey, “Pa ici-
pa o y mul i-objec i e op imiza ion o planning dense and g een
ci ies,” Jou nal o En i onmen al Planning and Managemen , ol. 64,
no. 14, pp. 2532–2551, 2021.
[111] G. D’O so and M. Miglio e, “A gis-based me hod o e alua ing
he walkabili y o a pedes ian en i onmen and p io i ised in es -
men s,” Jou nal o T anspo Geog aphy, ol. 82, p. 102555, 2020.
[112] I. Blecic, A. Cecchini, and G. A. T un io, “Compu e -aided d a ing
o u ban designs o walkabili y,” in Compu a ional Science and I s
Applica ions, Sp inge In e na ional Publishing, 2017, pp. 695–709.
110 Bibliog aphy
[113] S. Salcedo-Sanz, L. Cuad a, E. Alexand e-Co izo, S. Jiménez-Fe n-
ández, and A. Po illa-Figue as, “So -Compu ing: An inno a i e
echnological solu ion o u ban a ic- ela ed p oblems in mod-
e n ci ies,” Technological Fo ecas ing and Social Change, ol. 89,
pp. 236–244, 2014.
[114] N. Bol en and A. Caspi, “Towa ds ou ine, ci y-scale accessibili y
me ics: G aph heo e ic in e p e a ions o pedes ian access us-
ing pe sonalized pedes ian ne wo k analysis,” PLOS ONE, ol. 16,
no. 3, pp. 1–20, 2021.
[115] G. L. Ba czyszyn, L. M. D. O. Camena , D. D. F. D. Nascimen o, e
al., “A collabo a i e sys em o sui able wheelchai ou e planning,”
T ansac ions on Accesible Compu ing, ol. 11, pp. 1–26, 2018.
[116] M. S. Rahaman, Y. Mei, M. Hamil on, and F. D. Salim, “CAPRA:
A con ou -based accessible pa h ou ing algo i hm,” In o ma ion
Sciences, ol. 385-386, pp. 157–173, 2017.
[117] W. Sasaki and Y. Takama, “Walking ou e ecommende sys em
conside ing saw c i e ia,” in Con e ence on Technologies and Appli-
ca ions o A i icial In elligence, 2013, pp. 246–251.
[118] W. Zhang, K. Cao, S. Liu, and B. Huang, “A mul i-objec i e op i-
miza ion app oach o heal h-ca e acili y loca ion-alloca ion p ob-
lems in highly de eloped ci ies such as Hong Kong,” Compu e s,
En i onmen and U ban Sys ems, ol. 59, pp. 220–230, 2016.
[119] I. Delgado-Enales, P. Molina-Cos a, E. Osaba, S. U a-U ia e, and
J. Del Se , “Imp o ing he u ban accessibili y o olde pedes ians
using mul i-objec i e op imiza ion,” in Cong ess on E olu iona y
Compu a ion, IEEE, 2022, pp. 1–8.
[120] W. Toble , “Th ee p esen a ions on geog aphical analysis and mod-
eling,” 1993.
[121] C. A. C. Coello, G. B. Lamon , D. A. Van Veldhuizen, e al.,E o-
lu iona y algo i hms o sol ing mul i-objec i e p oblems. Sp inge ,
2007, ol. 5.
[122] A. Zhou, B.-Y. Qu, H. Li, S.-Z. Zhao, P. N. Sugan han, and Q.
Zhang, “Mul iobjec i e e olu iona y algo i hms: A su ey o he
s a e o he a ,” Swa m and e olu iona y compu a ion, ol. 1, no. 1,
pp. 32–49, 2011.
[123] A. J. Neb o, J. J. Du illo, F. Luna, B. Do onso o, and E. Alba,
“Mocell: A cellula gene ic algo i hm o mul iobjec i e op imiza-
ion,” In e na ional Jou nal o In elligen Sys ems, ol. 24, no. 7,
pp. 726–746, 2009.
