1. Introduction
Three-dimensional city models have become an important resource for planning, development, and policymaking in urban areas [
1,
2,
3,
4,
5]. A 3D city model is a digital model of an urban environment with a three-dimensional geometry of urban structures, as well as related objects belonging to urban areas [
6]. Applications using 3D city models have increased in their scope and complexity [
7], spanning from the analysis of electromagnetic propagation for telecommunications through environmental simulations analysing irradiation distribution [
8,
9] and noise propagation [
10] to virtual or augmented reality applications [
11,
12]. This proliferation of applications is, in turn, driving an increasing demand for the creation and maintenance of reliable 3D city models. A standard approach to creating city models at a large scale automatically or semi-automatically is to apply stereo vision on aerial or satellite remote sensing imagery [
3]. This, however, can be an expensive and/or time/labour-consuming process, particularly if high levels of accuracy in model outputs are required [
13]. As a result, large-scale 3D city models are mostly available in countries with developed economies and/or those with national mapping agencies, while countries, including many that are transitioning their economies (and where this information is perhaps of most value), do not have the resources available to produce them [
4]. An approach underpinned by suitable open data could fill this gap in capability.
Three-dimensional city models are characterized by their level of detail (LOD) [
14]. The CityGML standard defines five levels of details (LOD) from LOD0 to LOD4. The coarsest level, LOD0, represents the lowest level of geometry as a 2.5D DTM (digital terrain model) with building footprints or roof edge polygons. It is used for regional and landscape applications. LOD1, is well-known as a block model. In LOD1, the building height would be extruded with flat roofs. It is used for city and region coverage. In LOD2, buildings have differentiated roof structures and thematically differentiated boundary surfaces based on LOD1 models. It is applicable for city districts. LOD3 will add specific roof and wall structure details, such as doors and windows, to LOD2 models and it denotes architectural models. This one is widely used for landmarks. LOD4 gives interior structures, like doors, stairs, etc., within the buildings [
15,
16,
17]. An increase in the LOD of a model enables more applications, but it also increases data demands and their processing involves higher computational costs [
14,
18,
19].
Many applications of 3D city models require only low level of details—LOD1 (e.g., vulnerability models, disaster mitigation, climate change and energy models). Here, we investigate the production of spatially reliable and globally replicable 3D city models using open-licensed data in order to support that category of user. This research forms part of a wider project, ‘Sustaining Urban Habitats: An interdisciplinary approach’, which aimed to explore ways of combining environmental and economic modelling with social and cultural ethnographic work. The focus of the project was on two contrasting cities: a growth city in China (Shanghai) and a relatively stable city in Europe (Nottingham). During the implementation of the wider project, the dearth of accurate 3D models for many cities globally, including Shanghai, was observed. The project had very little budget to acquire data and thus raised the challenge of how to produce a 3D city model from open data. This study did not aim to alternate commercial 3D city models with 3D city models from open data, instead, it focused on presenting a method that produces 3D city models from open data (only) to serve those regions that cannot acquire commercial data. Given that open datasets are usually characterized by low resolution, we present a method capable of producing the desired LOD 1 city model for anywhere.
