Towa ds a Deep Au oma ic Gene a ion o Figu e-g ound Maps
Lukas A zoumanidis, Jona han Hech , Youness Dehbi
Compu a ional Me hods Lab, Ha enCi y Uni e si y, Hambu g, Ge many - (lukas.a zoumanidis, jona han.hech , youness.dehbi)
@hcu-hambu g.de
Keywo ds: Gene a i e Ad e sa ial Ne wo ks, Geog aphical Da a T ansla ion, Figu e-g ound Maps, U ban Mo phology, Buil
Densi y, Volun ee ed Geog aphic In o ma ion.
Abs ac
Figu e-g ound maps play a key ole in many disciplines whe e u ban planning o analysis is in ol ed. In his con ex , he au oma ic
gene a ion o such maps wi h espec o ce ain equi emen s and cons ain s is an impo an ask. This pape p esen s a i s s ep
owa ds a deep au oma ic gene a ion o igu e-g ound maps whe e he buil densi y o he gene a ed scenes is con olled and aken
in o accoun . This is p e o med building upon a Geog aphic Da a T ansla ion model which has been applied o gene a e less
a ailable geospa ial ea u es, e.g. building oo p in s, om mo e widely a ailable geospa ial da a, e.g. s ee ne wo k da a, using
condi ional Gene a i e Ad e sa ial Ne wo ks. A no el p ocessing app oach is in oduced o inco po a e he popula ion densi y and
he buil densi y acco dingly. Fu he mo e, he impac o bo h he le el o de ail o he s ee ne wo k, i.e. i s spa si y o densi y,
and he spa ial esolu ion o he aining da a on he gene a ed igu e-g ound maps has been in es iga ed. The gene a ed maps and
he quali a i e esul s e eal an ob ious impac o hese pa ame e s on he layou o buil and unbuil a eas. Ou app oach pa es he
way o he expansion o exis ing dis ic s by igu e-g ound maps o u u e neighbou hoods conside ing ac o s such as densi y and
u he pa ame e s which will be subjec o u u e wo k.
1. In oduc ion
To coun e ac he g ow h o sealed su aces and simul aneously
c ea e new li ing space in ci ies, u ban planning plays a i al
ole. Cu en pa adigms in u ban planning sugges o ebuild
o upg ade exis ing pa cels o buildings ins ead o c ea ing new
dis ic s on unde eloped land. In such case, a cons uc ion com-
pany ede elops a pa cel wi h a goal, e.g. o maximize e -
enue o o build he mos possible la s while espec ing ce ain
cons ain s, e.g. he building law o he physical dimensions
o he pa cel. In an ini ial planning phase, a d a ep esen -
ing he b oad layou o he s ee ne wo k and he a angemen
and geome y o he building oo p in s is laid ou . This ini ial
d a ing p ocess lays he ounda ion o he ollowing planning
phases, whe e each phase acili a es he d a u he wi h de-
ail o e en ually become a si e plan o a de elopmen p ojec .
Usually, he c ea ion o such d a s builds upon he so called
igu e-g ound maps as hose isualize he exis ing buil s uc-
u es while demons a ing he ela ionship be ween buil and
unbuil spaces in ci ies. In such maps, buildings a e depic ed
as black solid mass ( igu e) while s ee s o open spaces a e
ep esen ed as whi e oid (g ound) (Wang e al., 2024).
In p ac ice, i is a ely su icien o c ea e a single d a o a
u ban de elopmen p ojec . Ins ead, se e al hund eds o ini-
ial d a s migh be needed o isualize di e en and al e na -
i e design ideas. Howe e , d a ing is o en a edious and ime
consuming p ocess as i in ol es mainly analogous o digi al
manual d awing o c a ing. Depending on he scale and spe-
ci ica ions o he pa cel and he expe ience le el o he plan-
ne , such a ask could be o e whelming. Besides, he e a e
many u he equi emen s and cons ain s o conside such as
he desi ed le el o building densi y, he cons uc ion law o he
physical dimensions o he plo . To suppo planne s wi h his
endea ou , his pape p oposes a i s app oach owa ds a deep
au oma ic gene a ion o igu e-g ound maps wi h he possibili y
o con ol he buil densi y o u ban s uc u es. This app oach
pa es he way o he expansion o exis ing dis ic s by igu e-
g ound maps o u u e neighbou hoods conside ing in luencing
ac o s such as popula ion and buil densi y.
