evaluation criteria for recent lod proposals for city gml buildings · introduced a level of detail...

PFG 2016 / 1, 031–043 ArticleStuttgart, February 2016

© 2016 E. Schweizerbart'sche Verlagsbuchhandlung, Stuttgart, Germany www.schweizerbart.deDOI: 10.1127/pfg/2016/0282 1432-8364/16/0282 $ 3.25

Evaluation Criteria for Recent LoD Proposals for CityGML Buildings

Marc-O. Löwner, Braunschweig & Gerhard GröGer, Bonn

Keywords: Virtual 3D city models, semantical data models, CityGML, Level of Detail,evaluation criteria

for emergency driving training (Randt et al.2007). Some application areas such as emer-gency management (Zlatanova & li 2008) orindoor navigation (BeckeR et al. 2009) evenrequire information of the building’s interioron a city level. These models may vary withregard to geometrical and semantical com-plexity and to the degree of deviation from thecorresponding real world objects. Complexitylevels then are the result of specific data acqui-

1 Introduction

Virtual 3D city models represent single build-ings, city quarters, whole cities and even re-gions for applications such as noise propa-gation simulation and mapping (cZeRwinski

et al. 2007), fine dust distribution modelling(Ghassoun et al. 2015), urban and telecommu-nication planning (köninGeR & BaRtel 1998,knapp & cooRs 2008), or real-time simulations

Summary: The Level of Detail (LoD) concept is anessential part of the Open Geospatial Consortiumstandard CityGML and of 3D city models in gen-eral. New applications such as indoor navigationand energy performance estimation ask for a revi-sion of the current CityGML 2.0 LoD concept. Cur-rently, new approaches are discussed that need tobe evaluated. Here, new evaluation criteria for theassessment of recent Level of Detail concepts in thefield of semantical 3D city models are developed.They cover richness of aspects, completeness of theconcept, completeness of models in a particularLoD, avoidance of inconsistent models, freedom ofinterpretation, and feasibility and complexity oftransformation from CityGML 2.0 into the pro-posed concept. These criteria represent an addedvalue because user defined LoD profiles of the newCityGML version 3.0 LoD concept are likely andtherefore need to be evaluated. Applying the devel-oped criteria, we evaluate the most current propos-als on the further development of the CityGML 2.0LoD concept.

Zusammenfassung: Kriterien zur Bewertung ak-tueller Vorschläge für Gebäude-LoD in CityGML.Der Detaillierungsgrad (Level of Detail, LoD)stellt ein wesentliches Konzept von CityGML, deminternationalen Standard des Open GeospatialConsortiums für semantische 3D-Stadtmodelle,dar. Neue Anwendungsfelder wie die Innenraum-navigation oder die Energiebedarfsanalyse erfor-dern allerdings eine Überarbeitung der in City-GML 2.0 definierten Konzepte und haben bereitseine Folge von Verbesserungsvorschlägen nachsich gezogen. Hier werden neue Evaluationskriteri-en zur Bewertung aktueller LoD-Konzepte von se-mantischen 3D-Stadtmodellen vorgestellt. DieseKriterien umfassen die Reichhaltigkeit der Aspek-te, die Vollständigkeit des Konzeptes, die Vollstän-digkeit der Modelle innerhalb eines LoD, die Ver-meidung inkonsistenter Modelle, die Interpreta-tionsfreiheit sowie die Möglichkeit und Komplexi-tät der Transformation von CityGML 2.0 in dasvorgeschlagene Konzept. Die Definition solcherKriterien ist wichtig, weil für CityGML 3.0 dieMöglichkeit diskutiert wird, Profile der LoD Defi-nition zuzulassen. Mittels der neu entwickeltenKriterien werden die aktuellen Verbesserungsvor-schläge des LoD Konzeptes bewertet.

32 Photogrammetrie • Fernerkundung • Geoinformation 1/2016

sessment criteria for the comparison of LoDapproaches which are relevant for 3D citymodel collectors, providers and users are stillmissing.

Currently, there is a discussion in the City-GML community whether the fixed and stand-ardized geometrical representation for a par-ticular LoD, e.g. LoD1Solid or LoD2Multi-Surface for buildings, should be replaced bya more flexible and generic model where eachCityObject can be represented by any geom-etry type. According to this framework, re-stricting profiles may be defined for particularCityObjects including one official profile inthe CityGML Specification. Thus, evaluationcriteria become even more important.

Here, the assessment criteria richness of as-pects, completeness of the concept, complete-ness of models in a particular LoD, avoidanceof inconsistent models, as well as feasibilityand complexity of transformation from City-GML 2.0 for the evaluation of LoD conceptsare introduced and applied to the most recentapproaches. The term ‘aspect’ denotes a com-ponent of the data model that is relevant forthe LoD concept, for example geometry, se-mantics, appearance, or topology. However,since no formal definitions are given, we seethis contribution as a beginning of a discus-sion about requirements on LoD concepts.

In section 2 we give an overview on recentLoD concepts for semantical 3D city modelsfocussing mainly on the CityGML Buildingmodel. The criteria, which are the base of thecomparison, are introduced in section 3. Theevaluation and comparison of the approachesintroduced in section 2 is presented in section4. We will end with conclusions and an out-look.

2 Recent Level of DetailConcepts for SemanticalBuilding Models

This section gives an overview over three re-cent LoD concepts for semantical 3D buildingmodels that are related to the further develop-ment of CityGML. For a detailed descriptionof the CityGML 2.0 LoD concept reference ismade to löwneR et al. (2013), GRöGeR et al.(2012), and GRöGeR & plümeR (2012). A more

sition processes or they may be used to assessthe suitability of data for specific applications.