[124] K. Deb, K. Sindhya, and T. Okabe, “Sel -adap i e simula ed bina y
c osso e o eal-pa ame e op imiza ion,” in Gene ic and E olu-
iona y Compu a ion Con e ence, Associa ion o Compu ing Ma-
chine y, 2007, pp. 1187–1194.
[125] E. Dijks a, “A no e on wo p oblems in connexion wi h g aphs,”
Nume ische Ma hema ik, ol. 1, pp. 269–271, 1959.
Bibliog aphy 111
[126] A. de Ba celona, Diagnosis pa a la es a egia sob e cambio demog á-
ico y en ejecimien o: Una ciudad pa a odos los ciclos de ida
(2018-2030), 2018.
[127] O ganismo Au ónomo Cen o Nacional de In o mación Geog á ica
(CNIG), “Cen o de desca gas,” in h ps://cen odedesca gas
.cnig.es/Cen oDesca gas/ (Da e accessed: 13.10.2024).
[128] G. Boeing, “OSMnx: New me hods o acqui ing, cons uc ing, an-
alyzing, and isualizing complex s ee ne wo ks,” Compu e s, En-
i onmen and U ban Sys ems, ol. 65, pp. 126–139, 2017.
[129] Open s ee map,h ps: / /www. opens ee map .o g (Da e ac-
cessed: 13.10.2024).
[130] A. Bení ez-Hidalgo, A. J. Neb o, J. Ga cía-Nie o, I. O egi, and J.
Del Se , “jMe alPy: A Py hon amewo k o mul i-objec i e op i-
miza ion wi h me aheu is ics,” Swa m and E olu iona y Compu a-
ion, ol. 51, p. 100598, 2019.
[131] H. Ishibuchi, H. Masuda, Y. Tanigaki, and Y. Nojima, “Modi-
ied dis ance calcula ion in gene a ional dis ance and in e ed gen-
e a ional dis ance,” in In e na ional Con e ence on E olu iona y
Mul i-C i e ion Op imiza ion, Sp inge , 2015, pp. 110–125.
[132] H. Ishibuchi, H. Masuda, Y. Tanigaki, and Y. Nojima, “Modi-
ied dis ance calcula ion in gene a ional dis ance and in e ed gen-
e a ional dis ance,” in In e na ional Con e ence on E olu iona y
Mul i-C i e ion Op imiza ion, Sp inge , 2015, pp. 110–125.
[133] E. Zi zle and L. Thiele, “Mul iobjec i e op imiza ion using e olu-
iona y algo i hms – a compa a i e case s udy,” in Pa allel P oblem
Sol ing om Na u e, Sp inge , 1998, pp. 292–301.
[134] E. Zi zle , L. Thiele, M. Laumanns, C. Fonseca, and V. da Fonseca,
“Pe o mance assessmen o mul iobjec i e op imize s: An analysis
and e iew,” T ansac ions on E olu iona y Compu a ion, ol. 7,
no. 2, pp. 117–132, 2003.
[135] Y. Liu, J. Wei, X. Li, and M. Li, “Gene a ional dis ance indica o -
based e olu iona y algo i hm wi h an imp o ed niching me hod
o many-objec i e op imiza ion p oblems,” IEEE Access, ol. 7,
pp. 63881–63891, 2019.
[136] C. M. Fonseca, J. D. Knowles, L. Thiele, E. Zi zle , e al., “A u o-
ial on he pe o mance assessmen o s ochas ic mul iobjec i e op-
imize s,” in Thi d in e na ional con e ence on e olu iona y mul i-
c i e ion op imiza ion, ol. 216, 2005, p. 240.
[137] N. Riquelme, C. Von Lücken, and B. Ba an, “Pe o mance me -
ics in mul i-objec i e op imiza ion,” in La in Ame ican compu ing
con e ence, IEEE, 2015, pp. 1–11.
[138] J. F. Mi chell, “The “g eenhouse” e ec and clima e change,” Re-
iews o Geophysics, ol. 27, no. 1, pp. 115–139, 1989.
112 Bibliog aphy
[139] H. Rodhe, “A compa ison o he con ibu ion o a ious gases o
he g eenhouse e ec ,” Science, ol. 248, no. 4960, pp. 1217–1219,
1990.
