A Holistic Workflow for Semi-automated Object Extraction from Large-Scale Historical Maps

Vol.:(0123456789) 1 3 KN - Jou nal o Ca og aphy and Geog aphic In o ma ion (2023) 73:3–18 h ps://doi.o g/10.1007/s42489-023-00131-z A Holis ic Wo k low o Semi‑au oma ed Objec Ex ac ion omLa ge‑Scale His o ical Maps IngaSchlegel1 Recei ed: 23 Decembe 2022 / Accep ed: 13 Janua y 2023 / Published online: 10 Feb ua y 2023 © The Au ho (s) 2023 Abs ac The ex ac ion o objec s om la ge-scale his o ical maps has been examined in se e al s udies. Wi h he aim o esea ch u ban changes o e ime, semi-au oma ed and ans e able holis ic app oaches emain o be in es iga ed. We apply a combina ion o objec -based image analysis and ec o iza ion me hods on h ee di e en his o ical maps. By u he ma ching and geo e - e encing an app op ia e cu en geoda ase , we p o ide a concep o analyzing and compa ing hose aluable sou ces om he pas . Wi h mino adjus men s, ou end- o-end wo k low was ans e able o o he la ge-scale maps. The indings e ealed ha he ex ac ion and spa ial assignmen o objec s, such as buildings o oads, enable he compa ison o maps om di e en imes and o m a basis o u he his o ical analysis. Pe o ming an a ine ans o ma ion be ween he da ase s, an absolu e o se o no mo e han 72m was achie ed. The ou comes o his pape , he e o e, acili a e he daily wo k o u ban esea che s o his o ians. Howe e , i should be emphasized ha speci ic knowledge is equi ed o he p esen ed subjec i e me hodology. Keywo ds His o ical maps· Objec ex ac ion· Objec -based image analysis (OBIA)· Map compa ison· Vec o iza ion· Geo e e encing Ein holis ische Wo k low zu semi-au oma isie en Objek ex ak ion aus g oßmaßs äbigen his o ischen Ka en Zusammen assung Die Ex ak ion on Objek en aus g oßmaßs äbigen his o ischen Ka en is Gegens and zahl eiche Fo schungsp ojek e. Um den u banen Wandel im Lau e de Zei zu un e suchen, bedü en semi-au oma isie e und holis ische Ansä ze jedoch wei e en Un e - suchungen. In diese A bei we den Me hoden zu objek basie en Bildanalyse und Vek o isie ung au d ei e schiedene his o i- sche Ka en angewende . Mi hil e eines anschließenden Abgleichs sowie de Geo e e enzie ung eines en sp echenden ak uellen Geoda ensa zes s ellen wi ein Konzep o , das sowohl die Analyse als auch den Ve gleich de we ollen In o ma ionsquellen aus de Ve gangenhei e laub . Nu ge ing ügige Ände ungen wa en no wendig, um den ganzhei lichen A bei sablau au ande e g oßmaßs äbige Ka en zu übe agen. Unse e E gebnisse zeig en, dass die Ex ak ion und äumliche Zuo dnung on Objek en wie Gebäude ode S aßen einen Ve gleich zwischen Ka en e schiedene Zei al e e möglichen und somi eine G undlage ü wei e e his o ische Analysen scha en. Im Zuge eine a inen T ans o ma ion e gab sich eine maximale Abweichung on 72m zwischen beiden Da ensä zen. Die E gebnisse diese S udie e leich e n dami die ägliche A bei on z.B. S ad o sche n ode His o ike n. Dennoch soll e be ücksich ig we den, dass die o ges ell e subjek i e Me hodik spezi isches Fachwissen e o de . Schlüsselwö e His o ische Ka en· Objek ex ak ion· Objek basie e Bildanalyse (OBIA)· Ka en e gleich· Vek o isie ung· Geo e e enzie ung * Inga Schlegel inga.schlegel@hcu-hambu g.de 1 Lab o Geoin o ma ics andGeo isualiza ion, Ha enCi y Uni e si y Hambu g, Henning-Vosche au-Pla z 1, 20457Hambu g, Ge many 1 In oduc ion His o ical maps a e aluable sou ces when in es iga ing spa- ial changes o e ime (He old 2018). As an essen ial ool o communica ing geog aphic objec s and hei loca ions, hey 4 KN - Jou nal o Ca og aphy and Geog aphic In o ma ion (2023) 73:3–18 1 3 a e o en he only sou ce o in o ma ion o he unde s and- ing o spa io- empo al change (Sun e al. 2021; Kim e al. 2014). Wi h la ge-scale maps (app ox. > 1:20,000)—espe- cially “ci y maps”—we a e able o s udy u ban mo phology (Meinel e al. (2009), as ci ed in Muhs e al. 2016). F e- quen ly, geog aphical, poli ical, en i onmen al, and o he u baniza ion p ocesses can be back aced solely by means o his o ical maps. Fo he analysis o he u ban landscape o he pas , i is ine i able o make he in o ma ion om la ge-scale