The Open Geospatial Consortium (OGC)CityGML standard (GRöGeR et al. 2012,GRöGeR & plümeR 2012) for the representationof semantically enriched 3D city models hasintroduced a Level of Detail (LoD) concept inorder to support different applications of 3Dcity models. It does not only cover the geomet-rical detail level, but also the semantical one,i.e. the richness of feature types modelled.The current LoD concept of CityGML 2.0 pro-vides five discrete levels of detail. Its defini-tion is widely accepted in the scientific com-munity (e.g. BoGuslawski et al. 2011, Quinn etal. 2009, iwasZcZuk & stilla 2010, Fan et al.2009, GötZelmann et al. 2009, GueRke et al.2009). The term “LoDx model”, x ∈ {0, 1, 2, 3,4}, is frequently used to address the complex-ity of existing city models and their suitabilityfor specific applications. However, some defi-ciencies have been identified, which hamperthe use of CityGML for important applica-tions. In particular, indoor objects are coupledwith the highest LoD in CityGML 2.0, imply-ing highly complex semantics and geometry.Hence, applications requiring only coarse in-door models or indoor models in combina-tion with a coarse exterior are not supported.Further, there are no multiple LoD for indoorobjects in CityGML 2.0 (e.g. coarse or highlydetailed 3D representations, or 2D footprints)which are required for indoor navigation(domínGueZ et al. 2011, haGedoRn et al. 2009).In addition, the explicit representation of win-dows in an outer wall of a building is only pos-sible in LoD3, which requires an accurate geo-metrical representation of the building façade.To estimate a building’s energy demand, how-ever, explicit information on the area coveredby windows is needed, whereas a very coarserepresentation of the façade’s geometry is suf-ficient (dalla costa et al. 2011).To overcome these deficiencies, modifica-

tions or extension of the LoD concept havebeen proposed (BoeteRs et al. 2015, Biljecki

et al. 2013, Biljecki et al. 2014, BenneR et al.2013, löwneR et al. 2013, naGel 2014). Un-til today, none of these proposals or possiblecombinations of their single aspects have beenevaluated to be the best LoD concept for 3Dcity models in a comparative approach. As-

M.-O. Löwner & G. Gröger, Evaluation Criteria for Recent LoD Proposals 33

2.2 The Biljecki Approach

The Biljecki approach (Biljecki et al. 2013)does not only focus on CityGML, but also onproprietary approaches for 3D city modelsthat are developed by commercial companies.An LoD is defined as a quality measure withregard to a specific application. This meas-ure is related to a variety of aspects includingrichness of feature types, attribute richness,complexity of geometrical details, appearancequality, and positional accuracy. Separate hi-erarchies for geometry and semantics are pro-posed, which have to be defined by users. Fur-ther, it is proposed to define constraints foreach LoD, which assure its consistency. Anexample is a constraint that prevents interiorgeometries without exterior ones.Amodification of this approach with a simi-

lar set of six aspects is presented in Biljecki

et al. (2014). These aspects are applied to theexterior and the interior of features. They spana space of six dimensions, and an LoD is de-fined as a vector of six values or ranges of val-ues. Only consistent series of LoD0,…, LoDnare considered, which have to be monotonic ineach aspect. This means that from one LoD(i)to the next LoD(i+1) in the sequence, the val-ues of the six aspects increase or remain un-changed, but never decrease. Hence, there isa total order on the LoD in a series that thanallows for the comparison of two LoD. As anexample for the implementation of the frame-work, a series LoD0 to LoD9 is defined. Themain aspects are existence of features, geo-metrical correspondence between model andreality, and resolution of the appearance.These 10 LoD are roughly a refinement of theLoD0 to LoD4 in CityGML 2.0. The conceptis implemented as an extension of CityGML(formally, an Application Domain Extension).

2.3 The Nagel Approach

The Nagel approach (named after claus

naGel who was the first to propose this con-cept) is built on practical experience and di-rectly related to the further development ofCityGML 2.0 (löwneR et al. 2015, naGel

2014). It needs just two definitions. First, eve-ry city object has a spatial representation in

general overview of LoD concepts for seman-tic 3D city models can be found in BenneR etal. (2013).

2.1 The Benner Approach

The Benner approach (named after its maincontributor joachim BenneR) is directly re-lated to the further development of the City-GML standard (BenneR et al. 2013, löwneR etal. 2013). It exhibits two modifications: First,there is an explicit separation between a geo-metrical and a semantical LoD. Second, thecurrent LoD4 has been mapped to four LoDfor the interior. Hence, a building is parti-tioned into an exterior and an interior, bothwith one or more explicit LoD of geometricaland semantical aspects.

For geometry, the Benner approach sup-ports four different representations (LoD0 toLoD3) for all top-level features of the City-GML building model, whether they representthe building’s exterior shell (Building, Build-ingPart, …) or interior components (Room,…). BenneR et al. (2013) identify the semanti-cal structuring and classification as an impor-tant criterion for virtual 3D city models. Theydefine four different Semantical Levels (S0 –S3) for the Building model’s top level features.As for the Geometrical LoD, the SemanticalLevels of exterior shell and rooms may be dif-ferent. If a building refers to more than oneroom, all corresponding Room features musthave the same Semantical Level.

Since this extended concept allows forbuilding models representing the exteriorbuilding shell as well as interior rooms in dif-ferent geometrical and semantical LoD withonly some restrictions on acceptable combi-nations, new labels have been defined that ex-tend the existing CityGML 2.0 labels. Theyare a combination of geometrical and seman-tical labels where the latter is put in brackets,e.g. LoD2.1(S2.0) for a building with an exteri-or shell in LoD2 together with interior Objectsin LoD1 where Building, BoundarySurfacesand BuildingInstallations for the exterior andRooms for the interior are semantically clas-sified.


LoD, i.e. the richness of feature types such aswall surfaces, installations, or building partsrepresented in the model. In addition, otheraspects of a model can be considered, e.g. itsappearance, attribute data and so forth. How-ever, a useful LoD concept should cover morethan just geometry or, at least should be exten-sible to other aspects like semantic or appear-ance. An increasing complexity of the conceptitself and the labelling of instance documentsare inextricably linked to this criterion.

