Raster-based building representations provide the statistics of buildings in regular grids. The products are delivered as rasters showing aggregated building heights or total building volumes in a grid format, typically spanning tens to hundreds of meters. The results are often estimated from diverse data sources using simple regression models. For example, World Settlement Footprint (WSF) 3D () quantifies average building heights, total volumes and building fractions at 90 m resolution. Global Human Settlement Layers (GHSL) () model global building distributions, with GHS-BUILT-H () depicting building heights and GHS-BUILT-V () depicting volumes at 250 m resolution. provided global building heights at a spatial resolution of 150 m. Generally, these products are derived either from low-resolution data sources of multiple modalities, such as Sentinel-1 Synthetic Aperture Radar (SAR) images and Sentinel-2 multi-spectral images, or from global digital elevation model (DEM) sources at coarse resolution. To achieve higher resolution mapping, recently Google () demonstrated the potential of optical satellite imagery for the rapid retrieval of building information by generating building presence and height data at a 4 m resolution using only Sentinel-2 images. However, this dataset is not globally available yet and is primarily limited to regions in the Global South.

In summary, raster-based products of this kind usually have a large coverage or even global coverage, but with the compromise of a lower quality and resolution, restricting their use in applications requiring high granularity at the instance level.

Instance-level building representation maps delineate individual buildings and assign their corresponding 3D features. There already exist several 3D building model products at the instance level, albeit on a smaller scale (e.g., city or country). For instance, created 3D-BAG, the first open 3D building dataset based on Light Detection and Ranging (LiDAR) data covering the Netherlands; while the released a high-quality 3D semantic city model derived from high-resolution aerial photographs. Similar models are available in other developed areas worldwide, e.g., Philadelphia by , and 260 cities in Japan by the . Although they are derived from high-quality observations, and are capable of reconstructing detailed 3D building structures at Level of Detail 2 (LoD2) or higher, such products are not widely available globally due to their high operational costs and significant computational demands.

In addition to the aforementioned observation-based techniques, Volunteered Geographic Information (VGI) offers an alternative approach. While VGI – based on human annotations – is often recognized for its positional accuracy, it typically suffers from limited completeness, particularly with respect to building height information. Therefore, current mainstream products are primarily derived from aerial or satellite remote sensing data using algorithmic approaches. Microsoft () maintains a dataset of 1.4 billion building footprints, of which approximately 20 % include estimated building heights. provided 3D-GloBFP – a 3D building footprint dataset encompassing 1.66 billion buildings. Despite achieving substantial coverage in terms of building footprints and height estimates, none of these datasets are derived from complete or globally consistent sources. Consequently, the total number of buildings represented falls significantly short of the United Nations' estimate of approximately 4 billion buildings worldwide ().

While raster-based building representation products prioritize broader coverage, they suffer from low resolution and quality limitations. Instance-level representation products, on the other hand, offer detailed information on individual buildings but are constrained by limited coverage and incomplete datasets. Moreover, the production of high-level building representation products often involves multiple data sources, complicating the process of updating these datasets efficiently.

To address these gaps, we propose a novel pipeline to generate global LoD1 building models based solely on PlanetScope images. Leveraging this approach, we generated 3 m-resolution raster height maps alongside globally consistent building polygons. By fusing the generated building polygons with existing open-source building footprints based on quality assessments, we constructed a global building footprint dataset optimized for the highest possible quality. Finally, we assigned building heights to every individual building polygon using our generated height maps, which led to the first complete global LoD1 building models.

We collected global-scale PlanetScope Surface Reflectance (PSR) imagery to support polygonal building mapping and height estimation. PSR data are orthorectified, multi-spectral satellite images provided by Planet's satellite constellation. These images undergo atmospheric correction, ensuring that the reflectance values accurately represent the Earth's surface properties. The PSR data we utilized include four spectral bands – three in the visible RGB range and one near-infrared (NIR) band, with a spatial resolution of approximately 3 m. Additionally, the PlanetScope constellation offers a high temporal revisit frequency, capturing imagery of the same location up to daily, which is beneficial for monitoring dynamic urban environments.

We utilized building footprint datasets from various sources (see Table ), including OSM (), Google Open buildings (), Microsoft Building Footprints (), and CLSM (). Since none of the above-mentioned footprints is complete, we also generated our own global building polygons from an updated version of GlobalBuildingMap (), with a novel building regularization and polygonization pipeline applied. The building mapping and regularization networks were trained on PSR imagery using manually filtered OpenStreetMap (OSM) annotations with high completeness. The data distribution is illustrated in Fig. . To produce the final complete global building footprint dataset – GBA.Polygon – we integrated all available sources, including our own generated polygons, using a quality-based fusion approach.

As the most precise 3D measurement available, we used aerial LiDAR data covering a wide range of geolocations (see Fig. ) in developed countries. A comprehensive list of the covered countries is provided in Appendix . These data are publicly released by governments for open use. Due to the high operational cost of LiDAR observations, no such data are available in Africa. LiDAR observations were primarily acquired in 2019 to ensure temporal consistency with the PSR data. When 2019 acquisitions were not available, data from adjacent years were utilized. Since the coverage is concentrated in developed regions, where building changes are generally limited, the temporal discrepancies between the PSR data and the 3D labels are expected to be negligible. We processed the LiDAR data into normalized digital surface models (nDSMs), which represent the heights of ground objects. These nDSMs served as reference data for training our neural networks.