[140] M. Székely, L. Ca le o, and A. Ga ami, “The pa hophysiology o
hea exposu e,” Tempe a u e, ol. 2, no. 4, pp. 452–452, 2015.
[141] Z. Xu, G. Fi zge ald, Y. Guo, B. Jalaludin, and S. Tong, “Impac
o hea wa e on mo ali y unde di e en hea wa e de ini ions: A
sys ema ic e iew and me a-analysis,” En i onmen In e na ional,
ol. 89-90, pp. 193–203, 2016.
[142] H. H. Kim, “U ban hea island,” In e na ional Jou nal o Remo e
Sensing, ol. 13, no. 12, pp. 2319–2336, 1992.
[143] K. De Ridde , D. Lauwae , and B. Maiheu, “U bClim–a as u ban
bounda y laye clima e model,” U ban Clima e, ol. 12, pp. 21–48,
2015.
[144] T. R. Oke, “The ene ge ic basis o he u ban hea island,” Qua e ly
Jou nal o he Royal Me eo ological Socie y, ol. 108, no. 455, pp. 1–
24, 1982.
[145] M. Moka am, F. Ta ipanah, and T. M. Pham, “In es iga ing he
e ec o su ace u ban hea island on he end o empe a u e
changes,” Ad ances in Space Resea ch, ol. 72, no. 8, pp. 3150–
3169, 2023.
[146] T. Adulkongkaew, T. Sa apanaja u, S. Cha oenhi unyingyos, and
W. Singhi unnuso n, “E ec o land co e composi ion and building
con igu a ion on land su ace empe a u e in an u ban-sp awl ci y,
case s udy in bangkok me opoli an a ea, hailand,” Heliyon, ol. 6,
no. 8, p. 4485, 2020.
[147] A. Ka imi, P. Mohammad, A. Ga cía-Ma ínez, D. Mo eno-Rangel,
D. Gachka , and S. Gachka , “New de elopmen s and u u e chal-
lenges in educing and con olling hea island e ec in u ban a -
eas,” En i onmen , De elopmen and Sus ainabili y, ol. 25, no. 10,
pp. 10485–10531, 2023.
[148] G.-J. S eene eld, S. Koopmans, B. Heusink eld, L. Van Ho e, and
A. Hol slag, “Quan i ying u ban hea island e ec s and human com-
o o ci ies o a iable size and u ban mo phology in he ne he -
lands,” Jou nal o Geophysical Resea ch: A mosphe es, ol. 116,
no. D20, 2011.
[149] X. Zhou and H. Chen, “Impac o u baniza ion- ela ed land use land
co e changes and u ban mo phology changes on he u ban hea
island phenomenon,” Science o he To al En i onmen , ol. 635,
pp. 1467–1476, 2018.
[150] H. Zhang, J. Xu, and J. Wang, “P e aining-based na u al language
gene a ion o ex summa iza ion,” in P oceedings o he Con e -
ence on Empi ical Me hods in Na u al Language P ocessing, Asso-
cia ion o Compu a ional Linguis ics, 2019, pp. 789–797.
Bibliog aphy 113
[151] L. Lebano , K. Song, and F. Liu, “Adap ing he neu al encode -
decode amewo k om single o mul i-documen summa iza ion,”
in P oceedings o he Con e ence on Empi ical Me hods in Na u al
Language P ocessing, Associa ion o Compu a ional Linguis ics,
2018, pp. 4131–4141.
[152] Y. Liu and M. Lapa a, “Tex summa iza ion wi h p e ained en-
code s,” in P oceedings o he Con e ence on Empi ical Me hods in
Na u al Language P ocessing, Associa ion o Compu a ional Lin-
guis ics, 2019, pp. 3730–3740.
[153] K. Cho, B. Van Me iënboe , C. Gulceh e, e al., “Lea ning ph ase
ep esen a ions using nn encode -decode o s a is ical machine
ansla ion,” in P oceedings o he Con e ence on Empi ical Me h-
ods in Na u al Language P ocessing, Associa ion o Compu a ional
Linguis ics, 2014, pp. 1724–1734.
