PaperlayerView exact revision on arXiv

Source-linked AI summary

GlobalBuildingAtlas: An Open Global and Complete Dataset of Building Polygons, Heights and LoD1 3D Models

Xiao Xiang Zhu, Sining Chen, Fahong Zhang, Yilei Shi, Yuanyuan Wang

arXiv:2506.04106v1cs.CV

TL;DR

GlobalBuildingAtlas addresses uneven global availability of comprehensive building-height data by providing an open dataset of building polygons, heights, and LoD1 3D models. Its optical-satellite pipeline produces 3 m height maps and a global building model covering 2.75 billion buildings, with 97.7% having height estimates and RMSEs of 1.5–8.9 m.

  • Problem

    Comprehensive building-height data remains unevenly available across the globe.

  • Method

    GlobalBuildingAtlas uses an optical-satellite pipeline to produce LoD1 3D building models and pixel-wise height maps at 3 m resolution.

  • Results

    2.75 billion buildings are covered, 97.7% have height estimates, and height-estimation RMSEs range from 1.5 m to 8.9 m.

  • Takeaways & Limitations

    The resulting building volumes demonstrate significant advantages over existing products.

  • Takeaways & Limitations

    The height-estimation model was neither trained nor validated on African samples, so domain shift may limit its generalizability and accuracy there.

Abstract

from arXiv · show

We introduce GlobalBuildingAtlas, a publicly available dataset providing global and complete coverage of building polygons, heights and Level of Detail 1 (LoD1) 3D building models. This is the first open dataset to offer high quality, consistent, and complete building data in 2D and 3D form at the individual building level on a global scale. Towards this dataset, we developed machine learning-based pipelines to derive building polygons and heights (called GBA.Height) from global PlanetScope satellite data, respectively. Also a quality-based fusion strategy was employed to generate higher-quality polygons (called GBA.Polygon) based on existing open building polygons, including our own derived one. With more than 2.75 billion buildings worldwide, GBA.Polygon surpasses the most comprehensive database to date by more than 1 billion buildings. GBA.Height offers the most detailed and accurate global 3D building height maps to date, achieving a spatial resolution of 3x3 meters-30 times finer than previous global products (90 m), enabling a high-resolution and reliable analysis of building volumes at both local and global scales. Finally, we generated a global LoD1 building model (called GBA.LoD1) from the resulting GBA.Polygon and GBA.Height. GBA.LoD1 represents the first complete global LoD1 building models, including 2.68 billion building instances with predicted heights, i.e., with a height completeness of more than 97%, achieving RMSEs ranging from 1.5 m to 8.9 m across different continents. With its height accuracy, comprehensive global coverage and rich spatial details, GlobalBuildingAltas offers novel insights on the status quo of global buildings, which unlocks unprecedented geospatial analysis possibilities, as showcased by a better illustration of where people live and a more comprehensive monitoring of the progress on the 11th Sustainable Development Goal of the United Nations.

1 Introduction

Global building data remains uneven, incomplete, and often two-dimensional, limiting understanding of urban form and building-level conditions. GlobalBuildingAtlas addresses these gaps with globally consistent polygons, height maps, and LoD1 models derived from optical satellite imagery.

  • Motivation: 2D built-up-area metrics can misrepresent urban conditions because they omit vertical space and building heights.Different neighborhoods can share similar built-up area per capita despite contrasting spatial, social, and infrastructural conditions.
  • Research gap: Comprehensive building-height data remains unevenly available globally, especially in regions lacking observational infrastructure and technical capacity.These regions are often vulnerable and most in need of accurate, timely data.
  • Research gap: Existing global building products often lack true global completeness, building-level detail, or scalability because they rely on ancillary data.These limitations are particularly problematic in fast-changing urban contexts.
  • Contribution: GlobalBuildingAtlas uses optical satellite imagery to provide comprehensive building polygons, heights, and LoD1 3D building models.The dataset is designed for global-scale coverage and individual-building representation.
  • Contribution: 2.75 billion buildings are included in GBA.Polygon, addressing more than 40% of global buildings previously unaccounted for.GBA.Polygon is presented as a complete global building-polygon dataset.
  • