Real-world objects are represented by geographic features according to the definition in ISO 19109. Geographic features of the same type (e.g., buildings, roads) are modeled by corresponding feature types that are represented as classes in the Conceptual Model (CM). The objects within a 3D city model are instances of the different feature types.

In order to distinguish and reference individual objects, each object has unique identifiers. In the CityGML 3.0 CM, each geographic feature has the mandatory _featureID_ and an optional _identifier_ property. The _featureID_ is used to distinguish all objects and possible multiple versions of the same real-world object. The _identifier_ is identical for all versions of the same real-world object and can be used to reference specific objects independent from their actual object version. The _featureID_ is unique within the same CityGML dataset, but it is generally recommended to use globally unique identifiers like UUID values or identifiers maintained by an organization such as a mapping agency. Providing globally unique and stable identifiers for the _identifier_ attribute is recommended. This means these identifiers should remain stable over the lifetime of the real-world object.

CityGML feature types typically have a number of spatial and non-spatial properties (also called attributes) as well as relationships with other feature or object types. Note that a single CityGML object can have different spatial representations at the same time. For example, different geometry objects representing the feature's geometry in different levels of detail or as different spatial abstractions.

Many attributes have simple, scalar values like a number or a character string. However, some attributes are complex. They do not just have a single property value. In CityGML the following types of complex attributes occur.

In CityGML, the specific feature types like __Building__, __Tunnel__, or _WaterBody_ are defined as subclasses of more general higher-level classes. Hence, feature types build a hierarchy along specialization / generalization relationships where more specialized feature types inherit the properties and relationships of all their superclasses along the entire generalization path to the topmost feature type __AnyFeature__.

Note: A superclass is the class from which subclasses can be created.

In CityGML, objects can be related to each other and different types of relations are distinguished. First of all, complex objects like buildings or transportation objects typically consist of parts. These parts are individual features of their own, and can even be further decomposed. Therefore, CityGML objects can form aggregation hierarchies. Some feature types are marked in the conceptual model with the stereotype _«TopLevelFeatureType»_. These constitute the main objects of a city model and are typically the root of an aggregation hierarchy. Only top-level features are allowed as direct members of a city model. The information about which feature types belong to the top level is required for software packages that want to filter imports, exports, and visualizations according to the general type of a city object (e.g., only show buildings, solitary vegetation objects, and roads). CityGML Application Domain Extensions should also make use of this concept, such that software tools can learn from inspecting their conceptual schema what are the main, i.e., the top-level, feature types of the extension.

Some relations in CityGML are qualified by additional parameters, typically to further specify the type of relationship. For example, a relationship can be qualified with a URI pointing to a definition of the respective relation type in an Ontology. Qualified relationships are used in CityGML, among others, for:

The meanings of all elements defined in the CityGML conceptual model are normatively specified in the data dictionary in <<data-dictionary-section>>.

Spatial properties of all CityGML feature types are represented using the geometry classes defined in ISO 19107. Spatial representations can have 0-, 1-, 2-, or 3-dimensional extents depending on the respective feature type and Levels of Detail (LOD). The LOD concept is discussed in <<ug-levels-of-detail-section>> and <<ug-geometry-lod-section>>. With only a few exceptions, all geometries must use 3D coordinate values. Besides primitive geometries like single points, curves, surfaces, and solids, CityGML makes use of different kinds of aggregations of geometries like spatial aggregates (_MultiPoint_, _MultiCurve_, _MultiSurface_, _MultiSolid_) and composites (_CompositeCurve_, _CompositeSurface_, _CompositeSolid_). Volumetric shapes are represented in ISO 19107 according to the so-called _Boundary Representation_ (B-Rep). For further explanation see <<Foley2002>>.

The CityGML Conceptual Model does not put any restriction on the usage of specific geometry types as defined in ISO 19107. For example, 3D surfaces could be represented in a dataset using 3D polygons or 3D meshes such as triangulated irregular networks (TINS) or by non-uniform rational B-spline surfaces (NURBS). However, an encoding may restrict the usage of geometry types. For example, curved lines like B-splines or clothoids, or curved surfaces like NURBS could be disallowed by explicitly defining _null encodings_ for these concepts in the encoding specification (c.f. <<ug-modularization-section>> above).

Note that the conceptual schema of ISO 19107 allows composite geometries to be defined by a recursive aggregation for every primitive type of the corresponding dimension. This aggregation schema allows the definition of nested aggregations (hierarchy of components). For example, a building geometry (_CompositeSolid_) can be composed of the house geometry (_CompositeSolid_) and the garage geometry (_Solid_), while the house’s geometry is further decomposed into the roof geometry (_Solid_) and the geometry of the house body (_Solid_). This is illustrated in <<figure-recursiveaggregation>>.