his- o ical maps accessible. Single map objec s may p o ide insigh s in o o me names o oads and buildings o hei e olu ion o e ime. Bu gene ally, physical scans (bi - maps) o his o ical maps a e no machine- eadable. Manual a emp s o acqui e in o ma ion om his o ical maps a e no uncommon bu e o -p one, ime-in ensi e, and non- ans- e able (Xydas e al. 2022; Chiang e al. 2020; Gobbi e al. 2019). The e is a need o (semi-)au oma ed app oaches o sol e hese p oblems. In his s udy, we p o ide a holis ic wo k low o no only ex ac objec s om la ge-scale his o ical maps, bu also o de i e bene i s om he en i e y o geome ic, ela ional, and seman ic in o ma ion. Mo eo e , ou semi-au oma ed app oach demons a es how a spa ial assignmen be ween his o ical and cu en maps may be enabled and he e o e p o ides a basis o u he compa ison p ocesses be ween hese. An es ablished s a egy used o semi-au oma ically ex ac objec s om his o ical maps while minimizing he eade ’s subjec i e in luence s a s wi h image segmen a- ion, which ollows he p inciples o human pe cep ion: objec s wi hin an image a e di e en ia ed due o g aphi- cal a ia ions (e.g., in ligh in ensi y, ex u e, o spa ial con ex ), a i ac s, and de ia ions. Visually homogeneous image a eas o m so-called segmen s. By combining objec segmen a ion and classi ica ion, he concep o geog aphic objec -based image analysis (GEOBIA) is able o ep o- duce physically exis ing objec s, like buildings o oads, om as e maps (He old 2018; Hussain e al. 2013; Hay and Cas illa 2008; Neube 2005). Howe e , au ho s ag ee ha “ he e is no single ex ac ion me hod ha can be e ec- i ely applied o all di e en his o ical maps” (Sun e al. 2021). This is a complex ask and only ew s udies ha e shown sugges ions o u he p ocessing and he applica- bili y o hei esul s. Mos esea ch in his ield aims a ex ac ing and ec- o izing geome ies om his o ical maps o make hem analyzable, bu equen ly comes wi h se e al limi a ions and p econdi ions. Many s udies ocus on he ex ac ion o a single ea u e ype such as s ee s (Chiang and Knoblock 2013; Chiang and Knoblock 2012), i e bodies (Gede e al. 2020), o di e en land use classes (Gobbi e al. 2019; Za elli e al. 2019) like o es a eas (Os a in e al. 2017; He aul e al. 2013; Leyk e al. 2006) o we lands (Jiao e al. 2020). O he s assume homogeneously colo ed map egions (Chiang e al. 2011; Leyk and Boesch 2010; Ablameyko e al. 2002), which is a ely ue o his o ical maps. Less complex (“bina y”) maps con aining homogeneously black objec s o con ou s on whi e backg ounds we e in es iga ed by Xydas e al. (2022), Hei zle and Hu ni (2020), Le Riche (2020), Iosi escu e al. (2016), Muhs e al. (2016), and Kim e al. (2014). Bu di e en ia ing objec s solely based on colo di e ences is insu icien especially o widesp ead monoch ome his o ical maps o due o ancien pape ex u e, noise, o di on he hand-d awn maps (Jiao e al. 2020; Pel- le 2018; Muhs e al. 2016; A eaga 2013; Leyk and Boesch 2010). Labels o en emain unconside ed in he con ex o objec ecogni ion om his o ical maps as hey commonly su e om o e laps o g ay-scale alues simila o ex u es o con ou s o o he map elemen s (Hei zle and Hu ni 2020; Pelle 2018). O he au ho s p esume an exis ing coo dina e sys em (Le Riche 2020; Gobbi e al. 2019; Iosi escu e al. 2016) o a huge s ock o aining da a, which is needed o machine lea ning app oaches (Xydas e al. 2022; Hei zle and Hu ni 2020; Jiao e al. 2020; Gobbi e al. 2019; Za elli e al. 2019; Uhl e al. 2017). Mo eo e , ew s udies ha e ocused on la ge-scale bu a he small-scale maps (Gede e al. 2020; Hei zle and Hu ni 2020; Gobbi e al. 2019; Za elli e al. 2019; Lo an e al. 2018; Uhl e al. 2017; Muhs e al. 2016; He aul e al. 2013). As exis ing esea ch gene ally ocuses on sepa a e p o- cesses in ol ed in objec ex ac ion om his o ical maps, ou s udy sugges s a holis ic app oach composed o ex ac - ing, ec o izing, and linking objec s. We demons a e he bene i s o elimina ing and assigning labels o his whole p ocess and p esen applicabili ies o he esul ing geom- e ies. Because only by conside ing hese echniques as a whole, we a