Completeness of the conceptCompleteness of the concept is the degree towhich the concept allows for the representa-tion of data. A concept is complete if all data-sets can be represented. The more restrictionsapply, the lesser is the completeness. Evenmodels which may be considered as incon-sistent might lead to a higher degree of com-pleteness. An example is a 3D dataset withcoarse interior objects (blocks model rooms)with highly detailed outer shell penetrated bythe room geometries. Such a model is wellsuited for energy applications, where a roughestimation of the indoor volume is sufficient,but detailed roof structures with dormers arerequired for solar panel placement planning.In addition, this criterion plays a role for leg-acy data. Such datasets often violate consis-tency rules of data models or LoD concepts,e.g. thresholds for geometrical accuracy, to-pological inconsistencies, or missing semanti-cal classifications, but might be valuable sincethere are no other datasets available.

Since inconsistent models may increase thecompleteness of the concept, this criterion isinverse to the criterion ‘avoidance of inconsis-tent models’.

This criterion is crucial for users of 3Dcity models, which have to check whether thedata needed for an intended application canbe represented by a particular LoD concept.Likewise, for data collectors it is importantto check whether captured data can be repre-sented.

Completeness of models in a particular LoDThis criterion is the degree to which com-pleteness of instant models in a single LoD isforced by the concept. A model is completein a particular LoD if the concept states that

every LoD that refines its spatial representa-tion in higher LoD. That means secondly, thatthere is no restriction on the usage of any fea-ture type in an LoD. Thus, even feature typesthat have been limited to CityGML 2.0 LoD4,e.g. a CeilingSurface, can be used in any lowerLoD. In this approach, LoD0 stands for planarrepresentations and LoD1 for prismatic blocksmodel representations of a feature. Further,LoD2 models represent a generalised shape ofa CityObject whereas LoD3 represents it in itshighest geometrical complexity.In order to improve flexibility, a distinc-

tion is made between volumetric geographicfeatures and BoundarySurfaces. An Abstract-Building, for instance, is then modelled withzero to two instances of a GM_MultiSurfacein LoD0, representing the footprint or theedges of a roof. For the representation in LoD1to LoD3 it is modelled as a GM_MultiSurfaceor as a GM_Solid, respectively. An Abstract-BoundarySurface, which might be possible inall four LoD is represented by a GM_Multi-Curve in LoD0 and zero to one GM_MultiSur-faces in LoD1 to LoD3. Here, the curve repre-sentation stands for the footprint as a spatialabstraction of that wall surface.

The proposed concept stands out with itsclear and short definition. By allowing all fea-ture types being modelled in LoD0 to LoD3an explicit LoD for interior features becomesobsolete.

3 Definition of Criteria for theComparison of recent LoDConcepts

In order to compare the LoD concepts, sixcriteria are introduced: richness of aspects,completeness of the concept, completeness ofmodels in a particular LoD, avoidance of in-consistent models, freedom of interpretation,and feasibility and complexity of transforma-tion from CityGML 2.0.

Richness of aspectsA Level of Detail concept for semantical 3Dcity models may define more than just the as-pect of geometrical similarity of the model andthe real world feature. In contrast to comput-er graphics, it also may cover the semantical


metrical level if interior structures penetratethe exterior shell due to different interior andexterior LoD. Semantically, inconsistencymay occur if more semantical information isallowed to be attached to a coarse and undif-ferentiated geometry. Therefore, this assess-ment criterion results also from the complete-ness of concept.

Freedom of interpretationAn LoD concept may restrict the freedom ofinterpretation of the modeller. Therefore, itshould be clearly defined and avoid ambiguityto result in comparable results when applyingmodelling rules. Clarity of definition is partic-ularly of interest for the relationship betweenvendor and customer. Both parties shouldagree whether a concrete instance model issuited for a specific application and whetherthe model contains all the data to fulfil theuser’s demands. Further, a precisely definedLevel of Detail concept aiming at improvingdescription of the dataset can lead to a fasterand more robust development of applicationsoftware, e.g. with respect to support switch-ing between different LoD. This also entailsa comprehensible and unambiguous labellingof different LoD. Informative value of the la-bel is an essential element when restricting thefreedom of interpretation. Thus, the expres-sion of a clearly defined LoD concept leads tounambiguous instance models.

For collectors and providers of data, how-ever, an ambiguously defined model hasthe advantage of less effort to build up a 3Dcity model. In addition, legacy models mightfit better to an ambiguously defined model.Hence, the scope of such models is signifi-cantly wider.

Feasibility and complexity of transforma-tion from CityGML 2.0Since 2008 many companies, federal survey-ing agencies and municipalities, responsiblefor the representation and provision of their 3Dcity models use CityGML, worldwide. Hence,the required effort to transform these modelsto a newer version of CityGML with a modi-fied LoD concept (if this transformation is fea-sible at all) is an important criterion. If modelsconforming to the new LoD concept are syn-tactically and semantically compatible with

all objects or their parts specified in the par-ticular LoD and existent in the real world arerepresented in the dataset. One preconditionfor completeness is the LoD being defined forprincipal objects, e.g. buildings, primarily forthose, which have to represented completely.The LoD of the subordinated features (bound-ary surfaces, installations, etc.) is determinedby the LoD of the principal objects. For exam-ple, if a concept provides boundary surfacesand installations in LoD2 and defines that a‘LoD2 building’ has an complete outer shelland is completely bounded by boundary sur-faces, the LoD2 model of this concept is com-plete. Likewise, if an ‘LoD3 building’ requiresa complete outer shell with boundary surfacesand openings as feature types, it is complete.If the LoD is defined for each feature type sep-arately, there are no dependencies and, thus,the completeness of models in an LoD is low.This concept would allow representing a sin-gle opening or single furniture without anybuildings or rooms.