One ecen esea ch ield ha could pa e he way owa ds he
au oma ic gene a ion o igu e-g ound maps is Geog aphic Da a
T ansla ion (GDT) which ollows he idea o gene a ing less
abundan geospa ial da ase s by lea ning associa ions om o he
mo e plen i ul da ase s. In his con ex , wo new app oaches
GANmappe and Ins an Ci y which use condi ional Gene a i e
Ad e sa ial Ne wo ks (cGAN) o ackle GDT ha e been p o-
posed (Wu and Biljecki, 2023, 2022). These app oaches capi -
alize on mo e commonly ound geospa ial ea u es, e.g. s ee
ne wo ks, o gene a e less common ea u es such as building
oo p in s by le e aging on hei mu ual ela ionship.
This pape in es iga es whe he he ecen ly p oposed app oach
o Wu and Biljecki (2023) is able o cap u e and e lec highe
and lowe buil densi ies o u ban s uc u es. By le e aging a
la en ela ionship be ween he popula ion and he buil densi y,
we p esen a no el aining da a p ocessing ha p ecisely cap-
u es he desi ed le el o buil densi y which will be explained
in mo e de ail in Sec ion 3.2. Hence, he main con ibu ion o
his pape consis s in he inco po a ion o new ac o s in o he
au oma ic gene a ion o igu e-g ound maps. In pa icula , we
ocused on he assessmen o he impac o he le el o de ail
o he unde lying s ee ne wo k, he spa ial esolu ions and he
buil densi y.
The emainde o his pape is as ollows. Be o e going in o he
de ails o he me hodology o ou igu e-g ound map gene a ion
in Sec ion 3, we gi e a e iew o ela ed wo k in Sec ion 2.
The se up o ou expe imen s and hei indings a e p esen ed
in Sec ion 4. Sec ion 5 concludes he pape and gi es ou looks
o u u e esea ch.
ISPRS Annals o he Pho og amme y, Remo e Sensing and Spa ial In o ma ion Sciences, Volume X-4/W5-2024
19 h 3D GeoIn o Con e ence 2024, 1–3 July 2024, Vigo, Spain
This con ibu ion has been pee - e iewed. The double-blind pee - e iew was conduc ed on he basis o he ull pape .
h ps://doi.o g/10.5194/isp s-annals-X-4-W5-2024-33-2024 | © Au ho (s) 2024. CC BY 4.0 License.
33
2. Rela ed Wo k
The analysis o he su ounding u ban mo phology is unda-
men al o new building p ojec s. Acco ding o Wang e al.
(2024) u ban mo phology s udies hea ily ely on mo phome -
ics such as building oo p in s o s ee leng hs. In his con-
ex , hey de eloped a me hod o lea ning mo phology ea u es
based on igu e-g ound maps whe e hey compa e u ban o m
ypes in a ully unsupe ised manne . In hei app oach, hey
apply a isual ep esen a ion lea ning model called SimCLR o
cap u e he layou o building g oups (Wang e al., 2024). How-
e e , se e al ac o s may in luence he unde lying mo phology.
In his con ex , Hijazi e al. (2017) enginee ed a GIS-based
app oach o quan i y, o ins ance, he homogenei y o u ban
s uc u es as an impo an ac o by ex ac ing a ibu es such as
he angles o dis ances be ween buildings di ec ly om building
oo p in s. In o de o inco po a e such in luencing ac o s, ou
app oach builds upon ideas o Wu and Biljecki (2023, 2022).
These app oaches ha e been dedica ed o c ea e geospa ial da a
using condi ional Gene a i e Ad e sa ial Ne wo ks ollowing
he image- o-image ansla ion pa adigm. In hei wo k, s ee
ne wo k da a is used as inpu o gene a e building oo p in s
wi h he goal o suppo applica ions in he ields o u ban mo -
phology analysis and u ban simula ions which o en su e om
missing spa ial de ails.
Recen ly, an app oach de eloping a GAN-based end- o-end gen-
e a i e model o he 2D and 3D building layou gene a ion has
been p oposed (Jiang e al., 2023b). They s a ed ha mos ap-
p oaches o e look he impac o si e a ibu es on buil s uc-
u es, hampe ing hei po en ial o u he e alua ion and in-
o med decision making. The au ho s conduc ed expe imen s
wi h condi ional ec o s showing imp o ed pe o mance in di -
e en scena ios. Simila ly, Jiang e al. (2024) p oposed an
app oach o au oma ing si e planning wi h in eg a ed domain
knowledge o he buil en i onmen wi h he goal o imp o e
con ex -awa eness. Thei de eloped Gene a i e Ad e sa ial Ne -
wo k called CAIN-GAN is supposed o no only be capable
o syn hesized isually ealis ic and seman ically easonable
design solu ions bu also use ul o u ban sus ainabili y simu-
la ions.
