Machine Learning for the Built Heritage Archaeological Study

10
Machine Lea ning o he Buil He i age A chaeological S udy
AMAIA MESANZA-MORAZA, ISMAEL GARCÍA-GÓMEZ, and AGUSTÍN AZKARATE,
Buil He i age Resea ch G oup, Uni e si y o he Basque Coun y UPV/EHU, Vi o ia-Gas eiz, Spain
The p esence o a i icial in elligence in ou li es is inc easing and being applied o ields such as medicine, enginee ing,
elecommunica ions, emo e sensing and 3D isualiza ion. Ne e heless, i has ne e been used o he s a ig aphic s udy o
his o ical buildings. Thus a , a chaeologis s and a chi ec s, he expe s in a chaeology o a chi ec u e, ha e led his esea ch.
The me hod consis ed o isually—and, consequen ly, subjec i ely—iden i ying ce ain e idence ega ding he ele a ions o
such buildings ha could be a consequence o he passage o ime. In his a icle, we would like o p esen he esul s om one
o he esea ch p ojec s pu sued by ou g oup, in which we au oma ed he s a ig aphic s udy o some his o ic buildings using
mul i a ia e s a is ic echniques. To his end, we i s measu ed he building using su eying echniques o c ea e a 3D model,
and hen, we b oke down e e y s one in o quali a i e and quan i a i e a iables. To iden i y he s a ig aphic ea u es on
he walls, we applied machine lea ning by conduc ing di e en p edic i e and desc ip i e analyses. The p edic i e analyses
we e used o ule ou any blocks o s one wi h di e en cha ac e is ics, such as ough s ones, join ashla s, and oussoi s
o a ches; hese a e i egula i ies ha p obably show building p ocesses and whose iden i ica ion is c ucial in asce aining
he s uc u al e olu ion o he building. In supe ised lea ning, we expe imen ed wi h decision ees and andom o es —
and al hough he esul s we e good in all cases, we ul ima ely op ed o implemen he p edic i e model ob ained using he
las one. While iden i ying he e idence on he walls, i was also e y impo an o iden i y di e en con inui y solu ions
o in e aces p esen on hem, because al hough hese a e elemen s wi hou ma e iali y, hey a e o g ea alue in e ms o
imescale, because hey delimi di e en s a a and allow us o deduce he ela ionship be ween hem.
CCS Concep s: • Applied compu ing →A s and humani ies;A chi ec u e (buildings);
Addi ional Key Wo ds and Ph ases: A chaeology o A chi ec u e, building a chaeology, buil he i age, machine lea ning,
mul i a ia e analysis, s a ig aphic analysis, da a mining
ACM Re e ence o ma :
Amaia Mesanza-Mo aza, Ismael Ga cía-Gómez, and Agus ín Azka a e. 2020. Machine Lea ning o he Buil He i age A -
chaeological S udy. J. Compu . Cul . He i . 14, 1, A icle 10 (Decembe 2020), 21 pages.
h ps://doi.o g/10.1145/3422993
This con ibu ion is pa o he esea ch p ojec “The A cheology o A chi ec u e: new me hodological challenges and hei applica ion
in Eu ope and La in Ame ica” (HAR2015-64439-P), inanced by he Spanish Minis y o Economy and Compe i i eness and he p ojec
“A chaeology o A chi ec u e in he old and he new wo ld: om he s a ig aphy o he buildings o he s a ig aphy o he u ban ab ic”
(PID2019-109464GB-I00), inanced by he Spanish Minis y o Science and Inno a ion.
Au ho ’s add esses: A. Mesanza-Mo aza (co esponding au ho ), I. Ga cía-Gómez, and A. Azka a e, Buil He i age Resea ch G oup, Uni e si y
o he Basque Coun y UPV/EHU, Micaela Po illa Resea ch Cen e , Jus o Velez de Elo iaga S ee , 1s . 01006 Vi o ia-Gas eiz, Spain; emails:
{amaia.mesanza, ismael.ga cia, agus in.azca a e}@ehu.eus.
$
This wo k is licensed unde a C ea i e Commons A ibu ion-NonComme cial-NoDe i s In e na ional 4.0 License.
© 2020 Copy igh held by he owne /au ho (s).
1556-4673/2020/12-ART10
h ps://doi.o g/10.1145/3422993
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.
10:2 • A. Mesanza-Mo aza e al.
1 INTRODUCTION
In a chaeology, he main da ing me hod is s a ig aphic, while he es is s uc u ed a ound he ollowing: doc-
umen a y, ypological, o mal, s uc u al, o a chaeome ic sou ces [1–3]. This s a ig aphic analysis, which is
diach onic in na u e, equi es he iden i ica ion o all s a ig aphic uni s abo e g ound a chaeology—deposi s and
su aces [4]—as well as exis ing links be ween hem o be able o in e p e he his o ical e olu ion o a build-
ing. Toge he wi h s a ig aphic analysis, he e is a second me hod ha is ypological in na u e. In his case,
i is a synch onic analysis whe eby he o mal cha ac e is ics o a se ies o elemen s o ma e ial emains can
be compa ed, making he o ma ion o clus e s possible, i.e., g oups in which i s indi iduals sha e e y simila
cha ac e is ics, bu ha a e e y di e en om hose o he es o he componen s o o he g oups.
A chaeological ypologies in a chi ec u e a e c ea ed wi h he objec i e o da ing buildings. Thus, while he e
is a se ies o ch onological ins umen s bo owed om o he disciplines (Ca bon 14, he moluminescence, elec-
on spin esonance, e c.), he e a e o he ypes o indica o s, de eloped in he ield o a chaeology o a chi ec-
u e, which a e applied on a local basis, i.e., ha a e use ul in mo e o less limi ed geog aphic con ex s. This
akes he o m o he ch onology o s uc u al elemen s and mensioch onology [5]. This second echnique—
mensioch onology—a emp s o connec he dimensions o b icks o di e en his o ical pe iods, and o his end,
analyses hem s a is ically [6,7]. I undamen ally s udies he b icks a he han any o he ype o cons uc-
ion ma e ial due o i s being a ques ion o mass-p oduced p oduc s, and in which dimensional a iabili y is an
impo an ea u e [8].
The p ocedu e in ol es collec ing a su icien ly la ge sample o b icks, measu ing hem, analysing hem s a-
is ically, and compa ing hem wi h a se ies o known cases ha ha e been p e iously da ed wi h ce ain y based
on o he absolu e ch onological indica o s, such as w i en sou ces, insc ip ions, ce amics, and so on. In all cases,
his is undamen al o ensu e a good de ini ion o he a ea o applica ion, because i is a me hod ha wo ks e y
well in local a eas [9].
Al hough hese s udies ha e mos ly been ca ied ou wi h b icks only, o he ypes o es s ha e been ca ied
ou along he same lines, by changing he cons uc ion ma e ial, o example, in adobe [10,11], limes one [12],
o li hic ma e ials [13–21]. Howe e , he esul ob ained has no p o ed pa icula ly sa is ac o y. Occasionally,
he size o he s ones wi h which a building is buil may be limi ed by he a chi ec ’s inspi a ion, bu in mos
cases, demand will be ma ked by he maximum size o he block o ock a ailable, which is he main impedimen
acco ding o specialis s in he subjec in e ms o he de elopmen o mensioch onology p ojec s in ano he
ma e ial o he han b ick [7,22].
Ne e heless, e en hough he ew p ojec s ca ied ou so a wi h s one ma e ial ha e no gene a ed op imism
in e ms o hei alue, he p ojec ca ied ou by Tiziano Mannoni in 1974 [13] will se e as a good s a ing poin .
In i , ollowing on om a s a is ical s udy on he dimensions o he ashla blocks om he “Cas elo” de Genoa,
Mannoni con i ms he exis ence o di e en g oups o pieces ha , in he wo ds o he esea che , “could cons i u e
ch onological e idence.”
