Large-scale 3D building model datasets constructed from airborne LiDAR point clouds in Glasgow,UK

Abstract

Three-dimensional (3D) building models offer visual representation, interaction, analysis, and exploration of urban environment analysis. However, most cities in the United Kingdom (UK) do not have open building 3D model datasets. This study used large-scale high-density airborne LiDAR point clouds to produce 3D building models for Glasgow City. We proposed an open-source and efficient data analysis workflow that integrated a weakly supervised deep learning point cloud classification algorithm and a data-driven 3D model reconstruction method. The Glasgow 3D building model datasets include building footprints with height attributes and 3D models in the level of detail 1 (LoD1) and LoD2. The cross-reference results show that our building footprint aligned well with UK Ordnance Survey data (intersection over union of 82.67% for overlay, R = 0.93 and RMSE = 1.84 m for building height). Building models well represent outer shell features with an average RMSE = 0.54 m for the distance between point clouds and reconstructed models. This accurate 3D building model data can be used in multiple environmental applications for Glasgow, and the open-source data generation workflow can be extended to other major cities for similar applications.

Introduction

Three-dimentional (3D) city models are widely used in various urban applications such as urban planning and development, disaster management, navigation, environmental simulation, and decarbonization (Biljecki et al., 2015; Chen, 2011; Herbert and Chen, 2015; Lee and Zlatanova, 2008). Building 3D models are prominent components of the 3D city model. The City Geography Markup Language (CityGML) defines building models across five Levels of Detail (LoD), ranging from LoD0, which represents the two-dimentional (2D) building footprint, to LoD4, which includes detailed interior building structures (Gröger et al., 2012; Gröger and Plümer, 2012). Specifically, the LoD1 model represents the building footprint and associated height information with extrusion solid. The LoD2 model represents the building’s outer shell, especially roof shapes with simple geometry. Wysocki et al. (2024) reported that 15 countries offer governmental semantic 3D city models with-out cost. Previous studies published the generation methods and building 3D model data in Southeast Asian countries (Biljecki, 2020) and Netherland (Peters et al., 2022) with open footprint or cadastre data as auxiliary. However, most cities in the UK do not have openly available building 3D models to facilitate all the potential applications. Further-more, some of the existing building datasets of the UK, for example, UKBuildings (Verisk, 2021) was created by closed-source methods from bespoke commercial software, and the original data source, processing methods, and data accuracy remain unknown or commercially sensitive. Regarding the availability and reliability of building open data in the UK, while LoD1 models can be generated using building footprints and a constant height value, obtaining both footprints and height attributes with consistent geometries and temporal alignment in the desired areas remains challenging. Temporal misalignment occurs when building datasets are acquired at different times, during which buildings may be constructed, demolished, or modified, leading to inconsistencies in the representation of the same location. This is often due to variations in data sources or inaccuracies in the acquired products.

Remote sensing techniques such as multi-view optical imagery photogrammetry and Light Detection and Ranging (LiDAR) have been widely used to create large-scale 3D city models with demonstrated efficiency and satisfactory results (Cheng et al., 2011; Cho et al., 2008; Wu et al., 2017; Yu et al., 2021). Unlike multi-view optical imagery is often affected by cloud cover, shadow, glare, and uniform surfaces, LiDAR data can penetrate through foliage that allows to capture tree structures and detect buildings below canopies (Wang, 2013; Widyaningrum and Gorte, 2017). Airborne Laser Scanning (ALS) that is able to captured data in large scale with precise distance measurements becomes a superior source for 3D building modeling in city-wide scale (Ma et al., 2023; Peters et al., 2022). Constructing 3D building models from raw ALS point clouds commonly required multiple steps of processing. Firstly, building point segmentation is an essential process prior to constructing 3D building models in urban areas. Recent studies demonstrated point-based deep learning algorithms obtained high accuracy in semantic segmentation, nevertheless, most studies were carried out by using annotated benchmark point clouds instead of practical raw data that requires extensive annotation efforts (Hu et al., 2020; Landrieu and Simonovsky, 2018; Thomas et al., 2019; Qi et al., 2017). Secondly, 3D building model construction involves in extracting building footprints and heights for LoD1 model, and reconstructing rooftop features for LoD2 model (Arefi, 2009; Elaksher and Bethel, 2002; Rottensteiner, 2003). The results of each step can significantly affect the quality of final building models (Wang et al., 2023). Both model-driven methods and data-drive methods have been used in the existing literature to construct building LoD2 model. The model-driven methods select the best-fitted model from the predefined and existing building model library, but model generality is largely limited by the building model library (Henn et al., 2013; Schindler and Bauer, 2003; Zheng and Weng, 2015). The data-driven methods theoretically can reconstruct any shape as they extract roof geometries, that is, vertices, lines, or planes as primitives that are further intersected and optimized to form a 3D building model. However, most data-driven algorithms required extra input data such as geometrically highly consistent footprints as auxiliary data, or annotated rooftop points for deep learning feature extraction (Huang et al., 2022; Li et al., 2022; Peters et al., 2022). Since the UK does not have cadastre data, methods relying on footprints are difficult to implement. Most prior studies address only one of these steps in isolation. However, simply adopting existing methods for each individual step does not guarantee that their outputs can be seamlessly connected, nor does it produce an efficient or optimized workflow. There is a research gap to generate a data-driven workflow to optimally integrate advanced deep learning algorithms for 3D building models construction while considering the efficiency and feasibility of processing large-scale and real-world ALS data.