While the CityGML Conceptual Model does not employ the topology classes from ISO 19107, topological relations between geometries can be established by sharing geometries (typically parts of the boundary) between different geometric objects. One part of real-world space can be represented only once by a geometry object and is referenced by all features or more complex geometries which are defined or bounded by this geometry object. Thus redundancy can be avoided and explicit topological relations between parts are maintained.

Basically, there are three cases for sharing geometries.

In CityGML, objects of equal shape like trees and other vegetation objects, traffic lights and traffic signs can be represented as prototypes which are instantiated multiple times at different locations (see <<figure-prototypicshapes>>). The geometry of prototypes is defined in local coordinate systems. Every instance is represented by a reference to the prototype, a base point in the world coordinate reference system (CRS) and a transformation matrix that facilitates scaling, rotation, and translation of the prototype. The principle is adopted from the concept of scene graphs used in computer graphics standards. Since the ISO 19107 geometry model does not provide support for scene graph concepts, the CityGML class ImplicitGeometry has been introduced (for further description see <<ug-geometry-lod-section>>). The prototype geometry can be represented using ISO 19107 geometry objects or by referencing an external file containing the geometry in another data format.

In addition to the spatial representations defined in the _Core_ module, the geometry of physical spaces and of thematic surfaces can now also be provided by 3D point clouds using MultiPoint geometry. This allows, for example, spatially representing the building hull, a room within a building or a single wall surface just by a point cloud. All thematic feature types including transportation objects, vegetation, city furniture, etc. can also be spatially represented by point clouds. In this way, the ClearanceSpace of a road or railway could, for instance, be modeled directly from the result of a mobile laser scanning campaign. Point clouds can either be included in a CityGML dataset or just reference an external file of some common types such as LAS or LAZ.

CityGML is about 3D city and landscape models. This means that nearly all geometries use 3D coordinates, where each single point and also the points defining the boundaries of surfaces and solids have three coordinate values (x,y,z) each. Coordinates always have to be given with respect to a coordinate reference system (CRS) that relates them unambiguously with a specific position on the Earth. In contrast to CAD or BIM, each 3D point is absolutely georeferenced, which makes CityGML especially suitable to represent geographically large extended structures like airports, railways, bridges, dams, where the Earth curvature has a significant effect on the object’s geometry (for further explanations see <<Kaden2017>>).

In most CRS, the (x,y) coordinates refer to the horizontal position of a point on the Earth’s surface. The z coordinate typically refers to the vertical height over (or under) the reference surface. Note that depending on the chosen CRS, x and y may be given as angular values like latitude and longitude or as distance values in meters or feet. According to ISO 19111, numerous 3D CRS can be used. This includes global as well as national reference systems using geocentric, geodetic, or projected coordinate systems.

In the CityGML 3.0 Conceptual Model, a clear semantic distinction of spatial features is introduced by mapping all city objects onto the semantic concepts of spaces and space boundaries. A _Space_ is an entity of volumetric extent in the real world. Buildings, water bodies, trees, rooms, and traffic spaces are examples for such entities with volumetric extent. A _Space Boundary_ is an entity with areal extent in the real world. Space Boundaries delimit and connect Spaces. Examples are the wall surfaces and roof surfaces that bound a building, the water surface as boundary between the water body and air, the road surface as boundary between the ground and the traffic space, or the digital terrain model representing the space boundary between the over- and underground space.

To obtain a more precise definition of spaces, they are further subdivided into physical spaces and logical spaces. Physical spaces are spaces that are fully or partially bounded by physical objects. Buildings and rooms, for instance, are physical spaces as they are bounded by walls and slabs. Traffic spaces of roads are physical spaces as they are bounded by road surfaces against the ground. Logical spaces, in contrast, are spaces that are not necessarily bounded by physical objects, but are defined according to thematic considerations. Depending on the application, logical spaces can also be bounded by non-physical, i.e., virtual boundaries, and they can represent aggregations of physical spaces. A building unit, for instance, is a logical space as it aggregates specific rooms to flats, the rooms being the physical spaces that are bounded by wall surfaces, whereas the aggregation as a whole is being delimited by a virtual boundary. Other examples are city districts which are bounded by virtual vertically extruded administrative boundaries, public spaces vs. Security zones in airports, or city zones with specific regulations stemming from urban planning. The definition of physical and logical spaces and of corresponding physical and virtual boundaries is in line with the discussion in [<<Smith2000>>] on the difference between bona fide and fiat boundaries to bound objects. Bona fide boundaries are physical boundaries; they correspond to the physical boundaries of physical spaces in the CityGML 3.0 CM. In contrast, fiat boundaries are man-made boundaries. They are equivalent to the virtual boundaries of logical spaces.