Following in he I alian esea che ’s oo s eps, we applied he s a is ic o he his o ical-a chaeological s udy o
his o ical buildings, and he esul s ob ained hus a [23] ha e con i med wha Mannoni said, enabling a possible
ch onological sequence o hose buildings o be o mula ed. This ein o ced he impo ance o he ypological
s a egy o he a chaeological s udy o s one ele a ions. Bu o such econs uc ion, i was necessa y o collec
mo e e idence han he g oups ha had been o med. The e a e pieces o e idence ha a e mo e ypical o he
s a ig aphic s a egy o he analysis o walls ha may include, among o he hings, in e aces o join ashla s.
And s a ig aphy and ypology do no ha e o be sepa a ed, bu a he , can and should complemen each o he
[1], jus as hey did in he cou se o he es o a ion p ojec o he San a Ma ía Ca hed al o Vi o ia-Gas eiz (Spain)
in 1996 [24,25], in he ea ly medie al chu ches o Ala a (Spain) in 2007 [26], and in he Sal Valley o Salinas de
Añana (Ala a, Spain) one yea la e [27,28]. To he a o emen ioned mul i-laye ed na u e o his o ical buildings
should now be added hei plu i ipologized na u e [29].
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.
Machine Lea ning o he Buil He i age A chaeological S udy • 10:3
Howe e , ou p oposal is o ackle he a chaeological s udy o ele a ions by applying only he ypological
me hodology, albei no in he way i has been done up o now, bu a he , in an au oma ed way. O he wo
me hodologies desc ibed abo e, his is he only one ha can be au oma ed. To his end, a mul i a ia e analysis
will be used ha ocuses on applying di e en echniques, such as P incipal Componen s Analysis (PCA), e-
g ession analyses, disc iminan analysis, clus e analysis, among o he s [30]. This me hodology b eaks wi h he
adi ional way o handling his ype o p ojec , and i is necessa y o implemen quan i a i e and mul i a ia e
me hods acco dingly [31], mo ing om he habi ual way o p oceeding—quali a i e, subjec i e, and in ui i e— o
a mo e quan i a i e and objec i e one [32]. Toge he wi h his mo e quan i a i e a chaeology and he inco po-
a ion o s a is ics in o a chaeological wo k, machine lea ning and expe sys ems— wo b anches o a i icial
in elligence—will also be included.
The e a e ha dly any p ojec s whe e mul i a ia e analysis has been applied o he his o ical-a chaeological
s udy o buildings—pe haps he p ojec ca ied ou in 2013 by he A gen ine a chi ec G. Rolón [31,33], al hough
as i is a comple ely di e en a ea o applica ion, i is di icul o ind wha o compa e he ob ained esul o.
While i is ue ha he e a e inc easingly mo e p ojec s ha a e esea ching in o how o au oma ically iden i y
o classi y pa icula a chi ec u al elemen s o s yles [34–38], i is no easy o ind one ha ocuses on he
a chaeological s udy o his o ical buildings.
This is p ecisely he main inno a ion o his wo k ha we p esen he e, as i is he i s ime a his o ical building
has been s udied en i ely and independen ly by a compu e ollowing a pe iod o supe ised lea ning, based on
he e idence de ec ed in he walls. The sys em has lea ned o iden i y pa icula signs ha in ol e s uc u al
changes in a building, which enables a hypo hesis o be o mula ed abou i s his o ical e olu ion h oughou
his o y.
P edic i e and desc ip i e echniques we e applied as pa o he lea ning p ocess ha he compu e is pu
h ough. Wi h he o me , he goal was o y and iden i y and cha ac e ise pa e ns and/o beha iou om he
da a ha would enable key e idence o be pinpoin ed abou he ele a ions o in e p e he s uc u al e olu ion
o he building. A he same ime, desc ip i e echniques would p o ide he solu ions necessa y o p oduce
clus e s, i.e., homogenous g oups o ashla s ha , oge he wi h he p e iously ca ego ised e idence, would enable
a hypo hesis o be o mula ed abou he s uc u al sequence o he building being s udied.
Th oughou his a icle, we will explain how o apply p edic i e echniques o iden i y e idence abou he
ele a ions o his o ical buildings, using di e en cons uc ions ha ha e been app op ia ely selec ed in each o
he phases. The second pa o he esea ch, he clus e o desc ip i e analysis, is included in he a icle published
by he au ho s o Re e ence [23].
2 MATERIALS AND METHODS
2.1 Backg ound
2.1.1 Selec ed Buildings. The p oposed esea ch wo k equi ed ha ing buildings om which o ob ain alid
in o ma ion o au onomous lea ning abou he sys em and o he di e en buildings whe e his knowledge could
be applied. Bo h se s had o include a se ies o common ea u es, such as he li hology o s ones o he ype o
bonding, al hough i was no necessa y o hem o be o simila a chi ec u al s yles. This issue was impo an
o he ep oducibili y o he expe imen .
The li hology o he walls will be he common link be ween all he buildings selec ed o ake pa in he p ojec ,
as in all o hem limes one o Palaeocene limes one (Lumaquelas de Aja e), wo ked mos ly in he shape o ashla ,
appea as he p e e ed ock o hei cons uc ion—an aspec ha is qui e common in o he chu ches in Ala a
(Spain) [39].
We will he e o e ca y ou ou esea ch on h ee his o ic buildings: (1) he He mi age o he Pu ísima
Concepción, buil a he end o he 12 h cen u y and loca ed in he own o San Vicen ejo 14 km sou h o
Vi o ia-Gas eiz (Spain), in Bu gos p o ince. Fo many esea che s i is an “excep ional” emple wi hin he la e
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.
10:4 • A. Mesanza-Mo aza e al.
Fig. 1. Loca ion o he buildings selec ed and he Aja e qua y.
peninsula Romanesque s yle due mainly o i s ex ao dina y capi als [40]. The chu ch consis s o a single na e
wi h wo sec ions and ci cula chancel, all buil wi h e y well-cu ashla s made om Palaeocene limes one
om he nea by Aja e qua y. (2) The Basilica o San P udencio de A men ia, one o he bes examples o he
Romanesque s yle in he Basque Coun y. I s geog aphical loca ion, jus 3 km sou hwes o Vi o ia-Gas eiz, a a
c oss oads o ancien oads and i s link o he pa on sain o he p o ince, as well as i s signi icance as he mos
impo an spi i ual cen e o Ala a in he High Middle Ages, make i especially in e es ing om a his o ical poin
o iew [41]. And (3) San a Ma ía Ca hed al, loca ed in he mos ele a ed pa o he old ci y o Vi o ia-Gas eiz,
wi hin i s o iginal walled enclosu e. I is a La in c oss emple, wi h h ee na es, a chancel, an ambula o y, and
side chapels, buil using echniques common o di e en imes du ing he Go hic pe iod [24] (Figu e 1).
2.1.2 The B eaking Down o Buildings in o Va iables Based on Thei Geome ic Documen a ion. In he cou se
o eading ele a ions, i is essen ial o ha e geome ic documen a ion whe e he e idence disco e ed and he
esul s o he esea ch can be cap u ed [6,42,43]. The eco d ha is made mus be a comple e ep esen a ion
o he building and i s pa s, his being all he mo e use ul and acili a ing a chaeological wo k inso a as i will
be o g ea e quali y and p ecision [44]. L. Caballe o [45] said ha “ he quali y o he documen a ion does no
necessa ily ha e o a ec he deg ee o alidi y o he inal esul o he a chaeological analysis,” bu he mo e
de ailed he documen a ion, he mo e pe ec he esul .
I should be no ed ha , in he s udy o he his o ical-s uc u al e olu ion o a building, his is conside ed
a single elemen , so he g aphic model ob ained mus comply wi h his de e mining ac o . In o he wo ds,
ega dless o whe he he inal p oduc is a collec ion o d awings, he eco d should be pu oge he in such a
way ha he di e en pa s o he building may be linked [46].
When he 3D model is eady, ca ied ou using any o he mos widely used echniques in buil he i age (lase
scanne , au oma ic pho og amme y, s e eopho og amme y, ec i ied pho og aphy o opog aphy), i is usual o
de elop 2D d awings ha may se e as a ool whe e he esul s o a chaeological and/o a chi ec u al esea ch
can be cap u ed mo e isually. Howe e , o ou pu pose o ying o p oduce au oma ic lea ning, o ying
o p og amme a compu e so i may be able o simula e he way in which a chaeologis s o a chi ec u e ind
solu ions o p oblems, i is necessa y o go one s ep u he by making all his geome ic in o ma ion an ac i e
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.