Many esea che s applied o mal g amma s, pa icula ly shape
g amma s (S iny, 1980), as ano he pa adigm o gene a e eal-
wo ld man-made objec s. Fo ins ance, Gong e al. (2020) gen-
e a ed u ban ab ic in o hogonal and non-o hogonal u ban land-
scapes based on a p e-de ined se o shape ules. Howe e ,
hese app oaches a e in gene al su e ing unde he o e head o
he manual design o he g amma ules by expe s. Fo mo e
de ails on he opic o gene a i e u ban design, he in e es ed
eade is e e ed o Jiang e al. (2023a).
3. Me hodology
The ollowing sec ions will in oduce he acquisi ion and p o-
cessing o aining da a and desc ibe he applied model o he
gene a ion o building oo p in s based on di e en s ee ypes
building upon he ideas o Geog aphic Da a T ansla ion p esen-
ed by Wu and Biljecki (2022) and u he aking he popula ion
and s ee ne wo k densi y in o accoun .
3.1 Model a chi ec u e
The ne wo k a chi ec u e o he applied model is a ype o Image-
o-Image condi ional GAN (Isola e al., 2017) ha can ansla e
Figu e 1. Ins an CITY a chi ec u e which ou app oach builds
upon (Wu and Biljecki, 2023).
inpu image da a such as s ee ne wo ks o an ou pu image a-
cili a ed wi h gene a ed building oo p in s (Wu and Biljecki,
2023). Using di e en esolu ions, he unde lying ne wo k is
ained on wo esidual ne wo ks namely a Global Gene a o
and a Local Enhance . In each o wa d pass, he gene a o
ies o c ea e ou pu s ha could ’ ick’ he disc imina o in o
hinking he ou pu s a e ’ eal’ while he disc imina o will lea n
o classi y he ou pu s as ’ ake’ and he g ound u h as ’ eal’
(Wu and Biljecki, 2023). Disc imina o s in he model compa e
he encoded ea u es o m bo h he gene a ed images and he
g ound u h o indica e whe he he gene a ed image is eal o
ake. A he end o each o wa d pass, he loss o he gene a o
and he disc imina o is e alua ed and hei weigh s a e upda ed
acco dingly un il he gene a o ’s ou pu s a e compelling enough
so ha he disc imina o de ec s hal o hem as eal images (Wu
and Biljecki, 2023). Figu e 1 shows an o e iew on he ne wo k
a chi ec u e o Ins an CITY which builds he basis o ou pape
aking u he cons ain s in o conside a ion.
3.2 P ocessing
To gene a e igu e-g ound maps, he p e iously desc ibed model
is p o ided wi h aining da a consis ing o wo di e en se s
o image pai s. Each image pai consis s o an inpu and an
ou pu image whe e he model is supposed o lea n he isual
pa e ns o ans o m an inpu image o an ou pu image acco d-
ingly. An inpu image esembles he s ee ne wo k wi h di -
high densi y
mid densi y
1:4000
spa se ne .
dense ne .
1:8000
Figu e 2. Example o aining image pai s o di e en da ase s
showing spa ial esolu ions o 1:4000 s. 1:8000, spa se s.
dense s ee ne wo k and high s. mid buil densi y.
ISPRS Annals o he Pho og amme y, Remo e Sensing and Spa ial In o ma ion Sciences, Volume X-4/W5-2024
19 h 3D GeoIn o Con e ence 2024, 1–3 July 2024, Vigo, Spain
This con ibu ion has been pee - e iewed. The double-blind pee - e iew was conduc ed on he basis o he ull pape .
h ps://doi.o g/10.5194/isp s-annals-X-4-W5-2024-33-2024 | © Au ho (s) 2024. CC BY 4.0 License.
34
colo line wid h (in px) ype
# b301 12 & 9 unk, unk link
#840000 7.5 & 6 mo o way, mo o way link
# 1a01 12 & 6 p ima y, p ima y link
#014182 9 & 3 seconda y, seconda y link
#58d751 6 & 3 e ia y, e ia y link
#75bb d 3 access oads
Table 1. CRHD con igu a ion applied in his wo k.
e en s ee ypes indica ed by di e en line-wid h and colo -
ing acco dingly. The ou pu image ep esen s he igu e-g ound
map showing he s ee ne wo k including building oo p in s.
Based on he oad ne wo k and he building oo p in s ha a e
ex ac ed om OpenS ee Map (OSM), we can gene a e an a -
bi a y amoun o such aining image pai s. As men ioned,
he geome y o he s ee ne wo k has o ep esen he s ee
ype and he e o e is con e ed in o a Colou ed Road Hie a chy
Diag am (CRHD) as p oposed by (Wu and Biljecki, 2023). A
CRHD di e en ia es s ee ypes by hei assigned line-wid h
and colo ing as can be seen in Table 1 and Figu e 2 (Chen e al.,
2021). G ound u h images a e ende ed based on he build-
ing oo p in geome y as well as he s ee ne wo k geome y
whe e e e y geome y is colo ed in solid black and he s ee
ne wo k geome y which has he same line-wid h as de ined o
he CRHD used in he inpu images. The ypes o oads and
pa hs which ha e been selec ed a e also lis ed in Table 1. As
hey belong o he same le el o hie a chy, he ypes esiden ial,
se ice, li ing s ee , oo way, pa h, pedes ian and unclassi ied
a e g ouped in o access oads. Each hie a chy has been asso-
cia ed o a p e-de ined colo . Fo he sake o eplicabili y, he
used colo s and he line wid hs a e also lis ed in he same able.