Physical spaces, in turn, are further classified into occupied spaces and unoccupied spaces. Occupied spaces represent physical volumetric objects that occupy space in the urban environment. Examples for occupied spaces are buildings, bridges, trees, city furniture, and water bodies. Occupying space means that some space is blocked by these volumetric objects. For instance, the space blocked by the building in <<figure-occupiedandunoccupiedspaces>> cannot be used any more for driving through this space or placing a tree on that space. In contrast, unoccupied spaces represent physical volumetric entities that do not occupy space in the urban environment, i.e., no space is blocked by these volumetric objects. Examples for unoccupied spaces are building rooms and traffic spaces. There is a risk of misunderstanding the term OccupiedSpace. However, we decided to use the term anyway, as it is established in the field of robotics for over three decades [<<Elfes1989>>]. The navigation of mobile robots makes use of a so-called occupancy map that marks areas that are occupied by matter and, thus, are not navigable for robots.

The new space concept offers several advantages.

Semantic objects in CityGML are often composed of parts, i.e., they form multi-level aggregation hierarchies. This also holds for semantic objects representing occupied and unoccupied spaces. In general, two types of compositions can be distinguished.

The classification of feature types into OccupiedSpace and UnoccupiedSpace also defines the semantics of the geometries attached to the respective features. For OccupiedSpaces, the attached geometries describe volumes that are (mostly) physically occupied. For UnoccupiedSpaces, the attached geometries describe (or bound) volumes that are (mostly) physically unoccupied. This also has an impact on the required orientation of the surface normal (at point _P_ this is a vector perpendicular to the tangent plane of the surface at __P__) for attached thematic surfaces. For OccupiedSpaces, the normal vectors of thematic surfaces must point in the same direction as the surfaces of the outer shell of the volume. For UnoccupiedSpaces, the normal vectors of thematic surfaces must point in the opposite direction as the surfaces of the outer shell of the volume. This means that from the perspective of an observer of a city scene, the surface normals must always be directed towards the observer. In the case of OccupiedSpaces (e.g., Buildings, Furniture), the observer must be located outside the OccupiedSpace for the surface normals being directed towards the observer; whereas in the case of UnoccupiedSpaces (e.g., Rooms, Roads), the observer is typically inside the UnoccupiedSpace.

The CityGML Conceptual Model differentiates four consecutive Levels of Detail (LOD 0-3), where objects become more detailed with increasing LOD with respect to their geometry. CityGML datasets can - but do not have to - contain multiple geometries for each object in different LODs simultaneously. The LOD concept facilitates multi-scale modeling; i.e., having varying degrees of spatial abstractions that are appropriate for different applications or visualizations.

The classification of real-world objects into spaces and space boundaries is solely based on the semantics of these objects and not on their used geometry type, as the CityGML 3.0 CM allows various geometrical representations for objects. A building, for instance, can be spatially represented by a 3D solid (e.g., in LOD1), but at the same time, the real-world geometry can also be abstracted by a single point, footprint or roof print (LOD0), or by a 3D mesh (LOD3). The outer shell of the building may also be semantically decomposed into wall, roof, and ground surfaces. <<figure-buildinglods>> shows different representations of the same real-world building object in different geometric LODs (and appearances).

The biggest changes between CityGML 3.0 and earlier versions are as follows.

_Spaces_ and all its subclasses like _Building_, _Room_, and _TrafficSpace_ can now be spatially represented by single points in LOD0, multi-surfaces in LOD0/2/3, solids in LOD1/2/3, and multi-curves in LOD2/3. _Space Boundaries_ and all its subclasses such as _WallSurface_, _LandUse_, or _Relief_ can now be represented by multi-surfaces in LOD0/2/3 and as multi-curves in LOD2/3. See <<geometry-lod-section>> for further details on the different Levels of Detail.

Objects, which are not spatially represented by a volumetric geometry, must be virtually closed in order to compute their volume (e.g., pedestrian underpasses or airplane hangars). They can be sealed using a specific type of space boundary called a ClosureSurface. These are virtual surfaces. They are used when a closed surface is needed to compute volumes or perform similar 3D operations. Since they do not actually exist, they are neglected when they are not needed or not appropriate. For example, ClosureSurfaces would not be used in visualizations.