This criterion refers to the rules of an LoDconcept and to the degree to which the con-cept demands completeness. Hence, this cri-terion can be evaluated by inspecting the con-cept. It is another issue outside the scope ofthis discussion whether a particular dataset iscomplete with regard to an LoD specification(an issue of quality principles evaluation pro-cedures, c.f. ISO 19157:2013). This complete-ness typically can be assured based on refer-ence data only.

This criterion is of interest for applicationswhich require complete models, for exampleto assess the energy loss of a building, whichrely on models where each opening (door orwindow) is represented as a feature in thebuilding model.

Avoidance of inconsistent modelsThis criterion determines whether the LoDspecification contains a sufficient explicit orimplicit set of integrity rules to assure that themodel instances are meaningful. The risk toproduce inconsistent models can be avoidedby declaring an appropriate LoD concept forvirtual 3D buildings. Weak definitions maylead to inconsistent model instances even ifthe modeller follows all rules and restrictionsdefined. Inconsistency may occur on the geo-


4.1 Richness of Aspects

The CityGML 2.0 LoD concept enables therepresentation of a CityObject in differentgeometrical LoD and provides a refined se-mantical description with ascending LoD.However, geometric and semantic complex-ity of a building model is strictly coupled. InLoD0 and LoD1 no further decomposition ofa Building or BuildingPart into other featureclasses or semantic classification is possible.LoD2 represents the prototypic roof shape of abuilding, thematic ground, wall and roof sur-faces as well as installations, such as balconiesand dormers. LoD3 as the most detailed levelfor the outer shape allows for the semanticalrepresentation of openings. LoD4 adds inte-rior structures like rooms, furniture, interiorinstallations. While a certain LoD enforcesa specific geometric representation, the in-crease of semantic complexity is only option-al. Therefore, it cannot be stated that a seman-tical LoD is entirely independent.

Regarding geometry, the Benner approachsupports four different representations for alltop-level features of the CityGML buildingmodel regardless whether they represent thebuilding’s exterior shell (Building, Building-Part, etc.) or interior components. Next to theaspect of geometry BenneR et al. (2013) definefour different Semantical Levels (S0 – S3) for

CityGML 2.0, the complexity is zero. If thereis a simple one-to-one mapping between thenames of the elements in the model, the com-plexity is very low. The complexity is high ifcomplex structural or geometrical transforma-tions or classification processes are requiredto transfer datasets to the new concept.

This criterion is related to the freedom ofinterpretation. If it is higher in the sourcethan in the target concept, mapping requiresthe often complex classification of the vagueone. For example, if geometry details of thesource concept are vaguely defined but man-datory in the target concept, the geometry hasto be classified (for example, as LoD1, 2 or 3geometries) in order to assign the level in thetarget concept.

4 Comparison of four recent LoDConcepts

Based on the criteria introduced in section3, we now evaluate the discussed LoD ap-proaches and compare them. An overview ofthe evaluation results is given in Tab. 1. Theorder of evaluation is the same for all crite-ria: 1) CityGML 2.0, 2) Benner approach,3) Biljecki approach, (example LoD0 – LoD9instancing the Biljecki framework will be ex-amined additionally), and, 4) Nagel approach.

Tab. 1: Overview of the evaluation of LoD concepts according to the criteria developed in section3. Ratings range from ‘– –’ (not fulfilled at all) to ‘++’ (completely fulfilled). ‘/’ means that the crite-rion is not applicable.

CityGML 2.0 Benner Biljecki(framework)

Biljecki(LoD0 – 9)

Nagel

Richness of aspects ± + + + + –

Completeness ofthe concept ± + ± ± + +

Completeness of Models inan LoD ± + + + + + + – –

Avoidance of inconsistentmodels + + – – + + + + – –

Freedom of interpretation + + ± – – – – + +

Feasibility and complexityof transformation fromCityGML 2.0

/ + – – – – + +


els, at least in lower LoD. A floor plan may beconsidered as an example. Because all featuretypes can be used to represent the 2D repre-sentation of a story, i.e. IntBuildingInstalla-tion, InteriorWallSurface, Door, Window, etc.the floor plan is semantically rich and can bequeried for wall surfaces, doors, columns, andso forth.

4.2 Completeness of the LoDConcept

The completeness of the LoD concept is lowfor CityGML 2.0, since there are a lot of re-strictions between the two aspects of geom-etry and semantics. First, indoor objects canonly be represented in the most detailed LoD.Coarse room representations, which are rele-vant for energy applications, for example, areout of scope of the concept. Second, semanti-cal richness and complexity of the geometryare coupled. Not all combinations of geometri-cal and semantical complexity are supported.Hence, datasets which are suitable or requiredfor some applications cannot be representedby CityGML. For example, openings cannotbe represented in combination with coarserouter shell of a building. Also rooms cannot becombined with such a coarse outer representa-tion. On the other hand, incomplete instancemodels in an LoD (see next criterion) can berepresented by CityGML 2.0, for example asingle opening without any building or wall.Hence, the completeness of the LoD conceptis high for such ‘incomplete’ instance models.

In the Benner approach rooms can be rep-resented in any geometrical detail level, andcombinations of indoor objects and the outerbuilding shell are possible. A geometrical de-tail level can be combined with an arbitrary se-mantical level, due to the separation into geo-metrical and semantical LoD. Only a few re-strictions remain such that boundary surfaces,openings and building installations cannot begeometrically represented in LoD0 or LoD1.

In the Biljecki approach, the exterior andthe interior are separated. Hence, each com-bination of indoor and outdoor LoD can berepresented. However, the completeness is re-stricted by the monotonicity condition, statingthat the value of an aspect cannot decrease in

the Building model’s top level features. As forthe Geometrical LoD, the Semantical Levelsof exterior shell and rooms may be different.If a building refers to more than one room, allcorresponding Room features must have thesame Semantical Level.