The Eu opean Commission and he Fede al Minis y o T ans-
po and Digi al In as uc u e (Bundesminis e ium ¨
u Ve keh
und digi ale In as uk u ) di ided he popula ion densi y o
Ge many in o h ee di e en ca ego ies o con iguous g id cells
o 1km2as highligh ed in Figu e 3 Eu opean Commission
(2016); BMVI (2018). A high-densi y g id cell (ci y) co es-
ponds o a popula ion densi y o a leas 1.500 inhabi an s pe
km2wi h a minimum o al popula ion o 50.000. A mid-densi y
g id cell (subu b) e e s o a popula ion densi y o a leas 300
inhabi an s pe km2wi h a minimum o al popula ion o 5.000.
Alow-densi y g id cell has a popula ion densi y below 300 in-
habi an s pe km2. In his wo k, hese ca ego ies will be ap-
plied as a empla e o ex ac ing he aining da a om OSM
acco ding o hei popula ion densi y as highligh ed in Figu e 4.
As pe o med o he ci y o Hambu g, we o e lay he popula-
ion densi y and ex ac aining images acco ding o hei spa-
ial bounda y o he en bigges ci ies in Ge many acco ding o
hei o al popula ion. In he ollowing, he popula ion densi y
based ex ac ion o aining images summa ized in Figu e 4 will
be explained in mo e de ail.
To ende inpu and ou pu image pai s, he A las- ool a ail-
able in QGIS has been used. This ool allows o subsequen
map ende ing o au oma ically ex ac he aining images o
a speci ic spa ial ex en , scale and densi y ca ego y as can be
seen in bo h Figu e 2 and Figu e 4. Subsequen ly, he aining
images o each popula ion densi y a e ende ed in a esolu-
ion o 1024×1024 pixel and s o ed as a PNG ile o ma ch he
inpu size and ile ype equi emen s o he unde lying model
designed by Wu and Biljecki (2023).
In o al, we gene a e di e en aining da ase s wi h he ollow-
ing cha ac e is ics:
high densi y
mid densi y
low densi y
missing da a
popula ion >= 1500
popula ion >= 300
popula ion <300
Figu e 3. Popula ion densi y g id cells ep esen ing de ined
ca ego ies o he ci y o Hambu g, Ge many. Colo s dis inguish
he di e en popula ion densi ies. The ca ego ies ha e been
de i ed acco ding o he de ini ion o he Eu opean Commission.
•Spa ial esolu ion: 1:4000 o 1:8000
•Buil densi y: high o mid
•S ee ne wo k: spa se o dense
Fo each aining da ase , we ex ac ed he image pai s o a
scale o 1:4000 and 1:8000 in o de o p o ide high de ails o in-
di idual buildings o a neighbo hood o building blocks as high-
ligh ed in Figu e 2. Addi ionally, acco ding o Wu and Biljecki
(2023) a hese scales he model is supposed o be able o gen-
e a e a i icial images wi h sha pe co ne s ha would ep esen
eal images mo e closely. The added alue o his pape , is he
conside a ion o he densi y o he s ee ne wo k and, hence,
he in es iga ion o i s po en ial impac on he ained model
and he acco ding gene a ed igu e-g ound maps. The e o e,
we c ea ed wo di e en aining da ase s. The i s one does
no include access oads and, hence, ep esen s spa se s ee
ne wo k da a, whe eas he second comp ises such oads and,
hus, e lec s dense s ee ne wo k da a as can be seen in Fig-
u e 2. Fu he , he popula ion and he acco ding buil densi y
has been aken in o accoun in o de o gene a e igu e-g ound
maps depending on his pa ame e . Thus, he model has been
ained on wo di e en high and mid buil densi ies as al eady
depic ed in Figu e 2.
In gene al, ca ego izing he aining da a a e he abo e men-
ioned le els o popula ion densi ies esul s in he expec ed ou -
comes. Fo egions wi h high popula ion densi ies, we can ob-
se e a high buil densi y ep esen ed h ough closed building
de elopmen . The la e is cha ac e ized by a con inuous and
cohesi e buil s uc u e. Simila ly, mid buil densi ies show
he expec ed open building de elopmen ypically exp essed by
a mo e spacious and less densely cons uc ed buil s uc u e.