e able o answe loca ion- ela ed ques ions on he e olu ion o geog aphic ea u es and make his o ical maps “accessible o geospa ial ools and, hus, o spa io- empo al analysis o landscape pa e ns and hei changes” (Uhl e al. 2017). New quali a i e and quan i a i e analyses as well as compa isons o o he his o ical o cu en geo- da a become possible by sea ching h ough and p ocessing in o ma ion de i ed om his o ical maps (Gobbi e al. 2019; Chiang 2017; Iosi escu e al. 2016). Fo long- e m back ac- ing o indi idual buildings, o ins ance, shape-based com- pa isons ac oss di e en maps a e use ul (Le Riche 2020; Laycock e al. 2011). In his wo k, we p esen a semi-au oma ic solu ion o make la ge-scale his o ical maps usable o spa ial analysis while minimizing ime-in ensi e and labo ious manual use in e en ion. Based on ou p e ious indings on he needs o use s o his o ical maps (Schlegel 2019) as well as on he iden i ica ion and ex ac ion o map labels (Schlegel 2021), we demons a e he gene al easibili y o a comp ehensi e 5KN - Jou nal o Ca og aphy and Geog aphic In o ma ion (2023) 73:3–18 1 3 wo k low composed o (1) elimina ing labels, (2) ex ac - ing geome ies, (3) ec o izing and e ining hose, and (4) ma ching and spa ially assigning he ex ac ed map objec s wi h cu en ones. Po en ial u u e applica ions, which a e shown in he u he cou se, may be in ol ing seman ic in o ma ion om labels o anno a e co esponding map ea u es o an adjus men o a map’s isual appea ance. P ospec i ely, new da abases can be se up and compa a i e s udies be ween di e en da ase s become possible. 2 Li e a u e Re iew 2.1 Elimina ion o Labels Labels a e aluable componen s in his o ical maps holding impo an me ada a. Howe e , ex wi hin a map is ypically seen as a dis u bing ac o when ex ac ing geome ies. Mis- in e p e a ions in he con ex o segmen a ion may easily a ise due o o e laps, di ec adjacencies, o simila colo alues o map elemen s and s uc u es such as lines o ex- u es (Hei zle and Hu ni 2020; Bhowmik e al. 2018; Chi- ang 2017). Monoch ome maps, in pa icula , ha e a educed numbe o pa ame e s o di e en ia e be ween ex and o he elemen s. Howe e , an ini ial elimina ion o ex o labels om his o ical maps can be seen as a majo ad an age o u he objec ex ac ion p ocesses (Gede e al. 2020). P e- ious a emp s iden i ied labels wi h he help o ex ec- ogni ion—subsequen o objec ecogni ion and ec o iza- ion—o by shape ecogni ion algo i hms (Iosi escu e al. 2016). Ch yso alan is and Nikolaos (2020) used bina ized maps o sepa a e ex om o he objec s (see also Bhowmik e al. (2018)). By elimina ing small pixel g oups, hey we e able o emo e le e s. A GRASS GIS add-on de eloped by Gobbi e al. (2019) and Za elli e al. (2019) eplaces el- e an pixel alues by means o low-pass il e s wi hin old cadas e maps. Howe e , pixels mus al eady be de ined as “ ex ” in ad ance. Telea (2004) and Be almío e al. (2001) sugges di e en image inpain ing echniques, which a e o en applied o image es o a ion. Missing o damaged image egions a e illed o c ea e an image wi hou gi ing he iewe a hin o changes. In ou es ing, hese app oaches caused an unsa is ac o y blu ing o he inpu image. 2.2 Objec ‑Based Image Analysis Many me hodologies o (semi-)au oma ed objec ex ac- ion om his o ical maps we e demons a ed in ecen yea s bu p o en insu icien o a ious easons. Fo ins ance, a common his og am h esholding o colo space clus e ing (He aul e al. 2013) igno es any spa ial con ex , whe eas a i icial neu al ne wo ks equi e an inadequa e amoun o aining da a (Gobbi e al. 2019). Ch yso alan is and Nikolaos (2020) used GIS unc ion- ali ies o con e a his o ical mul icolo map in o a bina y image and hen o ex ac and ec o ize geome ies o build- ings. Howe e , ex u ed o co up polygons could no be handled and labels we e elimina ed only pa ially. A simi- la app oach was conduc ed by Iosi escu e al. (2016). By combining GIS ope a ions wi h Py hon lib a ies, Gede e al. (2020) segmen ed and ec o ized geome ies o i e s as a unc ion o hei colo whe eas Le Riche (2020) ex ac ed buildings om his o ical maps based on colo s and ex u es. Za