In comparison to the CityGML 2.0 LoDconcept the Benner approach reveals a higherrichness of aspects including a graduation ofinterior details, the independent representa-tion of different semantical LoD for exteriorand interior and an independent semanticalLevel of Detail for both, the interior and ex-terior building.The Biljecki approach defines a very rich

LoD concept involving six aspects, whichconsider semantics (features, attributes), ge-ometry (accuracy, dimensionality), appear-ance and the relation between geometry andsemantics. The exterior and the interior of fea-tures are considered separately. Combinationsof these aspects are restricted by the conceptof series of LoD, which forces an increase ofthe values of the particular aspects if the LoDnumber increases. The proposed implementa-tion of the approach (LoD0 – LoD9) uses onlya subset of the aspects of the general frame-work (feature complexity, dimensionality, andappearance). Nevertheless, the richness ofboth concepts is significantly higher than therichness of all other proposals.

The Nagel approach allows all CityObjectsto be modelled in every LoD, geometrically.Consequently, an extra LoD for interior fea-ture types, the CityGML 2.0 LoD4, becomesunnecessary. These results in four LoD fromLoD0 for surface representations of real-worldfeatures, e.g. the ground surface of a building,to LoD3 for the most detailed representationof modelled features. For instance a bounda-ry surface, which can only be modelled fromLoD2 in the CityGML 2.0 LoD concept, canbe represented as a curve in LoD0 that repre-sents the ‘surface of a wall’ in the real word.Therefore, the Nagel approach is more flexiblethan the CityGML 2.0 LoD concept.

No semantical LoD is introduced by theNagel approach. However, since the CityGML2.0 LoD concept restricts the usage of somefeature types in lower LoD, e.g. a boundarysurface in LoD0, it is argued, that the Nagelapproach allows for semantically richer mod-


geometry and for the semantical content. Thelatter is explicitly defined by a list of featuretypes which have to be present in the model.Hence, the models are complete with regard tothese labels.

The general framework of Biljecki providesan aspect ‘Feature Complexity’, which is ex-plicitly defined as ‘fineness of geometry withrespect to the real world’. As a straightforwardway to denote this, the minimal length of ob-jects is mentioned; if this condition is fulfilled,the object has to be represented in the model.Hence, the models are complete, at least forobjects which are larger than this minimal di-mension. Since the concrete implementationof the framework contains the aspect ‘FeatureComplexity’, it is complete as well.The Nagel approach defines LoD for each

feature separately and independently. If abuilding feature, for example, is labelled asLoD3, no statement on the occurrence of sub-ordinate feature types are made. Hence, com-pleteness of models in a particular LoD is low.

4.4 Avoidance of Inconsistent Models

The CityGML 2.0 LoD concept for buildingsis very straight-lined because the LoD attri-bute is obliquely devoted to the building ob-ject itself. As a result, higher detailed build-ings consist of an optional superset of featuresdescribing the building of the particular lowerLevel of Detail. In addition, the list of featuresthat represent a building in a specific LoD isdefined, thus preventing the assignment ofmore semantical information to a coarse andundifferentiated geometry. Further, the high-est geometrical representation for both, theexterior shell and the interior is required inLoD4 so that interior structures cannot pene-trate the exterior shell. Thus, a violation of se-mantics and geometry is virtually impossible.

The Benner approach determines the com-plete LoD of a building by the combination ofGeometric LoD and Semantical LoD. There-fore they define feasible combinations. Geo-metric LoD0 models can only be combinedwith Semantical Level 0, while geometricLoD1, LoD2 and LoD3 models are allowed torepresent four variants with increasing seman-tical complexity. These restrictions result in 13

a higher LoD. Hence, geometry and semanticsare coupled, as it is the case in CityGML 2.0.The concept is more restricted than the Ben-ner approach, which allows for combinationsof geometric and semantical aspects, despitethe restrictions mentioned.

The concrete implementation of the Biljeckiconcept by defining LoD1 to LoD9 is signifi-cantly more restricted than the general frame-work, since exterior and interior are coupled.Coarse indoor objects cannot be representedwith a fine exterior. Hence, the completenessis comparable to the completeness of the City-GML 2.0 LoD concept.

The Nagel approach is based on geometryonly and has no restrictions at all. Hence, thecompleteness of the LoD concept is maximal.

With regard to new application areas whichinitially were the motivation for a modifiedLoD concept, both, the Nagel and the Bennerapproach meet the requirements. The Nagelapproach is not restricted at all while in theBenner approach, the two aspects geometryand semantics are independent and can be ar-bitrarily combined. In the general Biljecki ap-proach, the combination of aspects is restrictedby the monotonicity condition. Since exteriorand interior are separated, more applicationsare supported. In the concrete implementation,however, interior and exterior are coupled.Hence, neither energy applications (coarse in-terior / fine grained exterior) nor indoor appli-cations (2D interior objects) are possible.

4.3 Completeness of Models in aparticular LoD

In CityGML 2.0, a LoD is defined for prin-cipal objects such as buildings and it is im-plied that the corresponding models are com-plete in that LoD. However, this is stated onlynarratively in the specification, but is not de-fined in the mandatory part: neither the UMLdiagram, nor the XML schemas or the con-formance requirements state this explicitly.Hence, the completeness of models in a par-ticular LoD is low.

In the Benner approach, an LoD is also de-fined for principal objects. The geometrical aswell as the semantical label explicitly state thecontent of such an object. This holds for the


set of such feature types is provided. Hence,the semantical richness is optional.

In addition to a geometrical representation,an LoD3 model might have boundary surfac-es, openings, and installations, which againmight have boundary surfaces. However, itis not strictly required that such features arepresent in an LoD3 building representation.Therefore, all of the following models wouldbe qualified as LoD3: building geometry only,building geometry and thematic surfaces, the-matic surfaces only, building geometry, the-matic surfaces and openings as well as the-matic surfaces and openings. Since Building-Parts and BuildingInstallations are other op-tional feature types, which both can be com-bined with these five options, the number ofdifferent LoD3 models is at least 20. Here,only semantical variability is considered. Ifadditionally geometrical aspects are takeninto account, this number increases signifi-cantly.