Howe e , his pa e n is some imes iola ed con as ing he un-
de lying aining da a.
To o e come his p oblem, we me ged bo h he high and he mid
buil densi y aining da a esul ing in wo da ase s o spa se
and dense s ee ne wo ks. Analogously, wo da ase s wi h di -
e en spa ial esolu ions, namely 1:4000 and 1:8000 ha e been
ISPRS Annals o he Pho og amme y, Remo e Sensing and Spa ial In o ma ion Sciences, Volume X-4/W5-2024
19 h 3D GeoIn o Con e ence 2024, 1–3 July 2024, Vigo, Spain
This con ibu ion has been pee - e iewed. The double-blind pee - e iew was conduc ed on he basis o he ull pape .
h ps://doi.o g/10.5194/isp s-annals-X-4-W5-2024-33-2024 | © Au ho (s) 2024. CC BY 4.0 License.
35
Ou pu
Inpu
OSM
Pixel-based
il e ing
High
Mid
High
Mid
Popula ion-based il e ing
High popula ion
Mid popula ion
densi y
densi y
Figu e 4. Illus a ion o ou aining da a p ocessing pipeline. High, mid and low popula ion densi y da a is p o ided as 1km2g id
cells by Eu os a bu only high and mid popula ion densi y da a is p ocessed.
gene a ed. To u he e ine he ou da ase s and hence e lec
well he le el o densi y, we apply a pixel-based il e ing. He e-
wi h, image pai s a e so ed acco ding o hei p opo ion o
unbuil a ea. An image comp ising mo e han 82% whi e pixels
is hen emo ed om he aining da ase . Images wi h less
han 72% whi e pixels ha e been assigned o he aining da a-
se o high buil densi ies whe eas hose ha ing a p opo ion
o whi e pixels be ween 72% and 82% ha e been associa ed
o he aining da ase wi h mid buil densi y acco dingly. The
h esholds used o e ine ou aining da ase s ha e been de e m-
ined based on a quali a i e isual inspec ion o he unde lying
images. Since he le els o buil densi y migh di e om coun-
y o coun y, he h esholds migh be e-e alua ed o ci ies
ou side o Ge many. In Ge man ci ies, a low popula ion dens-
i y can be p edominan ly s a ed in indus ial dis ic s such as
he sou h o Hambu g as depic ed in Figu e 3. In his pape ,
we ocused on dis ic s wi h mid and high densi y and omi ed
hose co esponding o low densi y. Fo each aining da ase ,
he image pai s a e spli in o aining and es ing da a wi h a
p opo ion o 90% and 10% espec i ely.
4. Expe imen al Resul s
The ollowing sec ion desc ibes he aining da a in mo e de ails
and gi es insigh in o he achie ed expe imen al esul s which
ha e been e alua ed by quali a i e isual inspec ion and quan -
i a i e me ics.
4.1 Da ase s
To p oduce ou aining da a, we use OSM1and popula ion
densi y da a2. In o de o e lec he u ban mo phology o Ge -
man ci ies, we ocus on he en bigges ci ies o aining and
gene a ing igu e-g ound maps. This allows o add essing di -
e en popula ion densi ies as impo an ac o in luencing he
design o such diag ams. The ollowing en mos popula ed ci -
ies in Ge many ha e been selec ed acco dingly: Be lin, Ham-
1h ps://download.geo ab ik.de/eu ope/ge many.h ml
2h ps://ec.eu opa.eu/eu os a /de/web/gisco/geoda a/
e e ence-da a/popula ion-dis ibu ion-demog aphy/
geos a
bu g, Munich, Cologne, F ank u , S u ga , D¨
usseldo , Leip-
zig, Do mund, Essen (S a is a, 2024). A e downloading he
OSM da a o each ci y, he s ee ne wo k da a and he build-
ing oo p in s a e c opped o he adminis a i e bounda ies o
each espec i e ci y. Analogously, he popula ion densi y da a is
s uc u ed in con iguous g id cells o size 1km2and ma ched o
he adminis a i e bounda ies o he a o emen ioned en bigges
ci ies in Ge many.
4.2 Con igu a ion & Accu acy Measu es
We applied he model p oposed by Wu and Biljecki (2023) ol-
lowing he ad ised hype pa ame e s which yielded he bes es-
ul s in hei expe imen s. The model was implemen ed in Py-
To ch and CUDA accele a o . In his con ex , ou expe imen s
ha e been conduc ed wi h PyTo ch e sion 2.2.0wi h CUDA
e sion 12.3. While aining, he model consumed oughly 11
GB o VRAM. A s a ing lea ning a e o 0.0002 and a ba ch
size o 1ha e been chosen wi h a aining ime amoun ing 100
epochs o each expe imen .