The objectives of this study include (1) constructing 3D building models at LoD1 and LoD2 levels with consistent geometries, temporal, and semantics aspects for the City of Glasgow in the UK, and (2) proposing an open-source methodology framework to generate large-scale 3D building models from raw ALS data. The Glasgow 3D city model dataset would bring opportunities for potential applications including but not limited to 3D digital twin creation, building energy efficiency analysis, flooding simulation, wind tunnel design and modeling, solar energy analysis, and ecosystem service assessment (Amirebrahimi et al., 2016; Guo et al., 2020; Johari et al., 2020; Katal et al., 2022; Prieto et al., 2019; Ridzuan et al., 2022). The proposed methodology framework would provide a guideline for future large-scale ALS data analysis from raw point cloud classification to 3D building model construction. The details of the methodologies can be found in the supplemental materials.

Data records and usage

This section describes the data products we generated in this study, including an overview of the data files, data formats, and data folder structure. It also provides information intended to assist users in accessing, processing, and reusing the data.

The Glasgow building 3D model datasets contain building footprints (LoD0) with height attributes and 3D building models at LoD1 for the entire study area of 375 km² and LoD2 models of 100 km² covering core areas of the Glasgow City. Specifically, building footprint polygons are stored in four 5 × 5 km² UK national grids in shapefiles saved in file geodatabase. The Table 1 describes building height attributes of footprint polygons. LoD1 and LoD2 building models are provided both individual buildings’ 3D mesh in OBJ format and CityJSON format. The collection of building models within 5 × 5 km² national grid are stored in multipatch shapefiles format. The attributes table of multipatch shapefile for LoD1 and LoD2 models can be found in Table 1. The structure of data record is illustrated in Figure 1, and an example of comparison of 3D building model at LoD1 and LoD2 levels is shown Figure 2. The Glasgow 3D building model dataset is available in Zenodo.

Table 1.

Description of attributes of footprint, LoD1 model, and LoD2 model data in shapefiles.

Data	Field name	Definition
Footprint	U_ID	Building footprint ID
	MIN	The height of the lowest point of the building (m)
	MAX	The height the highest point of the building (m)
	RANGE	The difference between the “MAX” and “MIN” values (m)
	MEAN	The average height of the building (m)
	MEDIAN	The median height of the building (m)
	STD	The standard deviation of BHM grids within the footprint (m)
	SUM	The sum value of BHM grids within the footprint (m)
	PCT90	The 90th percentile of BHM grids within the footprint (m)
	Ground_Z	The lowest absolute elevation of the building base (m)
	Shape_Area	The area of the building footprint in square meters (m2)
	Shape_Length	The perimeter of the building footprint (m)
LoD1	File Name	LoD1 mesh file name
	U ID	Building footprint ID
	Building Height	The average height of the roof (m)
	Footprint Area	The area of the building footprint in square meters (m2)
	Ground Z	The lowest absolute elevation of the building base (m)
LoD2	File Name	LoD2 mesh file name
	The remaining fields are the same as those in the footprint.

Figure 1.

The data structure of Glasgow 3D building models.

Figure 2.

Visualization of Glasgow 3D building models at LoD1 and LoD2 levels.