The concept of ClosureSurface can also be employed to model the entrances of subsurface objects. Those objects like tunnels or pedestrian underpasses have to be modeled as closed solids in order to compute their volume. An example would be for use in flood simulations. The entrances to subsurface objects also have to be sealed to avoid holes in the digital terrain model (see <<figure-closuresurfaces>>). However, in close-range visualizations the entrance should be treated as open. Thus, closure surfaces are an adequate way to model those entrances.

An important issue in city modeling is the integration of 3D objects and the terrain. Problems arise if 3D objects float over or sink into the terrain. This is particularly the case when terrains and 3D objects in different LODs are combined, when the terrain and 3D models are updated independently from each other, or when they come from different data providers [<<Kolbe2003>>]. To overcome this problem, the TerrainIntersectionCurve (TIC) of a 3D object is introduced. These curves denote the exact position where the terrain touches the 3D object (see <<figure-terrainintersectioncurves>>). TICs can be applied to all CityGML feature types that are derived from AbstractPhysicalSpace such as buildings, bridges, tunnels, but also city furniture, vegetation, and generic city objects.

If, for example, a building has a courtyard, the TIC consists of two closed rings: One ring representing the courtyard boundary, and one which describes the building's outer boundary. This information can be used to integrate the building and a terrain by ‘pulling up’ or ‘pulling down’ the surrounding terrain to fit the TerrainIntersectionCurve. The digital terrain model (DTM) may be locally warped to fit the TIC. By this means, the TIC also ensures the correct positioning of textures or the matching of object textures with the DTM. Since the intersection with the terrain may differ depending on the LOD, a 3D object may have different TerrainIntersectionCurves for all LODs.

An important design principle for CityGML is the coherent modeling of semantic objects and their spatial representations. At the semantic level, real-world entities are represented by features, such as buildings, walls, windows, or rooms. The description also includes attributes, relations and aggregation hierarchies (part-whole-relations) between features. Thus the part-of-relationship between features can be derived at the semantic level only, without considering geometry. However, at the spatial level, geometry objects are assigned to features representing their spatial location, shape, and extent. So the model consists of two hierarchies: The semantic and the geometrical in which the corresponding objects are linked by relationships (cf. <<Stadler2007>>). The advantage of this approach is that it can be navigated in both hierarchies and between both hierarchies arbitrarily, for answering thematic and/or geometrical queries or performing analyses.

If both hierarchies exist for a specific object, they must be coherent (i.e., it must be ensured that they match and fit together). For example, if a building is semantically decomposed into wall surfaces, roof surfaces and so forth, the polygons representing these thematic surfaces (in a specific LOD) must be part of the solid geometry representing the entire building (for the same LOD).

As described in <<ug-semantic-modeling-section>>, the bitemporal timestamps of all CityGML feature types allow representing the evolution of the real city and its model over time. The new _Versioning_ module extends this concept by the possibility of representing multiple, concurrent versions of the city model. For that purpose, the module defines two new feature types: 1) _Version_, which can be used to explicitly define named states of the 3D city model and denote all the specific versions of objects belonging to such states. 2) _VersionTransition_, which allows to explicitly link different versions of the 3D city model by describing the reason of change and the modifications applied. Details on the versioning concept are given in [<<Chaturvedi2015>>].

This approach not only facilitates the explicit representation of different city model versions, but also allows distinguishing and referring to different versions of city objects in an interoperable exchange format. All object versions could be stored and exchanged within a single dataset. Software systems could use such a dataset to visualize and work with the different versions simultaneously. The conceptual model also takes into account the management of multiple histories or multiple interpretations of the past of a city, which is required when looking at historical city developments and for archaeological applications. In addition, the Versioning module supports collaborative work. All functionality to represent a tree of workspaces as version control systems like _git_ or _SVN_ is provided. The Versioning module handles versions and version transitions as feature types, which allows the version management to be completely handled using the standard OGC Web Feature Service [<<Vretanos2010>>]. No extension of the OGC Web Feature Service standard is required to manage the versioning of city models.

The new Dynamizer module improves the usability of CityGML for different kinds of simulations as well as to facilitate the integration of devices from the Internet-of-Things (IoT) like sensors with 3D city models. Both, simulations and sensors provide dynamic variations of some measured or simulated properties such as the electricity consumption of a building or the traffic density within a road segment. The variations of the value are typically represented using time-series data. The data sources of the time-series data could be either sensor observations (e.g., from a smart meter), pre-recorded load profiles (e.g., from an energy company), or the results of some simulation run.