Consider a simple blocks model with the-matic boundary surfaces and openings as anexample for a problematic LoD classification.The question is how to classify such a model.Classification as LoD1 or LoD2 is impossible,since these schemata do not provide openings.Technically, it is possible to qualify such amodel as LoD3. Admittedly, this violates thedefinition of LoD3, which ‘denotes architec-tural models with detailed wall and roof struc-tures’ (GRöGeR et al. 2012). To sum up, it canbe said that freedom of interpretation is highfor the CityGML 2.0 approach.The Benner approach defines a compact

nomenclature incorporating exterior and in-terior Geometrical and Semantical Levels,which specializes the CityGML 2.0 LoD in-dicators and results in a substantially higherinformative value of the label. It indicates thegeometrical modeling style and the semanticmodeling depth of a Building or BuildingPart,consistently extending the CityGML LoD no-tation. Ambiguities are almost impossible andthe freedom of interpretation is low.

The aspects in the Biljecki approach are de-fined very precisely. The criterion ‘Presenceof city objects and elements’ defines the fea-ture types which have to be present in a par-ticular LoD. The ‘Feature Complexity’ allowsspecifing concrete minimal dimensions of fea-

feasible combinations of Geometrical LoD andSemantical LoD for the building exterior andinterior and make inconsistent models of geo-metric and semantic combination impossible.

The Benner approach distinguishes be-tween exterior and an interior Level of De-tails. Thus, geometrical problems might occurif the LoD of the geometrical detail level ofthe exterior and the interior are not identical.For instance, if rooms have a detailed geome-try, but the exterior is coarse, fitting problemsare likely to happen. Either the room geometrypenetrates the exterior shell, or empty spacesinside the building are not covered by rooms.Vice versa, if rooms have a coarse geometry,but the exterior shell is fine-grained, the roomgeometry might penetrate the exterior shell(rf. Fig. 15 in BenneR et al. 2013).

In the general framework of Biljecki, thereis a strict separation between indoor and out-door objects. Hence, inconsistent modelsmight also happen here. However, it is stat-ed in Biljecki et al. (2014) that ‘the interior isconstrained with the exterior’. However, it isnot completely clear what this exactly meansand whether those inconsistencies are prevent-ed. In the concrete implementation LoD0 toLoD9 of the Biljecki framework, the geome-try detail increases for indoor and outdoor ina similar way and are comparable to the City-GML 2.0 concept. Hence, violations betweenindoor and outdoor objects are not possible.The Nagel approach defines LoD not for

an entire building, but for every particu-lar CityObject with no restriction on how acomposed building should be modelled us-ing CityObjects in different LoD. Thus, thisproposal might lead to inconsistent models asalready identified for the Benner approach.Moreover, the label for a certain LoD does notbelong to an entire building any longer, but toits composing feature types.

4.5 Freedom of Interpretation

The CityGML 2.0 specification provides defi-nitions of the particular LoD in terms of ge-ometry and semantics that are very vague withrespect to geometrical complexity. Semanticsof an LoD is defined in terms of feature classesfor buildings and its parts, but only a maximal


if the conditions of LoD3 are completely ful-filled. If openings are not represented, it ismapped to LoD3(S2). If BoundarySurfacesare not represented, it is mapped to LoD3(S1).Finally, LoD4 is mapped to LoD3.3(S3.3), ifthe conditions of LoD4 are completely ful-filled. However, it is mapped to LoD3.x(S3.y),if the conditions for semantical completenessare fulfilled only partially. The value for x hasalready been defined in the mapping for LoD3,whereas y is set to 2 if openings are not rep-resented completely, y is set to 1 if in addi-tion the interior building installations and thebuilding furniture are missing. Finally, y is setto 0 if boundary surfaces are not present. Thefulfillment of the semantical conditions caneasily be checked by scanning the dataset.

The Biljecki approach provides strict, pre-cisely defined requirements for the geome-try. In order to check whether for instance aCityGML 2.0 LoD3 model satisfies the mini-mal dimension of the corresponding LoD ofBiljecki, it is not sufficient to consider the da-tasets only. Instead, the corresponding realworld objects have to be inspected, which isvery elaborate and expensive. For the transfor-mation into the concept of Biljecki, first theaspects on which the LoD concept is based onhave to be compared. The geometry aspect ofCityGML corresponds most likely to the ‘Fea-ture Complexity’. Whether semantics is mod-elled or not in a particular LoD is covered bythe ‘Presence of CityObjects and Elements’aspect. Hence, the CityGML 2.0 aspects area subset of the Biljecki aspects and, in gener-al, datasets can be transformed. However, theproblem of vague geometry definition in City-GML 2.0 versus a very precise definition ofthe Biljecki approach remains. Due to this, thetransformation is very elaborate.

For the example LoD 0 to 9, the mappingis as follows: For LoD0, there is no counter-part; at least not for buildings (there are no ar-eal representations for buildings). LoD1 cor-responds to LoD1, LoD2 to LoD3 (since thereis no interior in LoD2, the interior representa-tion in LoD3 is empty). LoD3 is representedas LoD8 (again without interior) and finallyLoD4 is represented as LoD9.

The transformation of CityGML 2.0 modelsaccording to the Nagel approach is straight-forward: The LoD of each feature can be de-

tures, and ‘Attribute data’ a concrete list of at-tributes. Hence, the freedom of interpretationis low. This also holds true for the concreteimplementation of the framework. For eachLoD (0 to 9), concrete values for the featurecomplexity such as minimal size of objects isgiven, e.g. 10 m in LoD0 for building blocksor 10 cm in LoD9.

In the Nagel approach, only one criterion,the detail level of geometry, is used. Howev-er, this criterion is not defined precisely, butonly intuitively by giving figures as examples.Hence, the freedom of interpretation is sup-posed to be high.