[154] K. Cho, B. Van Me iënboe , D. Bahdanau, and Y. Bengio, “On
he p ope ies o neu al machine ansla ion: Encode -decode ap-
p oaches,” in P oceedings o SSST, Wo kshop on Syn ax, Seman ics
and S uc u e in S a is ical T ansla ion, Associa ion o Compu a-
ional Linguis ics, 2014, pp. 103–111.
[155] J. Long, E. Shelhame , and T. Da ell, “Fully con olu ional ne -
wo ks o seman ic segmen a ion,” in Con e ence on Compu e Vi-
sion and Pa e n Recogni ion, IEEE, 2015, pp. 3431–3440.
[156] L.-C. Chen, G. Papand eou, I. Kokkinos, K. Mu phy, and A. L.
Yuille, Deeplab: Seman ic image segmen a ion wi h deep con olu-
ional ne s, a ous con olu ion, and ully connec ed c s, 2017.
[157] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia, Py amid scene pa sing
ne wo k, 2017. a Xi : 1612.01105.
[158] R. Meh a and T. A bel, “3d u-ne o b ain umou segmen a ion,”
in In e na ional MICCAI B ainlesion Wo kshop, Sp inge , 2018,
pp. 254–266.
[159] W. Chen, B. Liu, S. Peng, J. Sun, and X. Qiao, “S3d-une : Sepa able
3d u-ne o b ain umo segmen a ion,” in B ainlesion: Glioma,
Mul iple Scle osis, S oke and T auma ic B ain Inju ies, Sp inge
In e na ional Publishing, 2019, pp. 358–368.
[160] A. Saha, Y.-D. Zhang, and S. C. Sa apa hy, “B ain umou seg-
men a ion wi h a mu i-pa hway esne based une ,” Jou nal o G id
Compu ing, ol. 19, pp. 1–10, 2021.
[161] Y. Cao, S. Liu, Y. Peng, and J. Li, “Denseune : Densely connec ed
une o elec on mic oscopy image segmen a ion,” IET Image P o-
cessing, ol. 14, no. 12, pp. 2682–2689, 2020.
[162] H. Cao, Y. Wang, J. Chen, e al., “Swin-une : Une -like pu e ans-
o me o medical image segmen a ion,” in Eu opean con e ence on
compu e ision, Sp inge , 2022, pp. 205–218.

114 Bibliog aphy
[163] X. Yan, H. Tang, S. Sun, H. Ma, D. Kong, and X. Xie, “A e -une :
Axial usion ans o me une o medical image segmen a ion,” in
P oceedings o he IEEE/CVF win e con e ence on applica ions o
compu e ision, 2022, pp. 3971–3981.
[164] R. Augus aukas and A. Lipnickas, “Pixel-wise oad pa emen de-
ec s de ec ion using u-ne deep neu al ne wo k,” in In e na ional
Con e ence on In elligen Da a Acquisi ion and Ad anced Com-
pu ing Sys ems: Technology and Applica ions, ol. 1, IEEE, 2019,
pp. 468–471.
[165] J. Jing, Z. Wang, M. Rä sch, and H. Zhang, “Mobile-une : An e i-
cien con olu ional neu al ne wo k o ab ic de ec de ec ion,” Tex-
ile Resea ch Jou nal, ol. 92, no. 1-2, pp. 30–42, 2022.
[166] W. Yao, Z. Zeng, C. Lian, and H. Tang, “Pixel-wise eg ession us-
ing u-ne and i s applica ion on pansha pening,” Neu ocompu ing,
ol. 312, pp. 364–371, 2018.
[167] A. Soni, R. Kone , and V. G. K. Villu i, “M-une : Modi ied u-ne
segmen a ion amewo k wi h sa elli e image y,” in P oceedings o
he Global AI cong ess 2019, Sp inge , 2020, pp. 47–59.
[168] Z. Wang, P. Yang, H. Liang, e al., “Seman ic segmen a ion and
analysis on sensi i e pa ame e s o o es i e smoke using smoke-
une and landsa -8 image y,” Remo e Sensing, ol. 14, no. 1, 2022.