Machine Lea ning o he Buil He i age A chaeological S udy • 10:5
pa o he esea ch p ojec i sel [47]. To do his, i will be necessa y o ind a way o ex ac ing all his da a and
ans o ming i , di iding i in o simple o ms ha help us o unde s and i compu a ionally.
A e much e lec ion as o whe he o adop one elemen o ano he as a minimum uni o s udy, as his
choice condi ions he es o he p ocess and he esul ob ained om i , i was decided o de e mine ha he
minimum uni o s udy would be each block o s one ha makes up a building. This is because he la e would
be able o p o ide su icien in o ma ion when i came o “au oma ing” he p ocess o he s a ig aphic eading
o ele a ions, being ela i ely simple o b eak down in o a iables.
These a iables, in which each block o s one is going o be b oken down, mus be able o de ine hem unam-
biguously and unequi ocally. They a e a undamen al pa o he esea ch, as hey a e p esen in all phases o he
p ocess—obse a ion and measu emen , classi ica ion, analysis, p edic ion [48]. They need o be se ou clea ly,
by being de ined concep ually a i s , and hen mo ing on o he ope a ional phase whe e he p ocedu es a e
es ablished ha allow measu emen s o be ob ained.
In he same way as when i comes o b inging oge he he geome ic documen a ion o a building, i is
necessa y o conside i in i s en i e y, because i is also in e es ing o s udy he a iabili y in all i s s ones o
a leas he g ea majo i y o hem. I his is done pa ially, i will be di icul o iden i y s uc u al momen s o
ch onologies on he ele a ions.
Ini ially, o a ypological s udy, nei he he ype no he numbe o a iables o be used is es ablished, as hese
depend on he ci cums ances and needs a ached o each con ex o applica ion [49]. Howe e , i is ad isable o
ake all op ions in o conside a ion in his i s phase o he wo k as un il he whole p ocess is inished, i is
unknown which one o ones will be he g ea es de e mining ac o s [26]. I will be he same nume ical me hods
ha indica e hose a iables ha ca y mos weigh [32]. Gene ally speaking, in he s udy o his o ical buildings,
he a iables a e usually o a echnical-s uc u al ype (dimensions, ypes o ma e ials, ypes o bonding, ypes o
ins umen s used in sculp u es, s onemason ma ks, mo a , e c.), o mal ype ( ypes o capi als, a ches, i o iums,
windows, b acke s, e c.) o spa ial ype (loca ion, le el, g adien , e c.) [50].
When analysing he a o emen ioned a iables, p ac ically all o hem u ned ou o be quali a i e excep o
hose ha e idenced a geome ic o spa ial componen . The e o e, i was necessa y o add o hose quali a i e
a iables o he s o a quan i a i e na u e ha could be ob ained au oma ically. Finally, in he case o he s one
blocks, 33 a ibu es we e ex ac ed in o al, o which only 9 we e quali a i e, and he es , quan i a i e. Thus,
he e will be a i s g oup whe e 6 con ex ualisa ion a iables a e loca ed (building, ype o building, loca ion,
o ien a ion, exposi ion, li hology), he second o which would s udy 3 dicho omous a iables (a ch, join ashla s,
and ype o s one), and i will be hese ha will be used o classi y he s ones wi hin he supe ised lea ning phase.
Fo he hi d block, he e will be he quan i a i e a iables ha desc ibe bo h he geome ic cha ac e is ics o
he s one blocks (heigh , leng h, a ea, pe ime e , cen oid (X and Y coo dina es), diagonal 1, diagonal 2, heigh o
leng h a io), hose o hei co e ings1(a ea, pe ime e , and cen oid coo dina es) and exis ing ela ions be ween
he s one blocks and hei co e ings (a ea di e ence (% and absolu e), pe ime e di e ence (% and absolu e),
cen oid di e ence (%, absolu e and coo dina es) and diagonal di e ence (% and absolu e)).
We au oma ed he p ocess o gene a ing he co e ings and c ea ing and calcula ing he di ec and indi ec
a iables as much as possible o a oid subjec i i y. We used he “Model Builde ” A cGis applica ion o a oid
da a collec ion e o s. This was a undamen al pa o he esea ch p ocess. In addi ion, he models c ea ed we e
expo able di ec ly o Py hon in case modi ica ion was necessa y. Each s one had a unique iden i ica ion code.
2.2 Supe ised Classi ica ion
Wi h his p oposal app oach—documen ing he building in i s en i e y and di iding i in o mino elemen s de-
ined by mul iple a iables— he in o ma ion exceeds he capaci y o manual managemen and analysis, as has
1In his con ex , one synonym o “co e ing” could be “bounding box” o pe haps be e “minimum bounding ec angle.” This is an exp ession
o he maximum ex en s o a wo-dimensional objec ( ha is, he s ones o he buildings) wi hin i s coo dina e sys ems (min(x), max(x), min(y),
and max(y)).
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.

10:6 • A. Mesanza-Mo aza e al.
Fig. 2. Flowcha o he p ojec .
been he case un il now. Faced wi h such a olume o da a, he g ow h in ecen yea s o big da a, da a mining,
machine lea ning, and a i icial in elligence and i s p esence in almos all scien i ic ields makes he possibili y
o applying some o hese concep s o he a chaeology o a chi ec u e a e y in e es ing op ion (Figu e 2).
One o he pu poses o using p edic i e echniques is o ind pa e ns o beha iou wi hin a b oad se o da a.
In ou esea ch, hese me hodologies will be used wi h wo e y speci ic objec i es in mind:
• he classi ica ion o s one blocks in o wo b oad ca ego ies: ashla s and ough s ones.
• he analysis o he con inui y o cou sed ashla s, which in ol es iden i ying s ones wi h ce ain i egu-
la i ies, such as is he case wi h join ashla s.
In ac , wha is p oposed is he posi ioning o ou lie s wi hin he con igu a ion o he ele a ions, i.e., s ones ha ,
o a ious easons, do no ha e he egula shape o an ashla and ha may be an indica ion o possible s uc u al
ac ion. No mally, he cons uc ion o g ea his o ical buildings ends o be p olonged se e al yea s o decades,
and his makes he wall he e ogeneous. O he simple passage o ime makes i necessa y o edo o modi y he
ini ial design. All hese ac ions a e e lec ed in he ele a ions in he ype o e idence and i is p ecisely his ha
he sys em should au onomously de ec [51].
2.2.1 Classi ica ion o Ashla and Rough S ones. When building in s one, he e a e h ee basic ypes ha di e
depending on he deg ee o he cu and size: ough s ones, smalle ashla s, and ashla blocks. In o he wo ds,
ashla s a e hose blocks o s one in he shape o a pa allelepiped ha a e well cu and whose co ne s a e pe ec ly
squa ed, in con as o ough s ones ha a e comple ely i egula in shape. The smalle ashla s come somewhe e
in be ween.
The esea ch will be based on indi idualisa ion o he ashla s in con as o he ough s ones, because, as can be
seen, hey ha e e y une en ea u es ha a e ela i ely easy o de ec deduc i ely. The smalle ashla compa ed
o he ashla does no ha e a clea nuance ha dis inguishes i . In he li e a u e consul ed, he e is ha dly any
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.
Machine Lea ning o he Buil He i age A chaeological S udy • 10:7
Fig. 3. No h ele a ion o he Basilica o San P udencio de A men ia. The le pa is made using ashla s, while he igh
pa co esponds o ough ashla s.
wo k in his a ea, excep ha ca ied ou by Oses [52] in which an ini ial a emp was made o classi y he s one
blocks o walls based on a se ies o ou lines ex ac ed om a se ies o images. In he a o emen ioned p ojec , he
blocks we e classi ied in o h ee g oups, oughly equi alen o he h ee de ined by us: mason y, small ashla , and
ashla . Ve y good classi ica ions we e ob ained o bo h ex emes bu no such good ones o he in e media e
g oup, which does no ha e he cha ac e is ics o ei he one o he o he . This classi ica ion o he s one blocks,
appa en ly simple o unde ake isually, is no so easy o au oma e. Howe e , seeing ha he esea che s can
indeed ca alogue hem manually in an agile and accu a e way, we op ed o supe ised lea ning as being he
mos app op ia e me hodology o he pu pose o ying o build p edic i e models ha could classi y hem
au oma ically.