elli e al. (2019) and Gobbi e al. (2019) used GIS and R o segmen and classi y ea u es om his o ical land use maps by ega ding hei colo s, sizes, and shapes. Addi ional machine lea ning echniques we e applied by Gobbi e al. (2019). In ecen yea s, deep lea ning a emp s ia con olu ional neu al ne wo ks (CNNs) “ha e ecen ly ecei ed conside - able a en ion in objec ecogni ion, classi ica ion, and de ec- ion asks” (Uhl e al. 2017) om his o ical maps (Jiao e al. 2020, Hei zle and Hu ni 2020, and Xydas e al. 2022). How- e e , hey su e om majo d awbacks. Resul s om CNNs s ongly depend on he quali y and gene ally low quan i y o a ailable aining da a. O en, hese da a s ocks a e c ea ed manually and solely on he basis o he inpu bi map i sel , which is ime-consuming and impedes an applicabili y. O igina ing om he ield o emo e sensing, geog aphic objec -based image analysis may also be applied o scans o maps (Hay and Cas illa 2008). In he b oad ield o ca - og aphy, only ew au ho s use OBIA app oaches o c e- a e new geoda a. Whe eas Do nik e al. (2016) ep oduced soil maps om clima e and ege a ion maps, Ke le and de Leeuw (2009) ex ac ed poin -based popula ion da a om pape maps o es ima e long- e m popula ion g ow h. Edle e al. (2014) applied OBIA o ex ac and quan i y he p es- ence o oads, buildings, and land use classes and o u he e alua e he complexi y o opog aphic maps he eby. In con as o pixel-wise app oaches, OBIA ega ds no only spec al in o ma ion, bu also, e.g., he shape, size, o neighbo ly ela ions o objec s, and is, he e o e, much close o human pe cep ion. Hence, OBIA is o en sug- ges ed o objec ex ac ion om his o ical maps wi h he aim o make hem machine-in e p e able (Blaschke e al. 2014). Many s udies in he ield o OBIA ocus on maps o colo s and smalle scales, p esuppose a p eceding geo- e e encing (Ch yso alan is and Nikolaos 2020; Gede e al. 2020; Iosi escu e al. 2016) o well-de ined shapes o objec s (Ch yso alan is and Nikolaos 2020; Gobbi e al. 2019; Hei - zle and Hu ni 2020), o dis ega d in e sec ions be ween map ea u es. 6 KN - Jou nal o Ca og aphy and Geog aphic In o ma ion (2023) 73:3–18 1 3 2.3 Vec o iza ion andVec o Enhancemen As ec o da a can be be e p ocessed and analyzed han as e da a, a majo i y o he men ioned au ho s p oceed wi h a ec o iza ion o ex ac ed map objec s. B own (2002) and A eaga (2013) use speci ic so wa e ools o, espec i ely, ec o ize he ou lines o geologic s uc u es and buildings om his o ical maps. Vec o iza ion ools a e also p o ided wi hin A cGIS, GRASS GIS (Gede e al. 2020), and he GDAL lib a y (Jiao e al. 2020). To pu ge ec o ized objec s, u he simpli ica ion p o- cesses may ollow. Mul iple so wa e and ools, including eCogni ion, QGIS, A cGIS (God ey and E ele h 2015), SAGA GIS (Gede e al. 2020), R (A eaga 2013), and Py hon lib a ies, implemen p e-buil unc ions o smoo h o simpli y lines o polygons and o emo e ou lie s, spikes, and o he a i ac s. 2.4 Objec Ma ching Fo he di ec compa ison o ec o objec s om di e en maps om a ious imes, dis ance and simila i y measu es may be p omising (Xa ie e al. 2016). Ma ching geome ies be ween di e en inpu s is equen ly pe o med on he basis o shape o spa ial simila i ies (Tang e al. 2008) o iden ical a ibu e alues (F ank and Es e 2006). Howe e , seman ic simila i y app oaches a e no easible as scanned his o ical maps usually hold no ancilla y in o ma ion. E en i names o oads o buildings we e a ailable—e.g., by a p eceding ex ecogni ion— hey would need o be assigned o hei co esponding geome ies. When analyzing geome ic ela ions, such as o e lapping a eas o dis ance measu es (e.g., Euclidian o Hausdo dis- ance), only ela i e dis ances be ween objec s a e consid- e ed. This echnique is useless when compa ing no ye geo- e e enced da ase s (Xa ie e al. 2016). Region-based shape desc ip o s (e.g., a ea, con ex hull, Momen o G id desc ip- o , see Ahmad e al. (2014)) ega d all pixels wi hin a shape and may he e o e be p omising o a compa ison be ween iden ical eal-wo ld objec s om di e en inpu s. Bu due o unce ain ies, hey a e a he conside ed complemen a y ma ching app