[169] J. McGlinchy, B. Johnson, B. Mulle , M. Joseph, and J. Diaz, “Ap-
plica ion o une ully con olu ional neu al ne wo k o impe ious
su ace segmen a ion in u ban en i onmen om high esolu ion
sa elli e image y,” in In e na ional Geoscience and Remo e Sensing
Symposium, IEEE, 2019, pp. 3915–3918.
[170] M. M. Awad and M. Lau e i, “Sel -o ganizing deep lea ning (so-
une )— a no el amewo k o classi y u ban and pe i-u ban o es s,”
Sus ainabili y, ol. 13, no. 10, p. 5548, 2021.
[171] L. Wang, R. Li, C. Zhang, e al., “Une o me : A une -like ans-
o me o e icien seman ic segmen a ion o emo e sensing u -
ban scene image y,” ISPRS Jou nal o Pho og amme y and Remo e
Sensing, ol. 190, 196–214, 2022.
[172] P. Su, T. Abe a, Y. Guan, and P. Pellikka, “Image- o-image aining
o spa ially seamless ai empe a u e es ima ion wi h sa elli e im-
ages and s a ion da a,” Jou nal o Selec ed Topics in Applied Ea h
Obse a ions and Remo e Sensing, ol. 16, pp. 3353–3363, 2023.
[173] A. C. Lo enc, “Analysis me hods o nume ical wea he p edic ion,”
Qua e ly Jou nal o he Royal Me eo ological Socie y, ol. 112,
no. 474, pp. 1177–1194, 1986.
[174] P. P. Saha, K. Zeleke, and M. Ha eez, “Impac s o land use and cli-
ma e change on s eam low and wa e balance o wo sub-ca chmen s
o he Mu umbidgee Ri e in Sou h Eas e n Aus alia,” in Ex eme
Hyd ology and Clima e Va iabili y, 2019, pp. 175–190.
Bibliog aphy 115
[175] G. Fla o e al., “E alua ion o clima e models,” in Clima e change
2013: he physical science basis. Con ibu ion o Wo king G oup I
o he Fi h Assessmen Repo o he In e go e nmen al Panel on
Clima e Change, Camb idge Uni e si y P ess, 2014, pp. 741–866.
[176] M. Xue, K. K. D oegemeie , and V. Wong, “The ad anced e-
gional p edic ion sys em (ARPS)–a mul i-scale nonhyd os a ic a -
mosphe ic simula ion and p edic ion model. pa i: Model dynam-
ics and e i ica ion,” Me eo ology and A mosphe ic Physics, ol. 75,
pp. 161–193, 2000.
[177] A. Sa ka and K. De Ridde , “The u ban hea island in ensi y o
Pa is: A case s udy based on a simple u ban su ace pa ame iza-
ion,” Bounda y-laye me eo ology, ol. 138, pp. 511–520, 2011.
[178] P. Be g, O. Ch is ensen, K. Klehme , e al., “P ecipi a ion ex emes
in a eu o-co dex 0.11°ensemble a hou ly esolu ion,” Na u al Haz-
a ds and Ea h Sys em Sciences Discussions, pp. 1–21, Dec. 2018.
[179] M. J. Lipson e al., “E alua ion o 30 u ban land su ace models in
he u ban-PLUMBER p ojec : Phase 1 esul s,” Qua e ly Jou nal
o he Royal Me eo ological Socie y, ol. 150, no. 758, pp. 126–169,
2024.
[180] D. Lauwae e al., “Assessing he cu en and u u e u ban hea
island o b ussels,” U ban Clima e, ol. 15, pp. 1–15, 2016.
[181] N. Sou e ijns e al., “U ban hea in Johannesbu g and Eku huleni,
Sou h A ica: A me e -scale assessmen and ulne abili y analysis,”
U ban Clima e, ol. 46, p. 101331, 2022.
[182] C. Reis, A. Lopes, and A. S. Nou i, “Assessing u ban hea island
e ec s h ough local wea he ypes in Lisbon’s me opoli an a ea
using big da a om he Cope nicus se ice,” U ban Clima e, ol. 43,
p. 101168, 2022.