To be able o look o pa e ns o beha iou , i is necessa y o ha e a sample in which he da a has been
p e iously classi ied manually and is gua an eed—i will be his da a ha “ eaches” he compu e . This i s
pa o he p ocess is undamen al, as he esul s ob ained will be closely condi ioned by he esea che ’s p io
knowledge o his.
In his esea ch, o he lea ning and/o es phases, he s ones om he walls o he Basilica o San P udencio
de A men ia will be used, whe eby bo h elemen s—ashla s and ough s ones—a e dis inguished om each o he
wi hou any doub (Figu e 3). In his way, he da abase will con ain nea ly 9,000 s ones, all o hem o he same
li hology, limes one and om he same qua y in Aja e, and sp ead ou ac oss he ele a ions o he no h and
sou h ex e io walls o he a o emen ioned building.
The unde lying model in he disc iminan analysis equi es p io e i ica ion o he supposi ion o no mali y,
homoscedas ici y, and linea i y. These equi emen s a e no necessa y when logis ic eg ession o decision ees
a e applied, bu , e en so, s a is ical analysis o he da a can p o ide use ul in o ma ion ha may enable i s
s uc u e o be unde s ood, as well as he esul s ob ained.
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.
10:8 • A. Mesanza-Mo aza e al.
One o he i s supposi ions o be p o en will be mul i a ia e no mali y. In he absence o a s a is ical es
ha would allow such p oo , his e i ica ion will be ca ied ou indi idually in he case o all a iables.2In
his case, a e he Kolmogo o -Smi no and Ande son-Da ling es s we e ca ied ou o he g oups o ough
s ones and ashla s, none o he a iables used p o ed o be in acco dance wi h a no mal dis ibu ion. The lack
o no mali y is due la gely o he exis ence o some excep ionally high alues, undamen ally in he “high” and
“long” a iables ha make up he p ima y a iables, which will ha e o be analysed in case o any anomalous
alues ha would need o be emo ed om he sample. E en so, as he sample is la ge enough, we could u n o
he heo em o he cen al limi o be able o assume his no mali y [53].
Analysing hese wo a iables in de ail, in he case o he ca ego y de ined as ough s ones, bo h a ibu es—
high (μ=0.085m σ=±0.050m N =5,123) and long (μ=0.190m σ=±0.115m N =5,123)—a e biased o he igh ,
i.e., hey ha e posi i e symme y ha comes om an accumula ion o alues in he lowe a ea o he his og am,
which is logical gi en he cha ac e is ics o his ype o s one. The same analysis o he g oup o ashla s e lec s
a somewha di e en si ua ion, as none o he high (μ=0.267m σ=±0.080m N =3,557) o long (μ=0.421m σ=
±0.177m N =3,557) a iables appea s o be biased.
In e ms o homoscedas ici y, he Le ene es was applied, he esul o which allows us once again o ejec
he null hypo hesis o equali y o a iances o bo h g oups—a condi ion al eady obse ed in he boxplo s and
he co esponding his og ams.
As o mul icollinea i y, his occu s when one o he independen a iables can be explained as a linea combi-
na ion o one o mo e a iables p oducing a edundancy o in o ma ion. I his phenomenon occu s, i is no mal
ha he p edic i e models subsequen ly ob ained a e no e y explana o y, gi ing ise o majo e o s. I is ec-
ommended ha hose a iables whose Pea son eg ession index is equal o o g ea e han 0.80 [54]be emo ed
om he sample. Howe e , special ca e mus be aken no o emo e such many a iables om he sample, as
he e is a dange o la e c ea ing models ha a e no e y p edic i e. A e analysing he co a iance and co ela-
ion ma ix and ollowing he abo e indica ions, he a ea and pe ime e a iables o all co e ings we e emo ed
om he da abase.
The esul ob ained—nei he no mali y no homoscedas ici y is achie ed—obliges us o cas aside he disc im-
inan analysis as a echnique o classi ica ion o ashla s and ough s ones and o op , he e o e, o ano he ype
o ool.
No all a iables will ha e he same impo ance o in luence on he p edic i e models ha a e ob ained. The e-
o e, be o e s a ing he i s o he analyses, a sepa abili y s udy was ca ied ou (Figu e 4) ha enabled us ini ially
o asce ain which would be he mos signi ican —no only o ca ego ise he ype o s one, bu also o classi y he
join s. To do his, we used he no malized di e ence o mula based on mean alues and s anda d de ia ions [56].
As da a mining expe s poin ou [57–60], he pe cen age o eco ds used in he aining phase needs o be
conside ably highe han ha used o he es ing phase, i s being ecommended a 60%–40%. A e calcula ing
he models wi h di e en pe cen ages, and seeing he esul s ob ained, he p edic i e models we e gene a ed
based on 80% o he da a while he emaining 20% was le o he es ing phase. The elemen s in ol ed a each
momen we e andomly selec ed by he so wa e used.3
In his aining phase, we expe imen ed wi h classi ica ion ees, c oss- alida ion, and andom o es (Table 1).
Howe e , o y o op imise he models, we inally op ed o he andom o es algo i hm whe e, con a y o
wha happens o example wi h classi ica ion ees, as mo e ees a e in oduced in o he calcula ion, he algo-
i hm does no o e -adjus , bu , a he , se s a gene alisa ion e o limi [61]. Di e en models we e c ea ed using
2The assump ion o no mali y can be e i ied in a ious ways: g aphically (e.g., his og ams, box-plo s, Q-Q cha s), using s a is ical es s
(e.g., Kolmogo o -Smi no , Shapi o-Wilks, Ande son-Da ling), o by nume ical me hods (e.g., cen al endency measu emen s, coe icien s
o symme y, ku osis). Fo ou esea ch, we used he Kolmogo o -Smi no and Ande son-Da ling es s [54].
3To gene a e he p edic i e models, bo h in he aining- es phase and in he subsequen alida ion phase, we used he ee so wa e KNIME
Analy ics Pla o m, e sion 2.11.2.
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.
Machine Lea ning o he Buil He i age A chaeological S udy • 10:9
Fig. 4. His og ams wi h he mean and s anda d de ia ion o “high” and “cen oid di e ence” a iables. On he le , he
his og ams a e qui e sepa a e, and his is because he “high” a iable has a high sepa abili y in dis inguishing be ween
ashla s and ough s ones. Howe e , in he second g aph, he his og ams p ac ically coincide, and so his means ha he
“cen oid di e ence” a iable is no a good disc imina ing a iable o di e en ia ing be ween ashla s and ough s ones.
Table 1. Resul s o he T aining Phase o Ca ego isa ion o Ashla s and Mason y
Using Classi ica ion ees, C oss- alida ion, and Random Fo es
O e all eliabili y (%) Cohen’s kappa 1-measu e
Classi ica ion ees 97.335 0.947 0.960
C oss- alida ion 97.552 0.951 0.965
Random o es 99.662 0.993 0.999
10-25-100 ees and he esul s ob ained we e e y simila in all cases. All he es s ca ied ou wi h his algo i hm
p oduced less han 1% o e o s, achie ing an almos pe ec classi ica ion, and eaching he ou -o -bag (OOB)
s a is ic ha indica es he p edic ion e o alues in he es se [62] we e less han 3%.
