Level of detail in 3D city modelling

PhD research


Prof. dr. Jantien Stoter


Dr. Hugo Ledoux


Started in July 2012, estimated completion in first half of 2017.


Delft University of Technology, Faculty of Architecture and the Built Environment

Financed by:

Netherlands Organisation for Scientific Research (NWO), Dutch Technology Foundation STW

Summary and aim

The aim of my PhD research is to investigate and to improve the concept of level of detail (LOD) in 3D city modelling.
It is a research that ranges from high-level conceptual frameworks (e.g. developing specifications) to low-level technical activities (e.g. developing algorithms and writing software packages).
LOD is a concept available in various disciplines from computer graphics and cartography to electrical circuit design. For GIS practitioners, it refers to the degree of the abstraction of the real-world, and it is a well-known concept that is frequently mentioned in the context of acquisition and utilisation of 3D city models.
However, the concept is burdened with several shortcomings, e.g. it is not standardised, the performance of LODs in a specific 3D spatial analysis is not benchmarked, and the consistency of multi-LOD datasets is usually poor.
I am tackling these issues through a number of topics that are described below.

What are 3D city models?

3D city models are digital representations of buildings and other relevant features in the urban environment such as roads and terrain. Their form vary, from digital surface models to semantically rich advanced representations, the latter being in focus of this research. These models have dozens of different applications, supporting processes within a variety of industries. They are essential in use-cases such as the estimation of the solar potential of roofs and the visual enhancement of navigation.


The most prominent LOD concept is the one found in the OGC standard CityGML. It defines five standard LODs that increase in their spatio-semantic coherence [animation]. While the concept is famous, it has drawbacks, and discussions for its improvement are undergoing.
This research is associated with the OGC CityGML Standards Working group, and the results will be used for the upcoming version of the standard. I have developed a proposal for a new specification of geometric LODs that is available here.

CityGML 2.0 LODs


The concept of LOD is an ambiguous term in 3D city modelling. The goal here is to formalise the concept in a consistent framework, and to derive a specification format in order to unambiguously specify a discrete LOD which can be used, for instance, in contracting the procurement of 3D city models.
Selected publication: Journal paper Formalisation of the level of detail in 3D city modelling [DOI]

Error propagation

Propagation of errors is an important topic in GIS. The main research aim is to find the relation between the uncertainty in the input data, and the uncertainty in the output of a GIS operation or use-case.
However, state of the art is limited to rasters and 2D data. My research explores this topic in 3D and in semantically enriched geo-data. The developed methodology relies on the Monte Carlo method.
Selected publication: Conference paper Error propagation in the computation of volumes in 3D city models with the Monte Carlo method [DOI].

Hyper-dimensional integration

This part of the research deals with the integration of 3D geometry and LOD in a 4D model. The benefits of this approach are, for instance, achievement of consistency, and vario-scale 3D models.
Selected publication: An introduction is available in Stoter et al. (2012): Integrating scale and space in 3D city models [DOI] and Arroyo Ohori et al. (2013): Manipulating higher dimensional spatial information [DOI].

LOD management

3D city models (e.g. CityGML files) may be derived in multi-scale representations, all in the same file. Some of the features overlap, i.e. they are the same in two or more LODs. This part of the PhD investigates is it possible to link and re-use such features for consistency (e.g. update only one LOD and propagate the changes), and storage savings.
Selected publication: Conference paper Improving the consistency of multi-LOD CityGML datasets by removing redundancy [DOI].

Context-awareness in 3D city models

During run-time, LODs in computer graphics are selected on the basis of the context, i.e. factors such as distance from the observer and importance of the object. There is no such equivalent in 3D city modelling.
This part of the research attempts at translating this concept from computer graphics to 3D GIS, by defining relevant factors and selection criteria with respect a use-case.

LOD-A = {...}
LOD-B = {...}
LOD-C = {...}
factor = [...]
value = [...]
if value[0] < factor[0]:
elif value[0] > factor[0]:
elif value[0] == factor[0]:

Geometric references

Within the same LOD, there may be a number of variants of modelling practices. For instance, the top surface of the LOD1 block model may represent the top of a building or half of the height of the roof. This research investigates the different variants, and the influence that each variant has on a 3D GIS operation. For instance, the results of the computation of volume of buildings for energy demand estimation considerably vary between variants.
Selected publication: Conference paper Height references of CityGML LOD1 buildings and their influence on applications [DOI].

Random3Dcity engine

To support some of the aforementioned topics, especially the Monte Carlo method, I have created a procedural modelling engine that generates a large number of random buildings and realises them in multiple LODs in CityGML. The engine was programmed from scratch in Python with own procedures, and it is a standalone product of this research.
If you are interested in a sample of datasets generated by the engine, visit the Random3Dcity dataset page. The code of the engine was released open-source on Github.


Solar3Dcity is another example of a software that I have created in this research. It extracts the roof surfaces in a CityGML dataset and estimates their yearly solar irradiation. It takes into account the tilt, orientation and area of each roof surface, and it computes three irradiance components: the direct, diffuse and reflected radiation.
I am using this software to benchmark the performance of LODs.
The code of the engine was released open-source on Github.


This research is being conducted in the frame of the STW project 5D Data Modelling: Full Integration of 2D/3D Space, Time and Scale Dimensions funded by the The Netherlands Organisation for Scientific Research, and which is partly funded by theĀ Ministry of Economic Affairs (Project code: 11300).