[183] D. Hidalgo-Ga cía and H. Rezapou aghdam, “Va iabili y o hea
s ess using he u bclim clima e model in he ci y o Se ille (Spain):
Mi iga ion p oposal,” En i onmen al Moni o ing and Assessmen ,
ol. 195, no. 10, p. 1164, 2023.
[184] D. Hidalgo Ga cía and J. A co Díaz, “Mi iga ion and esilience o
local clima ic zones o he e ec s o ex eme hea : S udy on he ci y
o Ba celona (Spain),” U ban Science, ol. 7, no. 4, p. 102, 2023.
[185] E. J. I. Fe nandez, M. A. R. Díaz, and A. P. Mo ales, “E ec os
del sellado del suelo en la empe a u a del ai e en la ciudad de
mu cia,” in Geog a ía: cambios, e os y adap ación: lib o de ac as.
XVIII Cong eso de la Asociación Española de Geog a ía, Log oño,
Asociación Española de Geog a ía, 2023, pp. 375–384.
[186] M. Sha i i, M. H. Shamsi, Y. Ma, and D. Lauwae , “Quan i ica ion
o he impac o global wa ming on summe o e hea ing isk in a es-
iden ial building in u ban a eas in belgium,” in Jou nal o Physics:
Con e ence Se ies, IOP Publishing, ol. 2600, 2023, p. 092016.
116 Bibliog aphy
[187] S. Vulo a, F. Meie , D. Fenne , H. Nou i, and B. Kleinschmi ,
“Summe nigh s in be lin, ge many: Modeling ai empe a u e spa-
ially wi h emo e sensing, c owdsou ced wea he da a, and ma-
chine lea ning,” IEEE Jou nal o Selec ed Topics in Applied Ea h
Obse a ions and Remo e Sensing, ol. 13, pp. 5074–5087, 2020.
[188] M. Va en so , M. K ini skiy, and V. S epanenko, “Machine lea ning
o simula ion o u ban hea island dynamics based on la ge-scale
me eo ological condi ions,” Clima e, ol. 11, no. 10, 2023.
[189] M. Fu uya, D. Fu uya, L. Y. Oli ei a, e al., “A machine lea ning
app oach o mapping su ace u ban hea island using en i onmen-
al and socioeconomic a iables: A case s udy in a medium-sized
b azilian ci y,” En i onmen al Ea h Sciences, ol. 82, p. 325, 2023.
[190] D. E demi and T. Aya a, “P edic ion o empe a u e dec easing on
a g een oo by using a i icial neu al ne wo k,” Applied The mal
Enginee ing, ol. 112, pp. 1317–1325, 2017.
[191] A. Asadi, H. A e i, and H. Fa hipoo , “Simula ion o g een oo s and
hei po en ial mi iga ing e ec s on he u ban hea island using an
a i icial neu al ne wo k: A case s udy in aus in, exas,” Ad ances
in Space Resea ch, ol. 66, pp. 1846–1862, 2020.
[192] S. Chan and C. Chau, “De elopmen o a i icial neu al ne wo k
models o p edic ing he mal com o e alua ion in u ban pa ks in
summe and win e ,” Building and En i onmen , ol. 164, p. 106364,
2019.
[193] F. Toma is, F. J. Diez, M. S. Wilhelm, and L. M. Na as-G acia,
“P edic ion o daily ambien empe a u e and i s hou ly es ima-
ion using a i icial neu al ne wo ks in u ban allo men ga dens and
an u ban pa k in Valladolid, Cas illa y Leon, Spain,” Ag onomy,
ol. 14, no. 1, 2024.
[194] A. Ash iani, P. Mi zaei, and F. Haghigha , “Indoo he mal condi-
ion in u ban hea island: Compa ison o he a i icial neu al ne -
wo k and eg ession me hods p edic ion,” Ene gy and Buildings,
ol. 76, pp. 597–604, Jun. 2014.