Howe e , he e is a p oblem wi h o e all eliabili y o accu acy, and his can hide impo an di e ences be-
ween ca ego ies wi h a ying le els o accu acy, which is why i is impo an o analyse no only he diagonal
alues o he con usion ma ix, bu also he ma ginal alues. We analysed he e o o omission-EO ( ela ed o
he p ecision o eliabili y o he p oduce -RP) and he e o o commission-EC ( ela ed in his case o he ecall
o eliabili y o he use -RU). Fo ou p edic i e models, hese p ecision and ecall alues we e 0.99, which in-
dica ed a e y good classi ica ion; alues close o one imply ha he disc iminan a iables used allow o good
di e en ia ion be ween g oups, while hose close o ze o— he ex eme opposi e— e lec he ac ha he classi-
ica ion model is no alid. The Cohen’s kappa coe icien a emp s o delimi he deg ee o adjus men due only
o he accu acy o he classi ica ion, dispensing wi h ha caused by andom ac o s. In his case hose alues
exceed 0.99, implying almos pe ec ag eemen [64].
2.2.2 Join Ashla s and Ve ical In e aces. Once he ashla wo k and mason y had been sepa a ed, he da a-
base was much mo e homogeneous, and all he elemen s we e ashla s. Howe e , he e we e se e al blocks ha
had a sligh abbe , usually a a igh angle in one o he co ne s, which ended up b eaking ha uni o mi y. These
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.
10:16 • A. Mesanza-Mo aza e al.
Table 5. Con usion Ma ix as a Resul o Applying
he P edic i e Models Ob ained Using he Random
Fo es Algo i hm o he Ca ego isa ion o Join
Ashla s wi h Unno malized Da a
P edic ion
Ashla s Join ashla
Reali y Ashla s 3,458 266
Join ashla 6 74
Table 6. Con usion Ma ix as a Resul o Applying
he P edic i e Models Ob ained wi h Random
Fo es Algo i hm o Ca ego isa ion o Join
Ashla s wi h No malized Da a
P edic ion
Ashla s Join ashla
Reali y Ashla s 3,711 13
Join ashla 49 31
The con usion ma ix (Table 5) also p oduced some e y in e es ing conclusions. In his i s block, wha was
deduced is ha he classi ie wo ked excep ionally well in iden i ying he join s ha appea in he San Vicen ejo
da abase; because o he possible 80, i co ec ly iden i ied 74 o hem—a success pe cen age o 92% (RU). Bu
wha mus also be aken in o conside a ion is he ac ha i inco ec ly labelled 266 blocks in his ca ego y,
esul ing in an e o o omission close o 80% (RP), a d ead ul pe cen age. To sum up he join ashla s in he
alida ion sample we e e y accu a ely iden i ied, bu many e o s we e made; many ashla s we e labelled as
being join when ac ually hey we e no .
In he second ial, using no malized da a, an imp o emen in he o e all eliabili y ega ding aining was
obse ed (Table 6). Again, he alues o he kappa index we e a ec ed by he inequali y o he ca ego ies, al-
bei no as much as wi h non-s anda d da a. They all o alues o 0.49, which can be conside ed a mode a e
classi ica ion.
He e he p e iously end changes comple ely. While a li le o e 40% o he join s in he ele a ions o San Vi-
cen ejo (RU) ba ely classi y well, ne e heless no oo many inco ec alues a e inpu ed—only 13—which is well
below he p e ious example o 266. Logically, he s a is ics p o ided a ound he mino i y block we e jus as a -
ec ed as in he p e ious ound o ials. To sum up, he e a e ewe co ec ly iden i ied join s and also ew e o s.
As we ha e al eady men ioned, he p esence o join ashla s in an ele a ion may be an indica ion o some kind
o ac ion ca ied ou on i . As a esul , he appea ance o many inco ec ly classi ied ashla s as such p o ides
e y li le use ul in o ma ion o he esea che , since we would be poin ing ou possible s uc u al ac ions ha
ne e eally exis ed. The e o e, o keep he i s o he op ions would mean, in ou iew, including a lo o “noise”
in he s a ig aphic analysis p ocess.
Choosing he second op ion p o ides us wi h a ce ain secu i y, as we know ha he as majo i y o ashla s
classi ied as being join co espond o his label, and ha he numbe o e o s en e ed in he eading p ocess
is lowe .
Rega ding he loca ion o he e o s ha occu when iden i ying he join ashla s, he p edic ion model wo ks
qui e well wi h he join s ha ac ually exis in he ele a ions and he e o s a e undamen ally in he windowsills.
Thisiswhe ei isnecessa y olowe hes one obeable omake heshape ha hey eallyha e,whichmeans
hese blocks bea a ce ain esemblance o he join s.
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.

Machine Lea ning o he Buil He i age A chaeological S udy • 10:17
3.3 Discussion
J. A. Ba celó said in his publica ion Au oma ic a chaeology. A i icial In elligence in A chaeology [68], “A he
p esen ime, au oma ic a chaeology is mo e o a u u e p omise han a eali y; i is an unexplo ed ield ha will
p o ide ex ao dina y and unp edic able esul s in he coming yea s.”
Such a s a emen —made mo e han 20 yea s ago—is oday close o becoming a eali y, hanks o he echnical
p og ess o he pas ew decades ha is allowing aspec s such as a i icial in elligence, expe sys ems, and
machine lea ning o become inc easingly p esen in many di e se a eas o socie y, making i look inc easing less
like science ic ion.
When—e en be o e he p esen wo k—we we e beginning wi h his line o esea ch, we we e i mly con inced
ha a i icial in elligence was possible in a chaeology, and mo e speci ically in buil he i age [47]. Today, no
only ha e we no gi en up on his idea, bu we a e con inuing o s udy i mo e in dep h. The aim o his wo k
is o i o be a i s s ep owa ds he p omising u u e ha Ba celó p edic ed, al hough he e a e s ill lo s o
hings o do. Ou objec i e o acili a ing au oma ion o some a chaeological p ocesses is no in ended o eplace
ei he he a chaeologis o he a chi ec , bu simply o p o ide hem wi h an auxilia y ool o help hem in he
s a ig aphic analysis o a building—a ask no wi hou i s complexi y.
Th oughou ou esea ch, we ha e con i med how a chaeologis s a e o en excessi ely “imp essionis ic” when
desc ibing he easoning ha leads o hei s a ig aphic conclusions abou a building. This was one o he g ea
obs acles we ine i ably had o o e come, namely, he di icul y o clea ly isualizing he in e ence mechanisms
used by he esea che . This is because any a emp o au oma e a echnique equi es— i s —ha ing a clea “ low
cha ” o he in e nal p ocess o he echnique o be sys ema ised.
Be ha as i may, i we had o highligh a single key aspec as he undamen al d i ing o ce behind his
in e nal p ocess in ol ing he s a ig aphic analysis echnique, ha would be he ypological aspec . Logically,
we canno expec o educe all he complexi y in ol ed in he s a ig aphic eading o ele a ions o a single
ma e , bu we do so because, jus like Ba celó, we a e con inced ha he success o he applica ion o expe
sys ems in a chaeology depends on he implemen a ion o au oma ic ypologies [68]. This will make i easie o
a chaeologis s o inpu ma hema ical-s a is ical desc ip ions o hei disco e ies in o da abases, so i may be he
expe sys em ha decons uc s he objec o s udy s a ig aphically.
Some examples o au oma ic ypologies ha e exis ed o some ime in a chaeology; hey ela e o ce amics, he
li hic indus y, ibulas, and he o ien a ion o ombs, among many o he subjec s o in e es . E en so, we ha dly
ind any hing ela ed o buil he i age. The e a e, o cou se, ad ances made in he ield o mensioch onology—
bo h in b ick and s one— ha p o ed undamen al in planning his esea ch. Howe e , hey do no con empla e
he use o mul i a ia e echniques, bu a he , only basic s a is ical p ocedu es, as he s udy popula ion was
minimal—jus like he a iables ha de ined i , which can be summa ised in he dimensions o he pieces
analysed.
I he ypological key is undamen al o au oma ion o he s a ig aphic eading o ele a ions, i is also because
he ypology is, as i we e, he ac i e p inciple ha allows us o analyse ho oughly, o decons uc he building—
in an o ganised and sys ema ic manne — o ans o m all i s componen s in o magni udes, in o igu es ha may
hen be subjec o ma hema ical-s a is ical ea men .