4.6 Feasibility and Complexity ofTransformation from CityGML 2.0

The Benner and the Nagel approach havein common that LoD4 has been replaced byLoD3 for interior and exterior objects. The ad-ditional LoD4 outer geometry has deliberate-ly been introduced into CityGML in order toprovide the possibility to adapt the outer shellto the interior objects, in particular to fit theopenings of rooms. Although the LoD4 out-er geometry may differ from the LoD3 outergeometry, only one representation (LoD3) isprovided by both, the Benner and the Nagelapproach. Hence, there might be a loss of in-formation when transforming CityGML 2.0datasets containing LoD3 and LoD4 into ei-ther concepts.

A general problem when transforming da-tasets from CityGML 2.0 is the vagueness ofdefinitions with respect to geometry. Even ifit contradicts the common understanding ofLoD3, a coarse block can be used as a LoD3representation without breaking syntacti-cal rules or conformance requirements of theCityGML 2.0 concept.

Apart from the aforementioned problems,the mapping from CityGML 2.0 into the Ben-ner approach can be easily performed. LoD0is mapped to LoD0(S0) while LoD1 is mappedto LoD1(S1). Further, LoD2 is mapped toLoD2(S2), if the conditions of LoD2 (seman-tics) are completely fulfilled. If the semanticconditions of S1 respectively S0 are fulfilled,LoD2 is mapped to LoD2(S1) or LoD2(S0),respectively. LoD3 is mapped to LoD3(S3),


It is expected that the evaluation presentedhere will be very valuable for the definition ofthe LoD profiles of the new LoD concept ofCityGML 3.0. The group which will developthe official LoD profile can profit from the re-sult, as well as users which will define ownLoD profiles, in order to accommodate fortheir applications of 3D city models. This ap-proach is innovative, since up to now, evalua-tion criteria for LoD concepts have not beendeveloped and no systematic evaluation andcomparison of current LoD approaches werepossible.

References

BeckeR, t., naGel, c. & kolBe, t.h., 2009: A Mul-tilayered Space-Event Model for Navigation inIndoor Spaces. – lee, j. & Zlatanova, s. (eds.):3D Geo-Information Sciences, Lecture Notes inGeoinformation and Cartography: 61–77,Springer, Berlin, Heidelberg.

BenneR, j., GeiGeR, a., GRöGeR, G., häFele, k.h. &löwneR, m.o., 2013: Enhanced LoD Conceptsfor Virtual 3D City Models. – ISPRS Annals ofthe Photogrammetry, Remote Sensing and Spa-tial Information Sciences II-2/W1: 51–61.

Biljecki, F., Zhao, j., stoteR, j. & ledoux, h.,2013: Revisiting the concept of level of detail in3D City Modelling. – ISPRS Annals of the Pho-togrammetry, Remote Sensing and Spatial In-formation Sciences II-2/W1: 63–74.

Biljecki, F., ledoux, h., stoteR, j. & Zhao, j.,2014: Formalisation of the level of detail in 3Dcity modeling. – Computers, Environment andUrban Systems 48: 1–15.

BoeteRs, R., aRRoyo ohoRi, K., Biljecki, F. &Zlatanova, S., 2015: Automatically enhancingCityGML LoD2 models with a correspondingindoor geometry. – International Journal of Geo-graphical information Science 29 (12): 2248–2268; http://doi.org/10.1080/13658816.2015.1072201.

BoGuslawski, p., Gold, c. & ledoux, h., 2011:Modeling and analyzing 3D buildings with a pri-mal/dual data structure. – ISPRS Journal of Pho-togrammetry and Remote Sensing 66 (2): 188–197.

cZeRwinski, a., sandmann, s., stöckeR-meieR, e.& plümeR, l., 2007: Sustainable SDI for EUnoise mapping in NRW – best practice for IN-SPIRE. – International Journal for Spatial DataInfrastructure Research 2 (1): 90–111.

rived immediately from its geometry proper-ties. Only a simple syntactical transformationis required. Hence, it is the LoD concept withthe smallest transformation complexity.

5 Summary and Discussion

This paper evaluates and compares differentproposal for the revision of the CityGML 2.0LoD concept. Therefore, new evaluation crite-ria are developed reflecting the needs that arerelevant for stakeholders of 3D city models.These criteria cover such questions as: Canall 3D city building models be represented bythe concept? To which degree are inconsistentmodels prohibited, as they nevertheless mightbe adequate for some applications? Can userrely on models to be complete in terms of allparts defined in that LoD (likely to be a highburden for data collectors)? Are all aspects,which might be relevant for users, e.g. geom-etry, semantics, and appearance reflected bythe concept? How flexible and vague are thedefinitions? Vaguely defined concepts widenthe scope of models and alleviate data collec-tion. Strictly defined concepts are more reli-able for users. Is it feasible to transform cur-rent models into the new concept and what arethe transformation costs? However, the evalu-ation procedure can not formally be defined,since we consider concepts and their rules andspecifications. The evaluation whether or towhich degree a concept fulfils a specific cri-terion is performed by discursively inspectingthe specifications of the concepts.

Four current and relevant proposals fromacademia and the commercial sector as wellas the CityGML 2.0 concept are discussedand compared in relation to these criteria.The degree to which the criteria are fulfilledis very heterogeneous among the approaches.The Nagel approach is very flexible withoutany constraints, but allows for models withfew properties users can rely on. The Ben-ner approach indicates the semantic contentof models in a reliable way, and is nearly asflexible as the Nagel approach. Far more pre-cise and mandatory is the Biljecki approach,which, however, has some of the restrictionsof the current CityGML 2.0 model. The costsfor transforming CityGML 2.0 models to thisconcept are high.


iwasZcZuk, d. & stilla, u., 2010: A Concept for theAssignment of Textures to Partially occludedFaces of 3D City Models stored in CityGML. –kolBe, t.h. et al. (eds.): 5th ISPRS International3D GeoInfo Conference, Germany. – Interna-tional Archives of the Photogrammetry, RemoteSensing and Spatial Information Sciences XXX-VIII-4/W15: 57–62, Berlin.