oaches. Addi ional simila i y measu es a e necessa y (Xa ie e al. 2016). By ega ding he spa ial ela- ionship be ween objec s, S e anidis e al. (2002) quan i ied hei dis ances and ela i e posi ions. Samal e al. (2004) and Kim e al. (2010) consul ed hi d objec s o c ea e an o e all geog aphic con ex . Also, Sun e al. (2021) ega ded spa ial ela ionships by linking iden ical eal-wo ld objec s om di e en his o ical maps. Howe e , hei knowledge g aph app oach p esupposes he exis ence and assignmen o labels o hei co esponding geome ies. 3 Da a Fo a p oo o concep o ou sugges ed me hodologies, a la ge-scale (~ 1:11,000) his o ical map om he middle o he nine een h cen u y was chosen, which has al eady been objec o esea ch wi hin ela ed s udies (Schlegel 2019, 2021). The o iginal non-geo e e enced and undis o ed e sion o he map scan was c opped o a smalle ex en (~ 1000 × 800m in eali y) o easons o un ime comp es- sion wi hin all p ocesses. No map p ojec ion is known. The map subse in Fig.1 shows he ci y cen e o Hambu g wi h blocks o buildings, oads, and wa e a eas. Apa om sub- sequen ly colo ized wa e a eas, he map is d awn in black and whi e. Many da a supplie s p o ide hei as e scans wi h a esolu ion o 300 ppi which is conside ed adequa e o objec ex ac ion pu poses (Pea son e al. 2013). Lowe pixel densi ies induce blu ing and pixela ion, whe eas highe al- ues end o highligh in e e ing a i ac s om, e.g., olds in pape , discolo a ions, o smudges (Pelle 2018). We con in- ued o wo k wi h he TIFF o ma (wi hou comp ession) as i is lossless conce ning he image’s o iginal pixel alues (Gede e al. 2020). To demons a e he ans e abili y o he wo k low, wo mo e la ge-scale his o ical maps co e ing he same spa ial a ea we e used in he u he cou se (see Fig.9a, b). They all di e in hei isual appea ance and complexi y in e ms o con as s, ex u es, o he exis ence o labels and g idlines. Fo compa ing he desc ibed da a o a cu en coun- e pa , o icial ec o da ase s including ecen polygonal buildings (Landesbe ieb Geoin o ma ion und Ve messung 2022) and line- ype oads (Behö de ü Ve keh und Mobil- i ä swende (BVM) 2020) we e used. 4 Objec Ex ac ion 4.1 P epa a ion o  heElimina ion o Labels As simila colo alues and o e laps be ween labels and o he map objec s impede a clea disc iminabili y, an ini ial elimina ion o labels designa ing eal-wo ld objec s signi i- can ly con ibu es o a acili a ion o objec ecogni ion p o- cesses. We sugges o make use o he ou pu om p e ious label de ec ion a emp s (see Schlegel (2021)): ec o bound- ing boxes comp ising ex image a eas, which can be seen in Fig.1. An exempla y ex image a ea is shown in Fig.2a. Wi h he aim o elimina e i s con en om he map, i was ini ially c opped by means o i s o iginal bounding box (see Fig.2b) and o a ed o he ho izon al by i s angle o align- men (Fig.2c)—calcula ed by he used ex de ec ion ool S abo (Li e al. 2018; Chiang and Knoblock 2014). How- e e , hese ex image a eas do no only include cha ac e s, 7KN - Jou nal o Ca og aphy and Geog aphic In o ma ion (2023) 73:3–18 1 3 bu also edges o buildings, which is an ou g ow h o S abo (see uppe ma gin in Fig.2c). This is coun e p oduc i e wi hin he subsequen s ep o building segmen a ion as hese image a eas we e supposed o be en i ely elimina ed om he map. Thus, building edges would become dis o ed. To e ain hese impo an edges, all pixels wi hin a bounding Fig. 1 Map subse showing he ci y cen e o Hambu g (Ha a d Map Collec ion, Ha a d College Lib a y e al. (n.d.)) wi h bounding boxes con aining labels p oduced by a p e ious ex de ec ion Fig. 2 S eps o sepa a ing building edges om labels shown wi h an exempla y da ase : a inpu map wi h bounding box con aining ex image a ea, which hen was b c opped, c aligned ho izon ally, and d con e ed in o a bina y as well as e a h ee-class mask. The esul - ing bounding box excluding building edges was g u ned back o i s o iginal o ien a ion 8 KN - Jou nal o Ca og aphy and Geog aphic In o ma ion (2023) 73:3–18 1 3 box we e di e en ia ed by ex and pa s o buildings. A use -de ined h esholding helped o gene a e a bina y mask consis