4 CONCLUSION
E e y building o a minimum age is conside ed by a chaeological science o be a mul i-laye ed si e, because i s
s uc u e is he esul o successi e building ope a ions in ol ing con ibu ion and deg ada ion ha ha e been
supe imposed, lea ing an indelible ma k on i s walls. Reading o ele a ions is he mos common echnical p o-
cedu e used o de e mine his s a i ica ion, which has so a been ca ied ou by isual analysis. This analysis
in oduces a high deg ee o subjec i i y in o he p ocess since, depending on he esea che ’s expe ise o ex-
pe ience, he esul may a y. In his a icle, we in oduce a amewo k ha seeks o use machine lea ning o
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.
10:18 • A. Mesanza-Mo aza e al.
semi-au oma e he p ocess o eading ele a ions. Based on he mos common e idence ha specialis s in he
a chaeology o a chi ec u e we e looking o in he ele a ions o buildings, we ied o make he machine e-
sponsible o his by iden i ying and poin ing ou he di e en ypes o s ones, he exis ence o join ashla s and
in e aces, oussoi s o a ches, and e en clus e s ha simula e cons uc ion phases. In his way, we elimina e
he subjec i i y wi h which he me hod s a ed, and we inco po a e a mo e objec i e p ocedu es ha allow
o he s a ig aphic analysis o he building in an iden ical way o all esea che s. I will be la e when he
specialis s wi h all his e idence hen go on o unde ake hei his o ical in e p e a ion o he cons uc i e e o-
lu ion o he his o ical building. All he buildings ha ha e been used in his esea ch sha ed hei li hology, he
same qua y o o igin, and so as long as mo e buildings wi h simila cha ac e is ics can be inco po a ed (which
is ela i ely easy o do, as i is a widely used li hology in Ala a), he sys em will be able o con inue lea ning
and di e en ia ing be ween building signs wi h g ea e p ecision. This does no mean ha he me hod p oposed
only wo ks wi h Palaeocene limes one, as he me hod i sel can be ex apola ed o any building, albei no he
absolu e alues ob ained om i . In o he wo ds, he mos disc imina o y a iables will emain he same, bu no
he alue hey acqui e.
In his a icle, we ha e shown how o ain he sys em o iden i y and poin ou he mason y and ashla s,
he join ashla s, and he in e aces on he ele a ions, ob aining good esul s in all cases, by s a ing om he
p emise ha his issue had ne e been ackled be o e. I is impo an o see how he passi e ole ha , un il now,
he geome ic documen a ion o he building had acqui ed in a p ojec in ol ing eading ele a ions, changes
adically and becomes an ac i e pa o i , since i p o ides he da a equi ed o ha au onomous lea ning. We
expe imen ed wi h se e al algo i hms, bu , ul ima ely, we op ed o he andom o es , since i p o ided he bes
eliabili y in bo h he aining and alida ion phases.
Rega ding p edic i e models, hese wo ked sligh ly be e in he aining phase han in he alida ion phase.
Howe e , he e is an ob ious explana ion o his: To a la ge ex en , his is due o he imbalance be ween he
ca ego ies de ined in he San Vicen ejo he mi age da abase, because he elemen s o be classi ied ( ough s ones
and join ashla s) always ep esen ed he mino i y g oup wi hin he sample. This imbalance was e lec ed in he
p ecision ob ained when classi ying he di e en ca ego ies. Thus, while o he majo i y g oup— he ashla s—
use eliabili y exceeds 95%, in e ms o disc imina ion o he es o he elemen s his alue a ies be ween 40%
and 90%. In he classi ica ion o ough s ones, he esul ing o e all eliabili y p o ides alues a ound o 95%.
E en so—and his is he impo an hing— he p edic i e model c ea ed co ec ly classi ies 85% o he building’s
ough s ones and 99% o he ashla s.
The case o he join ashla s is somewha di e en , as we had wo p edic i e models—one ha was ob ained
using s anda dised da a and he o he ha was ob ained using non-s anda dised da a. I we only pay a en ion o
he nume ical alues, i seems logical o op o he second model, whe e p ecision, when i came o classi ying
hese ype o blocks, was 92% wi h a 7% e o in he classi ica ion o no mal ashla s, compa ed o he 40% ob ained
in he i s case o he ashla s, albei wi h only 0.4% e o in e ms o hei ca ego isa ion. Howe e , we ha e
chosen o keep he model ha , ap io i, o e s he leas p ecision, gi en he impo ance o hese ypes o pieces in
da ing possible al e a ions ca ied ou on he ele a ion. When in e p e ing he s uc u al e olu ion o a building,
i is p e e able o ha e ew join ashla s— ha a e well classi ied and ha p o ide in o ma ion— a he han many
e oneous ones ha may cause con usion.
4.1 Fu u e Wo k
As we ha e commen ed in he a icle, his is he i s ime ha machine lea ning has been used o s udy a
his o ical building om he a chaeological s andpoin , so logically he e is s ill oom o imp o emen . We will
ocus u u e wo ks a ound wo di e en lines: he a iables and he buildings.
Fo ou esea ch, we gene a ed he di e en p edic i e models om a se o 33 a iables, some o hem quali-
a i e, bu mos ly quan i a i e. One line o imp o emen in ol es he inco po a ion o new a iables ela ed o
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.
Machine Lea ning o he Buil He i age A chaeological S udy • 10:19
adiome ic aspec s, since, i bo h isible and he mal image in o ma ion a e a ailable, he digi al le els o he
images may help in he classi ica ion p ocess. Fo example, i we wo k wi h buildings made o b icks, he colou
may p o ide us wi h in o ma ion abou he amoun o clay hey ha e, he ex ac ion poin o he aw ma e ial,
o e en he i ing empe a u e. The di e en e lec i e and emi ing beha iou o he walls allows us o know
he composi ion o he ma e ials used in hei cons uc ion, as well as he mois u e o he exis ence o ma e ials
“hidden” om human sigh .
Ano he aspec o poin ou and which we ha e al eady ad anced in he conclusions, is he possibili y o es ing
his lea ning on mo e buildings. On one hand, i would be in e es ing o con inue wo king on buildings whose
ma e ial comes om he same qua y in Aja e o asce ain whe he he inc ease in he numbe o indi iduals
analysed imp o es he classi ica ion ob ained so a o o he wise. This would in ol e mainly ying o ob ain
da a o hose mino i y ca ego ies, such as he join ashla s o he oussoi s o he a ches. And, on he o he
hand, he e is he possibili y o applying he me hodology o buildings whose cons uc ion ma e ial is o ally
di e en . In his sense, we ha e gained a li le expe ience in buildings made en i ely o b ick in he ci y o Buenos
Ai es, and we will con inue o de elop his, since he ini ial esul s ob ained a e qui e encou aging.
REFERENCES
[1] Agus ín Azka a e, Luis Caballe o, and Juan An onio Qui ós. 2002. A queología de la a qui ec u a: de inición disciplina y nue as
pe spec i as. A queol. A qui ec . 1 (2002), 7–10.
[2] Ka e Giles. 2014. Buildings a chaeology. In Encyclopedia o Global A chaeology, C. Smi h (Ed.).– Sp inge , New Yo k, 1033–1041.
[3] Agus ín Azka a e. 2020. A chaeology o a chi ec u e. In Handbook o Global A chaeology, Ch. E. O se , P. Funa i, S. Law ence, A.
Za ankin (Eds). Rou ledge, Taylo & F ancis.
[4] Edua d C. Ha is. 2017. Ha is ma ices and he s a ig aphic eco d. In Encyclopedia o Geoa chaeology. Encyclopedia o Ea h Sciences
Se ies, A. S. Gilbe (Ed.). Sp inge , Do d ech New Yo k, London, 404–410.
[5] Juan An onio Qui ós. 1996. Indicado es c onológicos de ámbi o local: C ono ipología y mensioc onología. In A queología de La A qui-
ec u a. el Mé odo A queológico Aplicado al P oceso de Es udio E In e ención en Edi icios His ó icos, L. Caballe o and C. Esc ibano (Eds.).
–Jun a de Cas illa y León, Bu gos, 1–13.