knapp, s. & cooRs, v., 2008: The use of eParticipa-tion in public participation: The VEPs example.– cooRs, v. et al. (eds.): Urban and regional datamanagement. – Urban Data Management Socie-ty Symposium 2007 (UDMS Annual 2007),BALKEMA-proceedings and monographs inengineering, water and earth sciences: 93–104,Taylor & Francis, London, UK.

köninGeR, a. & BaRtel, s., 1998: 3D-GIS for Ur-ban Purposes. – Geoinformatica 2 (2): 79–103.

löwneR, m.o., BenneR, j., GRöGeR, G. & häFele,k.h., 2013: New Concepts for Structuring 3DCity models – an extended Level of Detail con-cept for CityGML Buildings. – muRGante, B. etal. (eds.): 13th International Conference onComputational Science and Its Applications,Part III, LNCS 7973: 466–480, Springer, Ber-lin.

löwneR, m.o., BenneR, j. & GRöGeR, G., 2015: Ak-tuelle Trends in der Entwicklung von City-GML3.0. – seyFeRt, e. et al. (eds.): Geoinforma-tionen öffnen das Tor zur Welt, 34. Wissen-schaftlich-Technische Jahrestagung der DGPF,Tagungsband 23, Hamburg.

naGel, c., 2014: Proposal for a revision of theCityGML LOD concept. – Presentation at the5th Meeting of OGC Working Package 3 for therevision of the LoD concept for CityGML 3.0,20. 10. 2014, Bonn; https://github.com/opengeospatial/CityGML-3.0/blob/master/WP%2003%20Resources/Meetings/1st/WP03_2014_07_09_Nagel_Proposal_for_a_Revision-of-the-LOD-Concept.pdf (1.11.2015).

Quinn, j.a., smaRt, p.d. & jones, c.B., 2009: 3Dcity registration and enrichment. – ISPRS-Workshop on quality, scale and analysis aspectsof city models. – International Archives of thePhotogrammetry, Remote Sensing and SpatialInformation Sciences XXXVIII-2/W11, Lund,Sweden.

Randt, B., Bildstein, F. & kolBe, t.h., 2007: Useof Virtual 3D Landscapes for Emergency DriverTraining. – 2007 IMAGE Conference, Scotts-dale, AZ, USA.

Zlatanova, s. & li, j. (eds.), 2008: Geospatial in-formation technology for emergency response.– ISPRS book series, Taylor & Francis, London,UK.

dalla costa, s., Roccatello, e. & RumoR, m.,2011: A CityGML 3D Geodatabase for BuildingsEnergy Efficiency. – International Archives ofthe Photogrammetry, Remote Sensing and Spa-tial Information Sciences XXXVIII-4/C21:19–24.

domínGueZ, B., GaRcía, Á. & Feito, F.R., 2011: Se-mantic and topological representation of build-ing indoors: an overview. – The Joint ISPRSWorkshop on 3D City Modelling & Applicationsand the 6th 3D GeoInfo Conference, Wuhan,China.

Fan, h., menG, l. & jahnke, m., 2009: Generaliza-tion of 3D Buildings Modelled by CityGML. –sesteR, m. et al. (eds.): Advances in GIScience.– 12th AGILE Conference, Lecture Notes inGeoinformation and Cartography: 387–405,Springer, Berlin, Heidelberg.

Ghassoun, y., löwneR, m.o. & weBeR, s., 2015:Exploring the Benefits of 3D city Models in theField of Urban Particles Distribution Modelling– a Comparison of Model Results. – BReuniG, m.et al. (eds.): 3D Geoinformation Science, the Se-lected Papers of the 3D GeoInfo 2014. – LectureNotes in Geoinformation and Cartography: 193–205, Springer.

GötZelmann, t., GueRke, R., BRenneR, c. & ses-teR, m., 2009: Terrain-Dependent Aggregationof 3D City Models. – ISPRS-Workshop on qual-ity, scale and analysis aspects of city models. –International Archives of the Photogrammetry,Remote Sensing and Spatial Information Sci-ences XXXVIII-2/W11, Lund, Sweden.

GRöGeR, G., kolBe, t.h., naGel, c. & häFele, k.h.(eds.), 2012: OGC City Geography Markup Lan-guage (CityGML), Encoding Standard, Version2.0.0., Open Geospatial Consortium. – OGCDoc. No. 12-019, Open Geospatial Consortium,2012.

GRöGeR, G. & plümeR, l., 2012: CityGML – In-teroperable semantic 3D City Models. – ISPRSJournal of Photogrammetry and Remote Sensing71: 12–33.

GueRke, R., BRenneR, c. & sesteR, m., 2009: Gen-eralization of 3D City Models as a Service. –ISPRS-Workshop on quality, scale and analysisaspects of city models, International Archives ofthe Photogrammetry, Remote Sensing and Spa-tial Information Sciences, XXXVIII-2/W11,Lund, Sweden.

haGedoRn, B., tRapp, m., GlandeR, t. & döllneR,j., 2009: Towards an Indoor Level-of-DetailModel for Route Visualization. – MDM ‘09,Tenth International Conference on Mobile DataManagement: Systems, Services and Middle-ware: 692–697, IEEE Computer Society, Piscat-away, NJ, USA.


Priv.-Doz. Dr. rer. nat. GeRhaRd GRöGeR, CPA Soft-ware GmbH, Auf dem Seidenberg 3a, 53721 Sieg-burg, Germany, e-mail: [email protected]

Manuskript eingereicht: August 2014Angenommen: Oktober 2015

Addresses of the Authors

Prof. Dr.-Ing. maRc-oliveR löwneR, Institute ofGeodesy and Photogrammetry, Technische Univer-sität Braunschweig, Pockelsstraße 3, 38106 Braun-schweig, Germany, e-mail: [email protected]

evaluation criteria for recent lod proposals for city gml buildings · introduced a level of detail...

Documents