The overall workflow for generating the GlobalBuildingAtlas dataset is illustrated in Fig. . The workflow can be divided into 4 sections, including global data acquisition, global building polygon generation, global building height estimation and post processing, i.e., global LoD1 model generation.

This section details the acquisition, organization and preprocessing of PSR data to prepare them for the subsequent deep learning models and LoD1 model generation. We divided the Earth's surface into grid cells of 0.2° × 0.2°. Grids overlapping with built-up areas, as defined by the Global Urban Footprint (GUF) dataset (), were selected as our areas of interest. For these regions, we acquired satellite imagery from 2019 that intersects with the selected grids, aiming to maximize coverage. In total, approximately 800 000 PlanetScope scenes were downloaded, each with a spatial resolution of 3 m and covering an area of about 287.5 km2.

To ensure data quality, we applied a cloud filter that retained only scenes with less than 10 % cloud cover. In areas where suitable cloud-free imagery from 2019 was unavailable, the dataset was supplemented with images from 2018. For each 0.2° × 0.2° grid cell, PSR imagery was mosaicked using a scene prioritization strategy guided by the corresponding Unusable Data Masks (UDMs), ensuring the selection of the clearest and most complete pixels across overlapping scenes.

This section outlines the global-scale building polygon generation process, as illustrated in the upper middle part of Fig. . We begin by describing the curation of the training data, followed by a detailed description of the building footprint extraction model, the regularization model, and the polygonization method used to generate the building polygons. Finally, we present the filtering process employed to remove false-positive polygons.

The proposed pipeline acquires 3D building information exclusively from optical satellite imagery, ensuring broad accessibility and ease of data acquisition. We utilized deep learning methods that are particularly well-suited for this task, as they can effectively leverage large-scale data while offering flexibility and scalability – key advantages for global geospatial applications. This section details the height information retrieval process, illustrated in the lower middle part of Fig. . We first describe the curation of the training data and then elaborate on the monocular height estimation process.

The generation of the global LOD1 building model relies on the availability of high-quality global building polygons and the corresponding building heights. Given the limited availability of accurate and detailed global building height datasets, GBA.Height represents the most suitable option for this purpose. However, to leverage existing large-scale building footprint datasets  – such as the vector sources listed in Table , which were derived from higher-quality input data  – we introduce a quality-guided fusion strategy. This approach, detailed in Sect. , enables the integration of multiple building polygon sources. The final global LoD1 building model was then constructed based on this fused footprints and our generated GBA.Height, as described in Sect. .

Our resulting GlobalBuildingAtlas dataset consists of global building polygons (GBA.Polygon), building height maps (GBA.Height) and LoD1 models (GBA.LoD1). Figure  gives an overview on the GlobalBuildingAtlas dataset. It illustrates the global distribution of building volumes (in the middle) as well as statistics on the building counts, areas and volumes and associated errors by continent (top), and showcases the resulting building height map and LoD1 model from selected areas in North America (middle-left) and Asia (middle-right), respectively. The bottom subfigures offer a closer look at the building volume distribution in 10 representative cities across the globe, using the same colormap as the global map for consistency. The total building count across the globe is 2.75 billion, corresponding to a total building area of 506.64 billion m2 (square meters) and a total building volume of 2.85 trillion m3 (cubic meters). This suggests that the actual global building count is lower than the UN's estimation of 4 billion. As anticipated, building volumes are concentrated in metropolitan regions, with notable clusters in East Asia, Europe, and North America. The global distribution of building-related metrics by continent reveals significant geographic variations in terms of quantity and scale.

Asia leads in nearly all metrics, with an estimated 1.22 billion buildings, the highest among all the continents. This reflects both the continent's large population and the extensive urban development across countries such as China, India, and those in Southeast Asia. Africa follows with approximately 540 million buildings, surpassing Europe (403 million) and North America (295 million) in terms of building count. South America contributes around 264 million buildings, while Oceania has the lowest count at 14 million buildings.

Despite Africa's relatively high number of buildings, its total building area is significantly lower, amounting to only 38 billion m2, in contrast to Asia's 218 billion m2, Europe's 138 billion m2, and North America's 79 billion m2. South America and Oceania contribute 28 and 5 billion m2, respectively. This disparity indicates that African buildings tend to be smaller in size, possibly reflecting the dominance of low-rise or informal structures – an expected pattern in less developed or peri-urban areas.

A similar pattern can be observed with the building volume statistics. Asia again ranks highest, with a cumulative building volume of approximately 1.272 trillion m3, followed by Europe (763 billion m3) and North America (538 billion m3). Africa, despite its large number of structures, contributes only 117 billion m3, further underscoring the prevalence of small-scale or single-story buildings. South America and Oceania contribute 123 billion m3 and 26 billion m3, respectively.