[6] Robe o Pa en i. 1995. His o ia, impo ancia y aplicaciones del mé odo de lec u a de pa amen os (His o y, impo ance and applica ions
o he s a ig aphic me hod). In o mes de la Cons ucción 46 (1995), 19–29.
[7] Daniela Pi aluga. 2009. La mensioc onologia dei Ma oni. Pe da a i e, Pe Conosce e e Pe Comp ende e le S u u e S o iche.ECIG.
[8] Nancy M. Mangiala di. 2015. The dimensional analysis o he b icks. The Case S udy o Luce a be ween he Thi een h and Fou een h
Cen u y. Bene en o, I alia.
[9] Lo enzo Can ini, Ma ia P e i ali, Rossella Moioli, and S e ano Della To e. 2019. The mensioch onology analysis suppo ed by he
new ad anced su ey echniques: Field es s in Milanese a ea. A ch. Pho og amm. Remo e Sens. Spa ial In . Sci XLII(2/W11), 359–365.
[10] José Ángel Asensio. 1995a. A qui ec u a de ie a y made a en la p o ohis o ia del alle medio del eb o y su elación con la del medi e -
áneo. Caesa augus a 71 (1995), 23–56.
[11] José Ángel Asensio. 1995b. La ciudad en el mundo p e omano en a agón. Caesa augus a 70 (1995), 1–461.
[12] Da id And ews. 1988. La mu a u a medie ale. Desc izione, analisi e s o ia economica. Cas um 2. S uc u es de l’habi a e occupa ion
du sol dans les pays Medi e aneens: Les me hodes e l’appo de l’a cheologie ex ensi e. Ac es de la econ e o ganisée pa l’ Ecole
F ançaise de Rome, Ghislaine Noyé (Ed.). Casa de Velázquez edi o es, 309–317.
[13] Tiziano Mannoni and Ennio Poleggi. 1974. Fon i sc i e e s u u e medie ali del “cas ello” di Geno a. A cheologia Medie ale 1 (1974),
171–194.
[14] Tiziano Mannoni, Au o a Cagnana, S e ano Falsini, Paola Ghislanzoni, and Daniela Pi aluga. 1991. A cheologia ed a cheome ia dei
mu i in Pie a: Supe ici e s u u e in ligu ia. In Le Pie e Nell’a chi e u a: S u u a e Supe ici: A i del Con egno Di S udi, b essanone.
Lib e ia P oge o Edi o e, 151–162.
[15] Au o a Cagnana and Tiziano Mannoni. 1995. Ma e iali e ecniche nelle s u u e mu a ie di Luni. Risul a i p elimina i. Quade ni di
S udi Lunensi 1(1995), 137–164.
[16] S e ano Della To e. 1996. S o ia delle echniche mu a ie e u ela del cons ui o. Espe ienze e Ques ioni di Me odo. Gue ina e Associa i,
B escia, Milano.
[17] Nicole a Benacchio and Paolo Cas ellani. 1997. Indagine sulle ipologie mu a ie di Assisi a XI e XIV secolo: p imi isul a i. A cheol.
A chi e . 2 (1997), 121–130.
[18] Y es Esquieu. 1997. L’a chéologie du bâ i en F ance. A cheol. A chi e . 2 (1997), 133–140.
[19] Michele Nuccio i. 2000. Le mu a u e medie ali di San a Fio a (Mon e Amia a-Toscana): Mensioc onologia dell mu a u e in Pie a, un
caso di s udio. A cheol. A chi e . 5 (2000), 65–85.
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.
10:20 • A. Mesanza-Mo aza e al.
[20] Daniel P igen and Ch is ian H. Sapin. 2004. La cons uc ion en pie e au moyen âge. In La Cons uc ion. Les Ma é iaux Du s: Pie e E
Te e Cui e. E an e ediciones, 117–148.
[21] Giuseppe Fiche a. 2005. A cheologia dell’a chi e u a di un edi icio ecclesias ico: La chiesa di San ’An imo sop a i canali a piombino.
A cheol. A chi e . 10 (2005), 153–171.
[22] Gian Pie o B ogiolo and Au o a Cagnana. 2012. A cheologia dell’A chi e u a—Me odi e In e p e azioni. All’Insegna del Giglio.
[23] Agus ín Azka a e, Ismael Ga cía-Gómez, and Amaia Mesanza-Mo aza. 2018. Análisis clús e : Un p ime paso sob e écnicas cuan i a-
i as en a queología de la a qui ec u a. A queol. A qui . 15 (2018), 1–18.
[24] Agus ín Azka a e, Leand o Cáma a, Juan Ignacio Lasagabas e , and Pablo La o e. 2001. Plan Di ec o de Res au ación de la Ca ed al de
San a Ma ía de Vi o ia-Gas eiz. Dipu ación Fo al de Ála a, Vi o ia-Gas eiz.
[25] Agus ín Azka a e and José Luis Solaun. 2013. A queología e His o ia de Una Ciudad. Los O ígenes De Vi o ia-Gas eiz. Vi o ia-Gas eiz.
Se icio edi o ial de la Uni e sidad del País Vasco/Euskal He iko Unibe si a ea.
[26] Leand o Sánchez-Zu iau e. 2007. Técnicas cons uc i as medie ales. Nue os Documen os A queológicos Pa a el Es udio de la Al a Edad
Media en Ála a. Se icio Cen al de Publicaciones del Gobie no Vasco, Vi o ia-Gas eiz.
[27] Albe o Pla a. 2008. Génesis de una illa medie al. A queología, Paisaje y A qui ec u a del Valle Salado de Añana (Ála a). Se icio Cen al
de Publicaciones del Gobie no Vasco, Vi o ia-Gas eiz.
[28] Albe o Pla a and Andoni E kiaga. 2018. El Sis ema de P oducción de Sal de Añana: Valle Salado (A aba-Ála a. País Vasco). Colección
Pa imonio, Te i o io y Paisaje. Vol. 3. Se icio edi o ial de la Uni e sidad del País Vasco/Euskal He iko Unibe si a ea, Vi o ia-
Gas eiz.
[29] Luis Caballe o. 2009. Edi icio his ó ico y a queología: Un comp omiso en e exigencias, esponsabilidad y o mación. A queol. A qui .
6 (2009), 11–19.
[30] B yan F. J. Manly. 2005. Mul i a ia e S a is ical Me hods. A P ime (3 d ed.). CRC P ess.
[31] Guille mo Rolón. 2013. La Vi ienda Popula Riojana del Ambi o Ru al: Pa ones A qui ec ónicos y Con ex o Social en los Valles du an e el
Pe iodo Republicano. PhD Disse a ion. Uni e sidad de Buenos Ai es, A gen ina.
[32] F ancisco Con e as. 1984. Clasi icación y ipología en a queología. In El Camino Hacia La Cuan i icación.Cuade nos de P ehis o ia y
A queología de la Uni e sidad de G anada 9 (1984), 327–385.
[33] Guille mo Rolón. 2014. Pa ones a qui ec ónicos, clús e es cons uc i os homogéneos y a iabilidad en el es udio de edi icios his ó icos.
Aspec os écnico- o males de la i ienda u al en la p o incia de La Rioja (A gen ina) du an e el pe iodo epublicano. A queol. A qui .
11 (2014), 1–24.
[34] Gayane Shalun s, Yll Haxhimusa, and Robe Sabla nig. 2011. A chi ec u al s yle classi ica ion o building acade windows. In P oceed-
ings o he 7 h In e na ional Symposium on Visual Compu ing. 280–289.
[35] Gayane Shalun s, Yll Haxhimusa, and Robe Sabla nig. 2012. A chi ec u al s yle classi ica ion o domes. In P oceedings o he 8 h
In e na ional Symposium on Visual Compu ing. 420–429.
[36] Maa en Bassie , Maa en Ve gauwen, and Bjo n Van Genech en. 2017. Au oma ed classi ica ion o he i age buildings o as-buil BIM
using machine lea ning echniques. ISPRS Annals o he Pho og amme y, Remo e Sensing and Spa ial In o ma ion Sciences,Volume
IV(2/W2), 26 h In e na ional CIPA Symposium 2017, O awa, 25–30.