ing o da k “ o eg ound” and b igh “backg ound” pixels (see Fig.2d). A u he “ o eg ound” di e en ia- ion was needed o sepa a e building edges om ex pix- els. Howe e , simila colo alues, o e laps, and smoo h ansi ions be ween ex and buildings we e challenging. Fo eclassi ying o me “ o eg ound” in o ei he “ ex ” o “building edge” pixels, mul iple h esholds and condi ions had o be applied (Fig.2e). As labels designa ing oads mos commonly un pa allel o nea by building edges, his s ep was pe o med ow-wise. As Fig.2 indica es, all pixels ep- esen ing “ ex ” and “backg ound” we e combined and ec- o ized. The esul ing polygonal bounding box was u ned back by i s ini ial o a ion angle (see Fig.2g) and hen used wi hin he ollowing objec ex ac ion s eps. 4.2 Objec De ec ion andRecogni ion To de ec homogeneous image egions and ex ac objec s such as buildings o wa e a eas om la ge-scale his o ical maps, we used objec -based image analysis. In con as o pixel-based app oaches (e.g., Maximum Likelihood, Clus e - ing, o Th esholding), which only ega d spec al di e ences be ween pixels, OBIA gene a es image objec s also based on common ex u es, shapes, con ex , e c. and is, he e- o e, mo e sui able o his o ical maps wi h limi ed spec al in o ma ion and he e ogeneous appea ances (Blaschke e al. 2014; Hussain e al. 2013). As none o he many ee and open sou ce packages a ailable o semi-au oma ed ea u e ex ac ion p oduces compa able esul s, we made use o he p op ie a y so - wa e eCogni ion De elope 10.2 o gene a e GIS compa - ible da a om a his o ical map ia OBIA (Kau and Kau 2014). eCogni ion con e s use -de ined ule se s—buil -up om unc ions, il e s, s a is ics, e c. o image segmen a ion and classi ica ion—in o machine- eadable code. These con- ca ena ions o algo i hms can be easily ans e ed o o he images (T imble Inc. 2022). As Fig.3a indica es, a i s ough di e en ia ion be ween da k ( o eg ound) map ea u es (e.g., buildings and labels) and he map’s b igh backg ound (wa e a eas, oads, and places) was enabled by h esholding he inpu TIFF. The con en o he labels’ bounding boxes, as shown in Fig.2 , was simply classi ied as “backg ound” and could he e o e be elimina ed (see Fig.3b). The de ec ion o u he map objec s is he e o e signi ican ly acili a ed on he one hand and building edges emain unal e ed on he o he hand. To ex ac con ou s o buildings, an edge de ec o was applied o he image. The building ex u e’s epea ing pa e n could be de ec ed by means o a g ay-le el co-occu ence ma ix—which measu es he e ical in a iance o adjacen pixel pai s—and analyzed by ex u e desc ip o s (Cha es 2021; T imble Inc. 2021). Rega ding he o iginal map in Fig.1, public buildings (e.g., he ownhall o chu ches) ha e a signi ican ly da ke ex u e and could, he e o e, clea ly be di e en ia ed om o he buildings based on hei g ay alues. Wa e a eas we e iden i ied by h esholding he RGB blue channel as well as applying supplemen a y ex u e desc ip o s o a oid alse posi i es. 4.3 Vec o iza ion Gene ally, OBIA esul s in as e iles con aining indi idual image objec s, subdi ided in o p ede ined single classes. Fo u he p ocessing and analysis pu poses, a ec o iza ion o his da a is ine i able. Based on expe iences o Iosi escu e al. (2016) and A eaga (2013), we applied GDAL’s polygonize unc ion o pe o m a as e - o- ec o con e sion Fig. 3 Fo eg ound objec s sepa a ed om he map’s backg ound a be o e and b a e elimina ing labels 9KN - Jou nal o Ca og aphy and Geog aphic In o ma ion (2023) 73:3–18 1 3 o each map class. Se e al unc ions o simpli y and smoo h he ec o ized map ea u es, o close inlying mino gaps, and elimina e small isola ed polygons we e compiled wi hin an end- o-end Py hon sc ip . This way, in e e ing a i ac s (e.g., islands, p o usions, o spikes) s emming om an imp ecise segmen a ion o unde ec ed labels could be han- dled. The esul ing polygons ep esen ing (public) buildings and wa e a eas a e shown in Fig.4 and can be p ocessed wi hin u u e analysis ope a ions. 5 Linking His o ical andCu en Da ase s Compa ed o p e ious s udies dealing wi h objec ex ac- ion om his o ical