The building height and volume estimation accuracy of the GlobalBuildingAtlas dataset across continents is mainly influenced by several factors: (1) variations in building morphology, (2) the representation of regional characteristics in the training data, and (3) the comprehensiveness and quality of the validation data. The global average height RMSE of GBA.LoD1 stands at 5.5 m. The lowest error is observed in Oceania (1.5 m), followed by Europe (4.1 m), North America (5.3 m), and Asia (5.9 m). South America, however, exhibits a significantly higher height RMSE of 8.9 m, suggesting a greater uncertainty or variation in building morphology, possibly due to a mixture of informal and vertical developments in urban regions. The global average volume RMSE of GBA.LoD1 is approximately 152.6 m3per100m2. South America again shows the highest error at 586.8 m3per100m2, which may be attributed to compounding uncertainties in both the building footprint and height estimation. Asia follows with 247.7 m3per100m2, then Europe (150.9 m3per100m2), North America (135.5 m3per100m2), and Oceania with the lowest error at 46.8 m3per100m2. Furthermore, GBA.Height enhances building volume estimation in South America and Asia, achieving RMSE values of 580.0 and 38.0 m3per100m2, respectively.

In summary, the statistics reflect significant continental differences in the scale of the built environment and estimation accuracy. While Asia dominates in terms of the building count, area, and volume, Africa's high building count but low aggregate volume and area underscores the heavy presence of small-scale and potentially informal constructions. Conversely, Europe and Oceania demonstrate relatively lower estimation errors, suggesting a higher consistency in terms of the building structures and their better representation in the training data. South America, despite its moderate building count, emerges as a region with notably high estimation errors, indicating potential challenges in urban morphology or limitations in the training data.

These findings provide valuable insights for global-scale urban modeling, infrastructure planning, and geospatial analytics.

The released GlobalBuildingAtlas dataset provides comprehensive information for both raster-based and instance-level building representation. For comparison, we compared it against three raster-based products, namely GHS-BUILT-H (), WSF-3D (), and Google 2.5D (), as well as two instance-level products, namely Microsoft () and 3D-GloBFP ().

The UN has estimated that there are approximately 4 billion buildings worldwide (), a figure that exceeds the number of building instances in any existing datasets, including our GBA. However, by using the N-ratio evaluation metric shown in Table , which reflects the completeness of GBA.Polygon, we provide an alternative estimate of the global building count based on our product to validate the UN's estimate.

Initially, N-ratios were computed continentally and globally against reference data, and ranged from 0.69 in South America to 1.19 in Europe, with a global average of 1.03. Using these ratios, we estimated the total number of buildings per continent by scaling the known building counts accordingly. For Africa, where no N-ratio is available, we computed three estimates using the minimum, maximum, and global average N-ratios from other continents.

Aggregate projections across continents yield a global estimate of 2.71 billion buildings, while the estimated bounds range from 2.64 billion to 2.97 billion buildings. This analysis suggests that the UN's estimate may exceed actual global building counts.

Based on the preceding analysis of our dataset, its key strengths and limitations can be summarized as follows.

To monitor progress toward SDG 11 – Make cities and human settlements inclusive, safe, resilient, and sustainable – several indicators have been defined (). Among these, Indicator 11.3.1 – Ratio of land consumption rate to population growth rate – plays a crucial role (). The indicator is computed by deriving the built-up area change and population change. Due to the absence of time series data, we were unable to directly compute the indicator to substantiate the rationale for employing volume-based indicators, though generating time series data using our developed pipeline is relatively straightforward. As an alternative approach, we examined the use of volume-based indicators by considering the secondary indicator defined in the indicator metadata () – built-up area per capita. In this context, we substituted built-up area with built-up volume to assess whether the revised indicator could provide a more accurate reflection of the urban development status.

Sustainable development is influenced not only by factors related to the built-up environment but also by a wide range of other social, economic, and environmental variables. Nevertheless, correlation analysis between built-up environment indicators and development-related measures can provide insights into the relative effectiveness of different indicators. In this study, we used gross domestic product (GDP) per capita in year 2019 () as a proxy for development status and compared its correlation with both the building area per capita and building volume per capita. The per-capita building metrics were computed using population data for the year 2019 ().

As Fig.  shows, GDP per capita exhibits a strong positive correlation with both building area per capita and building volume per capita, with correlation coefficients of 0.76 and 0.85, respectively. The significantly higher correlation coefficient between building volume per capita and GDP per capita suggests building volume per capita might be a more suitable measure.

To further validate this observation, we assessed the alignment between GDP per capita and building-based indicators – namely, building volume per capita and building area per capita – by randomly selecting pairs of countries and comparing their respective rankings. For each pair, we examined whether the ranking based on the indicator agreed with the ranking based on GDP per capita. As shown in Table , among 20 301 country or territory pairs sampled from a total of 202 countries and territories, using building volume per capita instead of building area per capita resulted in 788 additional ranking agreements. This substitution increased the agreement rate from 79.6 % (area-based) to 83.5 % (volume-based), demonstrating the superior discriminative power of building volume per capita in capturing differences in economic development.

In conclusion, volume-based indicators demonstrate greater potential than area-based metrics for assessing progress toward the SDGs related to sustainable urban development.