[37] Kun Guo and Ning Li. 2017. Resea ch on classi ica ion o a chi ec u al s yle image based on con olu ion neu al ne wo k. In P oceedings
o he IEEE 3 d In o ma ion Technology and Mecha onics Enginee ing Con e ence (ITOEC’17). 1062–1066.
[38] Ma hias Zeppealzaue , Mi osla Despo o ic, and Mun aha Sakeena. 2018. Au oma ic p edic ion o building age om pho og aphs. In
P oceedings o he ACM In e na ional Con e ence on Mul imedia Re ie al (ICMR’18). 126–134.
[39] Luis Miguel Ma ínez-To es. 2014. Li hologic combina ions in Romanesque chu ches o Ala a, No he n Spain. Ma e . Cons ucc. 64,
313 (2014), 1–6.
[40] Manuel Cas iñei as. 2012. San Vicen ejo de T e iño, un edi icio excepcional en la enc ucijada del a do ománico hispánico. In Viaje a
íbi a, ed. Ayun amien o del Condado de T e iño (1s ed.). Condado de T e iño, España, 230–328.
[41] Juan Ignacio Lasagabas e , Agus ín Azka a e, and Ande de la Fuen e. 2006. Plan di ec o de in e ención pa a la es i ución de los al-
o es his ó icos y simbólicos de la basílica de san p udencio de a men ia y su en o no. G upo de In es igación en Pa imonio Cons uido,
UPV/EHU, 1–176.
[42] Pe e Swallow, Ross Dallas, Sophie Jackson, and Da id Wall. 2004. Measu emen and Reco ding o His o ic Buildings (2nd ed.). Donhead,
London.
[43] Da id And ews, Jon Bed o d, Bill Blake, Paul B yan, Tom C omwell, and Richa d Lea. 2010. Measu ed and D awn Techniques and
P ac ice o he Me ic Su ey o His o ic Buildings. English He i age, Swindon.
[44] Luis Caballe o. 2010. Expe iencia me odológica en a queología de la a qui ec u a de un g upo de in es igación. In A queología Apli-
cada Al Es udio e In e p e ación de Edi icios His ó icos. Úl imas Tendencias Me odológicas Minis e io de Cul u a (Ed.). Sec e a ía Gene al
Técnica. Subdi ección Gene al de Publicaciones, In o mación y Documen ación, 103–120.
[45] Luis Caballe o. 1995. Mé odo pa a el análisis es a ig á ico de cons ucciones his ó icas o “lec u a de pa amen os.” In . Cons ucc. 46,
435 (1995), 37–46.
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.
Machine Lea ning o he Buil He i age A chaeological S udy • 10:21
[46] I an zu Ál a ez-González, Ane Lope egi, Amaia Mesanza-Mo aza, José Manuel Valle, and I a xe Vicen e. 2003. Di e en es p opues as
pa a la ep esen ación geomé ica de edi icios his ó icos. A queol. A qui . 2 (2003), 9–12.
[47] Ismael Ga cía-Gómez, Mi en Fe nández de Go os iza, and Amaia Mesanza-Mo aza. 2011. Láse escáne y nubes de pun os. Un ho izon e
aplicado al análisis a queológico de edi icios. A queol. A qui . 8 (2011), 25–44.
[48] Res i u o Sie a. 1994. Técnicas de In es igación Social. Teo ía y Eje cicios. Pa anin o, Mad id.
[49] Agus ín Azka a e. 2010b. Es udio his ó ico-a queológico del casco his ó ico de Vi o ia-Gas eiz. P o ocolos de in e ención, modelos de
ges ión y sis ema de in o mación pa imonial (SWIP). G upo de In es igación en Pa imonio Cons uido. UPV/EHU.
[50] Agus ín Azka a e. 2010a. El análisis es a ig á ico en la es au ación del pa imonio cons uido. Conside aciones concep uales e in-
s umen ales. In A queología Aplicada al Es udio e In e p e ación de Edi icios His ó icos. Ul imas Tendencias Me odológicas. Minis e io de
Cul u a (Ed.). Sec e a ía Gene al Técnica. Subdi ección Gene al de Publicaciones, In o mación y Documen ación, 51–64.
[51] Gian Pie o B ogiolo. 2010. In oduzione. In A cheologia dell’ A chi e u a 15 (2010), 11–14.
[52] Noelia Oses, Fadi Do naika, and Abdelmalik Moujahid. 2014. Image-based delinea ion and classi ica ion o buil he i age mason y.
Remo e Sensing 3 (2014), 1863–1889.
[53] Da id S. Moo e. 2004. Es adís ica Aplicada Básica. (2nd ed), An oni Bosch (Ed.). Ba celona.
[54] Pablo Rod íguez-Gonzál ez, Diego González-Aguile a, Da id He nández-López, Hec o González-Jo ge. 2014. Accu acy assessmen
o ai bo ne lase scanne da ase by means o pa ame ic and non-pa ame ic s a is ical me hods. IET Sci., Meas. Technol. 9, 4 (2014),
505–513.
[55] Cha lo e H. Mason and Willian D. Pe eaul . 1991. Collinea i y, powe and in e p e a ion o mul iple eg ession analysis. J. Ma ke .
Res. 28 (1991), 220–268.
[56] Felipe Oli a and Pa icia Ve dú. 2008. Re isión de la u ilización de dis in os índices espec ales y écnicas pa a la ca og a ía de á ea
quemada en imágenes modis. Uned. Espacio, Tiempo y o ma. Se ie VI, Nue a Época. Geog a ía 1:129–152.
[57] Michael R. Be hold, Ch is ian Bo gel , F ank Höppne , and F ank Klawonn. 2010. Guide o In elligen Da a Analysis. How o In elligen ly
Make Sense o Real Da a.Sp inge ,London.
[58] Jo ge Ga cía-Gu ié ez. 2012. In elligen Techniques on LiDAR o En i onmen al Applica ions. PhD Disse a ion. Uni e sidad de Se illa.
[59] Pang-Ning Tan, Michael S einbach, and Vipin Kuma . 2014. In oduc ion o Da a Mining. Pea son.
[60] Lio Rokach and Oded Maimon. 2015. Da a Mining wi h Decision T ees. Theo y and Applica ions (2nd ed.). Wo ld Scien i ic Publishing
Co.P e.L d.
[61] Leo B eiman. 2001. Random Fo es s. In Machine Lea ning 45 (2001), 5–32.
[62] Leo B eiman. 1996. Ou -o -Bag Es ima ion. Uni e si y o Cali o nia, Be keley.
[63] John James. 1989. The empla e-make s o he Pa is basin. Toichological Techniques o Iden i ying he Pionee s o he Go hic Mo emen
wi h an Examina ion o A -his o ical Me hodology. Wes G ins ead Publishing, Aus alia.
[64] J. Richa d Landis and Ga y G. Koch. 1977. The measu emen o obse e ag eemen o ca ego ical da a. Biome ics 33 (1997), 159–174.
[65] Ismael Ga cía-Gómez. 2017. Vi o ia-Gas eiz y su Hin e land. E olución de un Sis ema U bano en e los Siglos XI y XV. Colección Pa i-
monio, Te i o io y Paisaje. Se icio edi o ial de la Uni e sidad del País Vasco/Euskal He iko Unibe si a ea.
[66] Luis Miguel Ma ínez-To es. 2007. Li hological maps o chu ches in he Diocese o Vi o ia (Spain): Space- ime dis ibu ion o building
s ones and ancien qua ies. Build. En i on. 45 (2007), 860–865.
[67] Cha les A. Lan z and Ellio Nebenzahl. 1996. Beha iou and in e p e a ion o he k s a is ic: Resolu ion o he wo pa adoxes. J. Clin.
Epidemiol. 49 (1996), 431–434.
[68] Juan An onio Ba celó. 1996. A queología au omá ica, In eligencia a i icial en a queología. Cuade n. A queol. Medi e . 2 (1996), 13–149.
Recei ed Ap il 2020; e ised Augus 2020; accep ed Sep embe 2020
ACM Jou nal on Compu ing and Cul u al He i age, Vol. 14, No. 1, A icle 10. Publica ion da e: Decembe 2020.