[195] J. W. Oh, J. Nga ambe, P. N. Duhi we, G. Y. Yun, and M. San a-
mou is, “Using deep-lea ning o o ecas he magni ude and cha -
ac e is ics o u ban hea island in Seoul Ko ea,” Scien i ic Repo s,
ol. 10, no. 1, pp. 1–13, 2020.
[196] K. Gobakis, D. Koloko sa, A. Synne a, M. Salia i, K. Giannopoulou,
and M. San amou is, “De elopmen o a model o u ban hea island
p edic ion using neu al ne wo k echniques,” Sus ainable Ci ies and
Socie y, ol. 1, no. 2, pp. 104–115, 2011.
[197] G. Assa , X. Hu, and R. H. Assaad, “P edic ing u ban hea island
se e i y on he census- ac le el using bayesian ne wo ks,” Sus ain-
able Ci ies and Socie y, ol. 97, p. 104756, 2023.
Bibliog aphy 117
[198] S. Liu, J. Zhang, J. Li, Y. Li, J. Zhang, and X. Wu, “Simula ing
and mi iga ing ex eme u ban hea island e ec s in a ac o y a ea
based on machine lea ning,” Building and En i onmen , ol. 202,
p. 108051, 2021.
[199] F. B iegel, O. Makansi, T. B ox, A. Ma za akis, and A. Ch is en,
“Modelling long- e m he mal com o condi ions in u ban en i on-
men s using a deep con olu ional encode -decode as a compu a-
ional sho cu ,” U ban Clima e, ol. 47, p. 101359, 2023.
[200] H. E. Beck, N. E. Zimme mann, T. R. McVica , N. Ve gopolan, A.
Be g, and E. F. Wood, “P esen and u u e köppen-geige clima e
classi ica ion maps a 1-km esolu ion,” Scien i ic da a, ol. 5, no. 1,
pp. 1–12, 2018.
[201] J. A. Ace o and K. He anz-Pascual, “A compa ison o he mal com-
o condi ions in ou u ban spaces by means o measu emen s and
modelling echniques,” Building and En i onmen , ol. 93, pp. 245–
257, 2015.
[202] H. Hooybe ghs, J. Be ckmans, D. Lauwae , F. Le eb e, and K.
De Ridde , “Clima e a iables o ci ies in eu ope om 2008 o
2017,” Cope nicus Clima e Change Se ice (C3S) Clima e Da a
S o e (CDS), 2019.
[203] H. He sbach e al., “The ERA5 global eanalysis,” Qua e ly Jou -
nal o he Royal Me eo ological Socie y, ol. 146, no. 730, pp. 1999–
2049, 2020.
[204] J. Hidalgo and R. Jougla, “On he use o local wea he ypes clas-
si ica ion o imp o e clima e unde s anding: An applica ion on he
u ban clima e o Toulouse,” Plos One, ol. 13, no. 12, pp. 1–21,
Dec. 2018.
[205] Euskalme ,h ps://www.euskalme .euskadi.eus/di ulgacion
/manual-de-es ilo/ (Da e accessed: 13.10.2024).
[206] Eu opean cen e o medium- ange wea he o ecas s,h ps://co
n luence.ecmw .in /display/CKB/Clima e+ a iables+ o +c
i ies+in+Eu ope+ om+2008+ o+2017+documen a ion (Da e
accessed: 13.10.2024).
[207] Eu opean En i onmen Agency, DG Join Resea ch Cen e o he
Eu opean Commission, “Cope nicus land moni o ing se ice,” in
h ps://land.cope nicus.eu/en.
[208] F. Dembski, U. Wössne , M. Le zgus, M. Rudda , and C. Yamu,
“U ban digi al wins o sma ci ies and ci izens: The case s udy
o he enbe g, ge many,” Sus ainabili y, ol. 12, no. 6, 2020.
[209] G. Sch o e and C. Hü zele , “The digi al win o he ci y o
zu ich o u ban planning,” PFG–Jou nal o Pho og amme y, Re-
mo e Sensing and Geoin o ma ion Science, ol. 88, no. 1, pp. 99–
112, 2020.