One o he mos commonly applied measu e o assess he pe -
o mance o a GAN is he F ´
eche Incep ion Dis ance (FID)
p oposed by Heusel e al. (2017). The FID is a dis ance me -
ic which cap u es he simila i y o images c ea ed by gene -
a i e models o eal images and he eby exp esses he models
o e all quali y. Fo mally, wi h FID we measu e he di e ence
be ween wo Gaussian dis ibu ions using he F ´
eche dis ance
(Ha -Peled and Raichel, 2014). In his wo k, he i s dis ibu-
ion co esponds o he da ase o images gene a ed by ou ap-
p oach, while he o he dis ibu ion co esponds o he e e ence
da ase o images con aining he g ound u h, i.e. igu e-g ound
maps o iden ical spa ial loca ion. Each pa ame ic Gaussian
dis ibu ion is de ined by i ’s mean µand co a iance ma ix Σ
which ha e been de i ed a e ans o ming he eal and gene -
a ed images om he image space o la en ec o embeddings
using he inal coding laye o an incep ion model as can be
seen in Equa ion 1. This way, he FID compa es ea u es ha
co espond o eal-wo ld objec s a he han di ec ly compa -
ing gene a ed and eal images pixel-wise. As a esul , a highe
FID indica es a la ge di e ence be ween he gene a ed and he
eal image while a smalle FID indica es a highe simila i y.
ISPRS Annals o he Pho og amme y, Remo e Sensing and Spa ial In o ma ion Sciences, Volume X-4/W5-2024
19 h 3D GeoIn o Con e ence 2024, 1–3 July 2024, Vigo, Spain
This con ibu ion has been pee - e iewed. The double-blind pee - e iew was conduc ed on he basis o he ull pape .
h ps://doi.o g/10.5194/isp s-annals-X-4-W5-2024-33-2024 | © Au ho (s) 2024. CC BY 4.0 License.
36
In ou case, his is in gene al op ically e lec ed by mo e co -
ec ep esen a ions o shapes, sizes, and densi ies o gene a ed
building oo p in s.
F ID =||µX−µY||2−T (ΣX+ ΣY−2√ΣXΣY)(1)
4.3 In e p e a ion & Discussion
1)
gene a ion
eal
high buil densi y
spa se s ee ne .
dense s ee ne .
2)
eal
gene a ion
Figu e 5. Compa ison o wo gene a ed igu e-g ound maps wi h
a high buil densi y ained wi h spa se (le ) and dense ( igh )
s ee ne wo k da a wi h a spa ial esolu ion o 1:8000.
Va ying he abo e men ioned ac o s and he esolu ion, we
ained a model in o de o gene a e igu e-g ound maps ac-
co dingly. Expe imen al esul s shown in Figu e 5 highligh
he di e ences in he gene a ion o high densi y igu e-g ound
maps using spa se o dense s ee ne wo k aining da a. Pa -
icula ly, clea o see is ha he model is much be e o gene a e
inne cou ya ds which a e ypical o closed building de elop-
men in high buil densi y dis ic s in Ge many when small ac-
cess oads o hese inne cou ya ds a e p o ided o he model
du ing aining. Mo eo e , a dense , mo e de ailed s ee ne -
wo k seems o ob iously imp o e he gene a ion o buil s uc-
u es ha ollow along he geome y o s ee s as highligh ed
by Figu e 5. In e es ingly, he model does no seem o ge o e -
whelmed when lea ning on a dense s ee ne wo k as he s ee s
a e always gene a ed supe bly wi h only mino di e gence om
he inpu image.
1)
gene a ion
eal
mid buil densi y
spa se s ee ne .
dense s ee ne .
2)
eal
gene a ion
Figu e 6. Compa ison o wo gene a ed igu e-g ound maps wi h
a mid buil densi y ained wi h spa se (le ) and dense ( igh )
s ee ne wo k da a wi h a spa ial esolu ion o 1:8000.
Doing he same compa ison bu o mid densi y igu e-g ound
maps, he isual di e ence be ween he maps gene a ed by he
model ha was ained on dense s ee ne wo k da a and he one
ha was ained on spa se s ee ne wo k da a is less no iceable
as isualized in Figu e 6.
Figu e 7 example 2) demons a es a gene a ed igu e-g ound
map whe e e en indi idual pa hways o house en ances a e
conside ed by he model and a e clea ly dis inguishable. Fu -
he mo e, example 1) o he same igu e depic s esul s whe e
ISPRS Annals o he Pho og amme y, Remo e Sensing and Spa ial In o ma ion Sciences, Volume X-4/W5-2024
19 h 3D GeoIn o Con e ence 2024, 1–3 July 2024, Vigo, Spain
This con ibu ion has been pee - e iewed. The double-blind pee - e iew was conduc ed on he basis o he ull pape .
h ps://doi.o g/10.5194/isp s-annals-X-4-W5-2024-33-2024 | © Au ho (s) 2024. CC BY 4.0 License.