Possibilities of extracting building heights from open digital surface models (DSM) and digital elevation models (DEM) have previously been attempted [
20,
21]. These include extraction of building heights from the Shuttle Radar Topographic Mission (SRTM), the Advanced Spaceborne Thermal Emission and Reflection Radiometer Digital Elevation Model (ASTER DEM), Advanced Land Observing Satellite ALOS World 3D (AW3D) DSM, and TerraSAR-X add-on for digital elevation measurements (TanDEM-X). However, using DSMs alone cannot provide exact building heights or shapes. Rather it will result in more generalized individual building heights and distorted shapes due to issues of mixed pixels [
21]. Using 2D data of building footprints along with high resolution DSMs can be a possible solution to extract individual building heights without distorting the building shapes. The approach is predicated on the availability of open-source 2D spatial datasets, such as OpenStreetMap (OSM), albeit with varying degrees of completeness and reliability, to provide building footprint geometries. However, the third dimension is poorly represented in these datasets; less than 2.5% of the nodes in the OSM database carry an elevation attribute [
22,
23]. The recently available satellite-derived elevation datasets provide an opportunity for data fusion by incorporating the elevation data with open-licensed 2D building data to generate 3D models. Indeed Bagheri et al. generated LOD1 height values using multisensor and multimodal DEM fusion techniques—TanDEM-X DEM and Cartosat-1 DEM data were joined with OpenStreetMap building footprints [
24]. This study confirmed that simple, prismatic building models can be reconstructed by combining OpenStreetMap building footprints with remote sensing-derived geodata. However, the assumption of a flat terrain at a constant height restricts globally applicability of this approach. Furthermore, Cartosat-1 data are not currently global in availability. Required, therefore, is a methodology that considers the terrain underlying the urban area of interest and uses datasets that are available worldwide.
In this paper, we used open DSM data as a foundation dataset and utility in a globally replicable methodology to generate 3D city models. Recently available elevation datasets such as the AW3D DSM (with a horizontal spatial resolution of approximately 30 m) by the Japanese Aerospace Exploration Agency (JAXA) have an open license (a higher resolution (approx. 5 m) DSM is also produced, but only as a commercial product [
25]). Other common elevation-rich datasets include the ASTER DEM and that from the SRTM. Although these provide mainly terrain (a digital surface model includes all the natural and built features on the earth’s surface, whereas a digital terrain model is simply an elevation surface representing the bare earth referenced to a common vertical datum [
26]) elevation values that are freely available under permissive data licenses [
27]. We present a methodology that uses open data of 2D building footprints, along with DSM and DTM datasets, to generate 3D buildings in two geographically and morphologically diverse cities, namely the Huangpu district in Shanghai, China, which has a relatively flat topography, and Nottingham, United Kingdom, which has a more undulating terrain. Shanghai and Nottingham are inherently different from each other, not only in terms of physiography but also in terms of level of urbanization. While Shanghai is a rapidly urbanizing city, Nottingham is stabilized and saturated. Hence, these two cities provide end members to transfer the methods globally.
A secondary objective was to consider scenarios of data availability that could improve the overall accuracy of the open source 3D building model generated (which we call a foundation model). Here, we exploited that often higher resolution elevation data are available, though not always, or never, open source, and/or of limited spatial coverage. For instance, there are a number of examples where previously proprietary LiDAR datasets are now being opened, though often these are for cities in the global North [
28], or it may be the case that projects to produce 3D city models have a limited budget. Further, here we used the ALOS DSM to generate building heights. AW3D-30 DSM is produced by resampling the 5 m ALOS DSM, resulting in accuracy reduction. Thus, it is not possible to use this low resolution DSM directly in the same way you would with a high resolution commercial dataset. From high resolution DSMs, roof heights or building heights could be easily measured. Whereas, in low resolution ALOS DSM, this is not possible. This study thus also explored the optimal approach to using the ALOS-30m DSM.
4. Discussion
Three dimensional building models form useful data inputs for many analytical tasks, but their creation typically relies upon time-consuming editing, expensive proprietary datasets, or both. Here, we present a simple method of generating 3D buildings from open data that can be applied globally. The results presented in this paper show that AW3D-30 DSM data provide more accurate results in the case of low- and medium-rise buildings, and that errors can be improved through a calibrated enhancement process. Using OSM in combination with the medium-resolution AW3D-30 DSM, a set of building footprints with height information were created and their quality ascertained. We then evaluated enhancements to height accuracy through statistical analysis of a small sample area of high-resolution data (thus limiting expense where these data are not freely available).