Building footprints and height attributes data can be opened and processed in commercial or open-source GIS applications (e.g., ArcGIS or QGIS) or open-source geospatial packages (e.g., Geopanda in Python or sf in R) for spatial analysis purposes. Building LoD1 and LoD2 models in multipatch shapefiles can be accessed in ArcGIS or QGIS. Building LoD1 and LoD2 models mesh file in OBJ format can be viewed in software such as CloudCompare, FME, easy3D, Sketchup, and Rhino, and can be processed in Trimesh Python library and CAD software. Building LoD1 and LoD2 models in CityJSON format can be viewed and edited in CityJSON Ninja, ArcGIS CityEngine, etc. LoD2 models cannot be built due to missing points or large data size, and some other LoD2 models contained minor error planes, a note of problematic models is provided in the database (file name: lod2 model notes.csv). Some large buildings have been divided into several parts to construct LoD2 model and the name of each part has been modified with suffixes such as “a,” “b,” “c,” etc. after the building’s U ID. Therefore, the LoD2’s building ID does not fully correspond to the U ID of footprint and LoD1. For the best result in associating individual LoD2 building to footprint and LoD1 building, the “Spatial Join” or “Selection By Location” functions are suggested to be used. The details of the data technical validation can be found in the supplemental materials.

Conclusion

This paper presents a city-wide and comprehensive building 3D model composed of in LoD0 (i.e., footprint), LoD1 and LoD2 levels for the Glasgow City in the UK. The dataset was derived from high-density ALS data, which generated accurate footprint, height attributes, LoD1 model, and LoD2 models in consistent geometries, temporal, and semantics aspects that are rarely found for a UK city. The building 3D models are provided in various formats, including mesh in OBJ, multipatch shapefile, and CityJSON, which allows the convenient application for different communities, for example, geography, planning, architecture and engineering, and digital modeling. This dataset not only provides a valuable resource for urban environmental analysis but also exemplifies the generation of city-scale building 3D models completely from raw ALS data. The designed workflow balances efficiency and accuracy to manage large-scale, high-density ALS data, covering the entire process from raw data handling and semantic segmentation to the production of building 3D models. By integrating weakly supervised deep learning for point cloud semantic segmentation, the workflow significantly reduces the manual effort required for data preprocessing. Additionally, we optimized the workflow to produce essential outputs that support subsequent processing, minimizing data redundancy and reducing manual labor in data preparation and integration. While this workflow was demonstrated using UK-specific datasets and a combination of open-source packages and commercial software, the overall methodological framework is transferable and could be adapted to other cities, provided that comparable point cloud data are available. However, in this study, validation could not be performed against ground truth data or manually generated 3D models due to the lack of available validation datasets. Future validation with ground truth data would enhance the understanding of model accuracy and help refine the modeling methods. Furthermore, green spaces are a vital component of urban environments. Future work could include generating tree canopy 3D models, extracting biophysical parameters such as tree height, volume, leaf area index, and biomass, and identifying tree species to enrich the dataset’s environmental value.

Code availability

We mainly used open-source resources such as Python, R, C++, CloudCompare (2.12.4), to process ALS data and generate datasets. Commercial software ArcGIS Pro (3.2) and FME (2022.1.1) are used for regularization and format conversion. Our codes are available on our Github page. Specifically, SQN code for point cloud classification has been implemented with Python 3.6, TensorFlow 1.11.0, CUDA 9.0, and cuDNN 7.4.1 on Ubuntu 18.04.6. The code of applying SQN with our data can be found in this Github page. The C++ program City3D used to construct the LoD2 building models has been implemented with Qt5 5.15.9, CGAL5.5.2, OpenCV 4.7.0, and Gurobi 10.0.1, Vision Studio 2022, 17.5.5 on Windows 10 and with Qt5 5.15.3, CGAL5.4.1, OpenCV 4.5,1 and Gurobi 11.0.0 on Ubuntu 22.04.4. All underlying algorithms and implementation details used in the workflow are documented in the published paper and fully accessible in the accompanying GitHub repository, ensuring that the code remains transparent, reproducible, and easily upgradable as software libraries evolve in the future.

Supplemental material

Supplemental material - Large-scale 3D building model datasets constructed from airborne LiDAR point clouds in Glasgow, UK

Supplemental material for Large-scale 3D building model datasets constructed from airborne LiDAR point clouds in Glasgow, UK by Qiaosi Li and Qunshan Zhao in Environment and Planning B: Urban Analytics and City Science.