maps, we go one s ep u he and p e- sen an exempla y way o how quali a i e and quan i a i e e alua ions o long- e m changes wi hin a ci yscape may be p ac ically enabled. We he e o e spa ially assigned a mo e ecen ec o da ase o he his o ical coun e pa as shown in Fig.7. Ou aim was o au oma e his coa se geo e e encing p ocess as a as possible. Due o changing names o oads and buildings o e ime, he lack o in- dep h in o ma ion, o simply imp ecise scales, dis ances, and di ec ions wi hin his o ical maps, we used he p e i- ously ex ac ed geome ies o geo e e encing pu poses (Rumsey and Williams 2002). As can be seen om Fig.5, chu ches and o he municipal buildings s ill exis o e ime and, beyond ha , do no subs an ially change hei basic shape and geog aphic loca ion o e ime. The e o e, hei objec shapes could be ma ched and used o he de ini ion o con ol poin s in he u he cou se o geo e e encing (Skopyk 2021; Ha licek and Caj haml 2014). 5.1 Shape Ma ching To de ine ma ching geo e e encing con ol poin s be ween he his o ical and cu en da ase , iden ical eal-wo ld objec s a e o be iden i ied. We, he e o e, measu ed he shape simi- la i y be ween he ex ac ed public buildings shown in Fig.5 (Sun e al. 2021; MacEach en 1985). A ma ching based on Fig. 4 Vec o ized and e ised ea u es o he his o ical map 10 KN - Jou nal o Ca og aphy and Geog aphic In o ma ion (2023) 73:3–18 1 3 spa ial o seman ic (a ibu e-based) simila i ies was imp ac- ical due o he lack o a coo dina e sys em as well as u he in o ma ion conce ning he his o ical map. As Fig.5 indica es, a side-by-side compa ison be ween geome ies o public buildings ex ac ed om he his o ical map on he one hand and he o icial ec o da ase con ain- ing cu en buildings on he o he hand was pe o med. We implemen ed a ma ching o hei shapes based on hei In e - sec ion o e Union (IoU). A e adjus ing he aspec a ios o co esponding coun e pa s ia ec angula bounding boxes (“en elopes” (Es i 2022)), hei espec i e de ia ions could be quan i ied ia IoU. As can be seen om Fig.6, a building geome y and i s en elope oge he o m a bina y mask—consis ing o he alues 1 (building geome y) and 0 (en elope). A inal supe imposi ion o hese masks helped o de e mine hei o e lapping a ea (in e sec ion) p opo ion- ally o hei common a ea (union) (see Fig.6). All “building” pixels wi h a alue o 1 we e conside ed o he IoU cal- cula ion, which was conduc ed wi h he help o Py hon’s numpy lib a y. Table1 summa izes he IoU esul s o all de ec ed public buildings con inued o use o geo e e enc- ing pu poses. 5.2 Geo e e encing 5.2.1 Me hod O e iew The cen oids o hose geome ies wi h he closes ma ches (see highligh ed cells in Table1) we e de ined as con- ol poin s o a semi-au oma ed, ough geo e e encing be ween he his o ical and cu en da ase . To p ese e he objec s’ shapes and o keep spa ial de o ma ions o a minimum wi hin he his o ical da a, an a ine ans o ma- ion o all cu en buildings and oads was conduc ed. This Fig. 5 Vec o ized public buildings (chu ches and ownhall) om he his o ical map (uppe ow) and hei coun e pa s om he cu en da ase (bo om ow) Fig. 6 In e sec ion o e Union be ween he his o ical S . Pe i Ki che and a i s cu en coun e pa as well as b he cu en S . Ka ha inen Ki che. The aspec a io o he geome ies’ en elopes was adjus ed o one ano he , espec i ely 11KN - Jou nal o Ca og aphy and Geog aphic In o ma ion (2023) 73:3–18 1 3 was done ia QGIS Vec o Bende (Dalang 2019) using he h ee ma ching objec pai s highligh ed in Table1 as well as Fig.7. In ou es case, only h ee con ol poin s wi h su icien poin ing accu acy could be ound—such a small numbe is qui e ypical o his o ical maps. Howe e , i a ailable, a la ge quan i y o con ol poin s is ad isable Table 1 Nume ical esul s om In e sec ion o e Union be ween his o ical and cu en buildings’ geome ies His o ical map S . Ka ha inen Ki che S . Pe i Ki che Ra haus Cu en da ase S . Ka ha inen Ki che 83,9% 74,9% 45,0% S . Pe i Ki che 79,3% 84,4% 43,1% Ra haus47,5% 45,9% 58,5% S . Jacobi Ki che 82,1% 80,2% 42,6% Mahnmal S . Nikolaia54,0% 49,8% 26,1% aThe S . Jacobi Ki che was no classi ied as public building by eCogni ion whe eas he (Mahnmal) S . Nikolai was econs uc ed in ano he ci y dis ic a e being mainly des oyed du ing Wo ld Wa II and lea ing only i s owe un il oday (Claussen n.d.). a The S . Jacobi Ki che was no classi ied as public building by eCogni ion whe eas he (Mahnmal) S . Nikolai was econs uc ed in ano he ci y dis ic a e being mainly des oyed du ing Wo ld Wa II and lea ing only i s owe un il oday (Claussen n.d.) Fig. 7 Geo e e enced cu en buildings and oads based on he cen oids o he highligh ed S .