The approach presented can be applied by any user that has 2D building footprint data and AW3D data and terrain information (i.e., from GMTED2010). AW3D-30 is the most suitable open DSM for building height generation, in comparison with ASTER, SRTM, and TanDEM-X [
21]. However, while using AW3D-30 DSM there is a challenge of dealing with mixed pixels due to instances when buildings in the AW3D-5 digital building height range with a ground footprint of approximately 30 m or less were split into adjacent 30 m resolution pixels, each with a lower height than the original [
21]. Thus one of the important advantages of using OSM together with AW3D-30 DSM is that it helps to avoid the issues of mixed pixels and provides more accurate individual building heights and shapes. To the authors’ knowledge, this is the first attempt at combining OSM data with AW3D data to generate 3D models. We built upon previous studies that fused OSM with satellite-derived elevation data [
24], however in our study, we provided a method to generate 3D models for both flat and undulated terrain using open data, which makes it feasible to replicate globally with any kind of terrain. Our study also demonstrated ways to increase the accuracy of the generated 3D city models using a sample high resolution DSM and DTM data. Our work also demonstrated that the usage of high resolution DTM for ground elevation extraction can result in higher accuracy of building height values. This paper recommends the use of high-resolution digital terrain models (DTMs) wherever possible and in the absence of the same, GMTED 2010 data shall be used as ground elevation for undulating terrain and can use mean elevation value as ground elevation for flat terrains like Shanghai. The study also highlighted the need for a geospatial community to generate a global open access high-resolution DTM. The need for generating global high-resolution DEM in open access was also highlighted by Schumman and Bates, 2018 [
45]. There are also initiatives like ‘Open Topography’, which facilitates community access to high-resolution topographic data [
46]. These high-resolution data (metre to sub-metre scale) are derived from LiDAR and other technologies. This free access to high-accuracy terrain data further sheds light to the extensive potential of generating highly accurate 3D city models using open data.
The accuracy assessment of the two distinctive cities shows that the 3D model developed using this methodology will have higher accuracy in cities like Nottingham, where majority of the buildings are of a low rise and where growth is relatively saturated. Whereas in cities like Shanghai, where the percentage of very tall buildings is high, the accuracy will be reduced. In our study, for Nottingham, we could generate 27.7% of buildings with +/−1 m accuracy, 51.45% with +/−2 m accuracy, and 84.27% with +/−5 m. In Shanghai, the accuracy was much lower than that of Nottingham—the percentage of buildings within the accuracy levels of +/−1 m, +/−2 m, and +/−5 m were 17.66, 32.96, and 62.26 respectively. The accuracy reduction in Shanghai is explained by the increased number of tall buildings compared to city of Nottingham. It is significant to observe that the AW3D-30 DSM provides more accurate results for low- and medium-rise buildings, but exhibits relatively large errors in height for very tall buildings. This result echoes findings of Alganci et al. [
36], but contradicts the finding of Misra et al. [
21]. Accuracy assessments of different DSMs by Alganci et al. [
36] revealed that the AW3D-30 DSM performed worse for high-rise buildings compared to SPOT DSM and PHR DSM, and that AW3D-30 DSM has a high accuracy level in residential areas. In contrast, Misra et al. [
21] reported that AW3D-30 is most suitable for observing buildings taller than 9 m in height. However, this is in comparison with ASTER- and SRTM-based building heights, which are less suitable for extracting building height variation [
21]. In our study we considered all buildings with height above 2 m and results showed good accuracy. Hence, using the presented method, even without any accuracy enhancement, will provide better accuracy in cities with low- and medium-rise buildings compared to cities with high-rise buildings. Once our accuracy enhancement method was applied (by way of a sample of high resolution elevation data), this improved the reliability of 3D models from open data (in Nottingham we demonstrated enhancements in the percentages of buildings within an accuracy level of +/−1 m from 27.7% to 32.81%, and for accuracy level of +/−2 m from 51.45% to 57.43%). However, this method is limited to the containment of only systematic errors; random errors are not accounted for.