37

eal
gene a ion
1)
2)
3)
4)
inpu
Figu e 7. Resul s om ou di e en high densi y es images
wi h a spa ial esolu ion o 1:8000 and a dense s ee ne wo k.
The inpu image, he gene a ed igu e-g ound map and he eal
igu e-g ound map a e displayed.
la ge indus ial acili ies a e accu a ely gene a ed due o he a -
angemen , ype and wid h o he s ee s. Fo he mid buil dens-
i y, we can u he s a e ha he gene a ed scenes a e cha ac e -
ized by o ganic and easonable building oo p in shapes as can
be depic ed in Figu e 8.
Analyzing he igu e-g ound maps o high and mid densi ies a
a spa ial esolu ion o 1:4000 issued om a ained model on a
dense s ee ne wo k e eals a la ge disc epancy o hose gene -
a ed a 1:8000 as indica ed by Figu e 9. A a spa ial esolu ion
o 1:4000, he model seems o be incapable o p oduce accu a e
building geome ies bo h o high and mid densi y scena ios.
To conclude ou expe imen al esul s, we calcula ed he FID
sco es. As indica ed by Table 2, igu e-g ound maps gene a ed
wi h a model ha is ained on a spa ial esolu ion o 1:4000
esul s in wo se pe o mance compa ed o he ou pu o a model
ha is ained on a spa ial esolu ion o 1:8000, suppo ing he
indings o he isual analysis. The FID sco es o he igu e-
g ound maps esul ing om high and mid densi ies o a spa ial
esolu ion o 1:8000 show ha aining he model wi h spa se
s ee ne wo k da a hampe s he pe o mance o he model sig-
ni ican ly.
densi y s ee ne wo k spa ial esolu ion FID
high dense 1:4000 153.14
mid dense 1:4000 140.20
high dense 1:8000 71.25
mid dense 1:8000 75.52
high spa se 1:8000 116.46
mid spa se 1:8000 84.77
Table 2. FID sco es o he conduc ed expe imen s.
eal
gene a ion
1)
2)
3)
4)
inpu
Figu e 8. Resul s om ou di e en mid densi y es images
wi h a spa ial esolu ion o 1:8000 and a dense s ee ne wo k.
The inpu image, he gene a ed igu e-g ound map and he eal
igu e-g ound map a e displayed.
1)
2)
3)
eal
gene a ion
eal
gene a ion
high buil densi y
mid buil densi y
4)
Figu e 9. Resul s om ou di e en high and mid densi y es
images wi h a spa ial esolu ion o 1:4000 and dense s ee
ne wo k da a.
5. Conclusion and Ou look
This pape in oduced an app oach owa ds a deep au oma ic
gene a ion o igu e-g ound maps commonly used in u ban plan-
ISPRS Annals o he Pho og amme y, Remo e Sensing and Spa ial In o ma ion Sciences, Volume X-4/W5-2024
19 h 3D GeoIn o Con e ence 2024, 1–3 July 2024, Vigo, Spain
This con ibu ion has been pee - e iewed. The double-blind pee - e iew was conduc ed on he basis o he ull pape .
h ps://doi.o g/10.5194/isp s-annals-X-4-W5-2024-33-2024 | © Au ho (s) 2024. CC BY 4.0 License.
38
ning. We show ha ou app oach is able o au oma ically gen-
e a e igu e-g ound maps ha can be adjus ed o high and mid
buil densi ies. The gene a ed maps consis o a i icially gen-
e a ed building oo p in s and a p ede ined s ee ne wo k. This
opens up new oppo uni ies o suppo he as ye edious si e
plan d a ing and has he po en ial o accele a e his ime con-
suming ask which di e en s ages o u ban de elopmen ely
on. The ained model u ns ou o cap u e he buil densi y
which has been e lec ed by he esul ing maps. In his con ex ,
a aining and da a p ocessing on a dense and mo e de ailed
s ee ne wo k wi h a mid spa ial esolu ion impac ed he es-
ul ing a angemen and sha pness o building oo p in s. As o
now he models backbone is an unmodi ied condi ional GAN.
A goal o u u e esea ch is he inco po a ion o u he ypical
si e plan equi emen s, e.g. he physical dimensions o a plo
o cons ain s on he a io be ween buil and unbuil a ea. Fo a
mo e accu a e ep esen a ion o he popula ion and buil dens-
i ies, he inco po a ion o he desi ed numbe o loo s will be
add essed as well. Addi ionally, de ailed expe imen al in es ig-
a ion including pa ame e uning and op imiza ion will be sub-
jec o u u e wo k.