-Ka ha inen-Ki che, S .-Pe i-Ki che, and Ra haus 18 KN - Jou nal o Ca og aphy and Geog aphic In o ma ion (2023) 73:3–18 1 3 Pelle P (2018) F om pape map o geospa ial ec o laye : demys i y- ing he p ocess. IASSIST Q 42(3):1–24 Rumsey D, Williams M (2002) His o ical maps in GIS. In: Knowles AK (ed) Pas ime, pas place: GIS o his o y. Es i P ess, Red- lands, CA, pp 1–18 Samal A, Se h S, Cue o K (2004) A ea u e-based app oach o con la- ion o geospa ial sou ces. IJGIS 18(5):459–489. h ps:// doi. o g/ 10. 1080/ 13658 81041 00016 58076 Sammlung Ch is ian Te s egge (n.d.) 1722, on C. F i zsch, pho o-li h- og aphisches Replika om Ve lag S umpe & Co., 1880. h ps:// www. c h is ian- e s egg e. de/ hambu g/ ka en_ hambu g/ iles/ 1722_ i z sch_ s um pe _ 300dpi. jpeg. Accessed 4 Feb 2022 Schlegel I (2019) Empi ical s udy o a deploymen o a me hodology o imp o ing he compa abili y be ween his o ical and cu en maps. KN - J Ca og Geog In o 69(2):121–130. h ps:// doi. o g/ 10. 1007/ s42489- 019- 00016-0 Schlegel I (2021) Au oma ed ex ac ion o labels om la ge-scale his- o ical maps. AGILE GIScience Se ies. h ps:// doi. o g/ 10. 5194/ agile- giss-2- 12- 2021 Skopyk B (2021, Ma ch 8) Geo e e encing His o ical Maps: Me hods, bes p ac ices, and a u o ial in A cGIS P o. A cGIS S o yMaps. h ps:// s o y maps. a cgis. com/ s o i es/ dd75d 0398 7d4de d924d 30316 1895b 8b. Accessed 30 June 2022 S e anidis A, Agou is P, Geo giadis C, Be olo o M, Ca swell JD (2002) Scale- and o ien a ion-in a ian scene simila i y me ics o image que ies. IJGIS 16(8):749–772. h ps:// doi. o g/ 10. 1080/ 13658 81021 01485 52 Sun K, Hu Y, Song J, Zhu Y (2021) Aligning geog aphic en i ies om his o ical maps o building knowledge g aphs. IJGIS 35(10):2078–2107 Tang W, Hao Y, Zhao Y, Li N (2008) Fea u e ma ching algo i hm based on spa ial simila i y. In: P oceedings o SPIE 7147, Geoin o ma ics 2008 and Join Con e ence on GIS and Buil En i onmen : Classi ica ion o Remo e Sensing Images, 714704. h ps:// doi. o g/ 10. 1117/ 12. 813204 Telea A (2004) An image inpain ing echnique based on he as ma ch- ing me hod. JGT 9(1):25–36. h ps:// doi. o g/ 10. 1080/ 10867 651. 2004. 10487 596 T imble Inc. (2021) Ha alick ex u e. h ps:// docs. ecogn i ion. com/ 10.0. 2/# eCogn i ion_ docum en a ion/ Re e ence Book/03 Fea u es/2 Objec ea u es/4 Tex u e/Ha alick ex u e. h m?Highligh =ha alick. Accessed 25 May 2022 T imble Inc. (2022) The Powe o eCogni ion. h ps:// docs. ecogn i ion. com/# eCogn i ion_ docum en a ion/ Modul es/1 eCogni ion a a glance/01 In oduc ion o he Powe o eCogni ion.h m. Accessed 24 May 2022 Uhl JH, Leyk S, Chiang YY, Duan W, Knoblock CA (2017) Ex ac ing Human Se lemen Foo p in om His o ical Topog aphic Map Se ies Using Con ex -Based Machine Lea ning. In: P oceedings o he 8 h In e na ional Con e ence o Pa e n Recogni ion Sys- ems. h ps:// doi. o g/ 10. 1049/ cp. 2017. 0144 Xa ie EMA, A iza-López FJ, U eña-Cáma a MA (2016) A su ey o measu es and me hods o ma ching geospa ial ec o da ase s. ACM Compu Su 49(2):1–34. h ps:// doi. o g/ 10. 1145/ 29631 47 Xydas C, Kesidis A, Kaloge opoulos K, Tsa sa is A (2022) Buildings Ex ac ion om His o ical Topog aphic Maps ia a Deep Con o- lu ion Neu al Ne wo k. In: P oceedings o he 17 h In e na ional Join Con e ence on Compu e Vision, Imaging and Compu e G aphics Theo y and Applica ions 5, pp 485–492. h ps:// doi. o g/ 10. 5220/ 00108 39700 003124 Za elli P, Gobbi S, Ta oni C, La Po a N, Ciolli M (2019) Objec -based image analysis o his o ic maps classi ica ion. In A ch Pho o- g amm Remo e Sens Spa In Sci XLII-4/W14:247–254