Using OSM in combination with AW3D-30 DSM data has substantial potential for future scientific research due to the former’s ever-growing size and the latter’s global coverage [
24,
34,
35,
36,
47]. Studies have reported that there has been a considerable increase in OSM building data in recent years. For example, from 2012 to 2017 alone there has been a 20 times increase in OSM building data in China [
48]. Effective derivation of elevation values for OSM data will likely extend its utility [
22]. However, the absence of a global completeness assessment may hamper the use of OSM for urban planning and development, unless it is resolved [
49]. One of the major concerns in using OSM data is the quality. Most OSM data are provided by nonprofessionals and hence both the coverage and the quality of the data are questionable [
50,
51,
52]. Despite this disadvantage, OSM is a good source of 2D building data, especially where free 2D building data are unavailable, as in China, where authorized building data are not freely available [
48]. Studies have also revealed that the rate at which OSM is receiving contributions from users has been constantly increasing and is continuing to grow; complemented by collaborative mapping efforts amongst the OSM community to check and improve the quality of contributions [
53].
AW3D-30 DSM also has considerable future potential, particularly for low- and middle-income countries, given its global coverage and open license. The JAXA released its first version AW3D-30 m DSM with a horizontal resolution of approx. 30 m mesh, free of charge in May 2015. This dataset was generated from the DSM dataset (5 m mesh version) of the precise global digital 3D map ALOS World 3D” (AW3D), which was the world’s first and the most precise 3D map covering all global land scales with a 5 m mesh [
37]. Although the AW3D-30 DSM had a 30 m grid spacing, it could be deduced that this was due to the acquisition of strong signals from the original 5 m DSM, which was produced from the 2.5 m images [
36]. In March 2017, version 1.1 was released, filling the void height values with existing DEMs in cloud and snow pixels between 60° north and 60° south. In April 2018, AW3D was upgraded to version two [
54]. Continuous enhancements of AW3D-30 DSM are expected, improving its future utility.
Thus, one of the great advantages of our methodology is that 3D models can be generated from any 2D building data in combination with any DSM, which means not just using OSM and AW3D-30 DSM data. Wherever any 2D building data are available, the user will be able to generate the building elevation in combination with DSM data. Currently, only AW3D provides free DSM data. Even though ASTER DEM and SRTM provide elevation data, they are not surface models, hence those datasets are not usable to generate 3D building elevation. However, in the future there will likely be higher resolution DSMs. LiDAR DSM and ICESat-2 data are examples. Many countries are already providing accurate LiDAR DSM data. For example, LiDAR DSM data are already available for about 70% of England from the UK Environmental Agency [
41]. ICESat-2 (ICE, CLOUD, and Land Elevation Satellite) is an ambitious mission from NASA, which will provide a global distribution of geodetic measurements of both the terrain surface and relative canopy heights and it will also survey urban areas [
55]. Further, Global Ecosystem Dynamics Investigation (GEDI) LIDAR from NASA, with its dense track sampling and precise geolocation, forms the basis of an important dataset of ground control points to validate and calibrate global and regional DEMs and serves as a reference for surface elevation change [
56]. Thus, we hope that when more accurate DSMs become available, it will enable the user to produce more accurate 3D models with better shape descriptions of buildings, especially roof modelling, thereby generating higher LODs using the defined methodology. Knowing the nature of the terrain in the modelling area is a factor in our method. For cases with flat terrain (e.g., Shanghai), the mean ground elevation is deducted from the DSM data to obtain the building height, whereas in cases of undulated terrain (e.g., Nottingham), terrain elevation can be obtained from multiple sources, such as contour topographic data or from satellite-based sources like GMTED2010 and LiDAR DTM.