Acknowledgemen s
The au ho s would like o exp ess hei g a i ude o he open
code om Wu and Biljecki (2023).
Re e ences
BMVI, 2018. Regionals a is ische Raum ypologie (RegioS a )
des BMVI ¨
u die Mobili ¨
a s- und Ve keh s o schung. Fede al
Minis y o T anspo and Digi al In as uc u e. bmd .bund.de
(15 Feb ua y 2024).
Chen, W., Wu, A. N., Biljecki, F., 2021. Classi ica ion o u ban
mo phology wi h deep lea ning: Applica ion on u ban i ali y.
Compu e s, En i onmen and U ban Sys ems, 90, 101706.
Eu opean Commission, 2016. Coun y summa y o Ge -
many. Tes ing he deg ee o u banisa ion a he global le el.
ghsl.j c.ec.eu opa.eu (15 Feb ua y 2024).
Gong, Q., Li, J., Liu, T., Wang, N., 2020. Gene a ing u ban ab-
ic in he o hogonal o non-o hogonal u ban landscape. En-
i onmen and Planning B: U ban Analy ics and Ci y Science,
47(1), 25–44.
Ha -Peled, S., Raichel, B., 2014. The F ´
eche dis ance e isi ed
and ex ended. ACM T ansac ions on Algo i hms (TALG), 10(1),
1–22.
Heusel, M., Ramsaue , H., Un e hine , T., Nessle , B., Ho-
ch ei e , S., 2017. Gans ained by a wo ime-scale upda e ule
con e ge o a local nash equilib ium. P oceedings o he 31s In-
e na ional Con e ence on Neu al In o ma ion P ocessing Sys-
ems, NIPS’17, 6629–6640.
Hijazi, I., Li, X., Koenig, R., Schmi , G., El Meouche, R., L ,
Z., Abune’meh, M., 2017. Measu ing he homogenei y o u ban
ab ic using 2D geome y da a. En i onmen and Planning B:
U ban Analy ics and Ci y Science, 44(6), 1097–1121.
Isola, P., Zhu, J.-Y., Zhou, T., E os, A. A., 2017. Image- o-
image ansla ion wi h condi ional ad e sa ial ne wo ks. 2017
IEEE Con e ence on Compu e Vision and Pa e n Recogni ion
(CVPR), 5967–5976.
Jiang, F., Ma, J., Webs e , C. J., Chia adia, A. J., Zhou, Y.,
Zhao, Z., Zhang, X., 2023a. Gene a i e u ban design: A sys-
ema ic e iew on p oblem o mula ion, design gene a ion, and
decision-making. P og ess in Planning, 100795.
Jiang, F., Ma, J., Webs e , C. J., Li, X., Gan, V. J., 2023b. Build-
ing layou gene a ion using si e-embedded GAN model. Au o-
ma ion in Cons uc ion, 151, 104888.
Jiang, F., Ma, J., Webs e , C. J., Wang, W., Cheng, J. C., 2024.
Au oma ed si e planning using CAIN-GAN model. Au oma ion
in Cons uc ion, 159, 105286.
S a is a, 2024. S a is a. de.s a is a.com (15 Feb ua y 2024).
S iny, G., 1980. In oduc ion o shape and shape g amma s. En-
i onmen and planning B: planning and design, 7(3), 343–351.
Wang, J., Huang, W., Biljecki, F., 2024. Lea ning isual ea-
u es om igu e-g ound maps o u ban mo phology disco e y.
Compu e s, En i onmen and U ban Sys ems, 109, 102076.
Wu, A. N., Biljecki, F., 2022. GANmappe : geog aphical da a
ansla ion. In e na ional Jou nal o Geog aphical In o ma ion
Science, 36(7), 1394-1422.
Wu, A. N., Biljecki, F., 2023. Ins an CITY: Syn hesising mo -
phologically accu a e geospa ial da a o u ban o m analysis,
ans e , and quali y con ol. ISPRS Jou nal o Pho og amme y
and Remo e Sensing, 195, 90–104.
ISPRS Annals o he Pho og amme y, Remo e Sensing and Spa ial In o ma ion Sciences, Volume X-4/W5-2024
19 h 3D GeoIn o Con e ence 2024, 1–3 July 2024, Vigo, Spain
This con ibu ion has been pee - e iewed. The double-blind pee - e iew was conduc ed on he basis o he ull pape .
h ps://doi.o g/10.5194/isp s-annals-X-4-W5-2024-33-2024 | © Au ho (s) 2024. CC BY 4.0 License.
39