The method presented in this paper affords the development of 3D models with LOD1 for any urban setting globally. High-resolution 3D datasets with higher LODs are, of course, possible but are very expensive to produce and many applications do not require very precise height datasets. Often, a model with LOD1 data is enough. Studies shows that LOD1 models provide a relatively high information content and usability compared to their geometric detail [
57,
58]. LOD1 model is the simplest volumetric 3D city model and fundamentally considered coarse and inferior to an LOD2. However, it may be more valuable than an LOD2 model for certain scenarios, especially when a finer footprint is more useful than the acquired roof shape [
59]. Examples of such cases include: climate change and urban climate modelling, property registration, energy modelling [
60], energy demand estimation [
61,
62], shadowing simulations [
63,
64], navigation, estimation of noise pollution [
65], design of urban green spaces, crisis management, vulnerability assessments for disaster mitigation and management, simulating floods [
66], for analysing wind comfort [
67], global change assessments [
68,
69], and visualisation [
70]. Computation of the net internal area of a building is another application area of LoD1 data, useful for energy estimations, real estate valuation, and population counts [
71,
72,
73,
74].
LoD requirements are task-specific and data volume-dependent [
75,
76,
77,
78]. Should LOD1 be appropriate, the method presented here will also allow users to generate data in a cost-effective manner. Indeed, studies have attempted to assess the possibilities of 3D model generation from the OSM data used here and have already identified the huge potential of OSM for fulfilling the requirements for CityGML LoD1 [
79,
80,
81]. Further, free and open earth observation data (e.g., Landsat and Sentinel) offer great potential for large-area mapping of human settlements [
68]. As our method relies on open datasets, we hope that it will be of great use in developing and low-income countries to generate 3D data at no cost and with minimal effort. Furthermore, as this method uses freely downloadable open source datasets, it helps to save time and effort. Usually, generating 3D data is very tedious and time consuming and also requires a through lengthy process of data procurement procedures. Many applications, like hazard and risk management or crisis management, require faster results and our data generation technique will be very handy in these circumstances.
As this study intended to develop LOD1 models, we did not consider topological errors. If the 2D topological relationships between the footprints are not taken into account, the resulting 3D city models will not necessarily be topologically consistent (i.e., primitives shared by 3D buildings will be duplicated and/or intersected and overlapped building parts etc.) [
82,
83,
84]. Models with topological inaccuracies often cannot be accepted by downstream analytical applications that demand 2-manifold exterior shells [
82,
83]. However, our objective was to develop LOD1 3D city models, which do not require higher levels of accuracy. We removed all incomplete and irregular buildings after creating the 3D city model. However, we did not check for any minor errors, nor for topological accuracy. We used 2D polygons from OSM, hence, if there are any topological errors in this dataset then these errors will be reflected in our results. We recommend consideration of topology should higher accuracy in resultant models be required (i.e., LOD2+).
One of the main disadvantages observed related to using AW3D-30m data was that the accuracy limitation with high-rise buildings. As the accuracy of very tall buildings (more than 100 m) were found to be less in AW3D-30 DSM, building height data from websites like Skyscraper (which publishes the tall building information data from Council on Tall Buildings and Urban Habitat) can be used to replace the height values of these buildings, thereby increasing overall accuracy. Further, as the accuracy level reduces with the increasing percentage of tall buildings, it would be advantageous to know about the characteristics of a particular city before applying this methodology. Further, though we used AW3D-30 DSM data that were published in 2017, this dataset utilized the 2011 satellite data as base data. Hence, there could be accuracy difference for the buildings that were constructed after 2011. While using this method, it is also recommended to cross-check the results of building elevations with low height values for larger 2D footprints, as tall buildings may have a large low height podium. This can be done by visual interpretation from Google Earth satellite images. The original AW3D-30 DSM has some data-void regions and these values are filled with the values from adjacent pixels [
44]. So, some accuracy difference could arise due to this procedure. Large digitization errors or shifts in the 2D building footprints can result in the misrepresentation of height information.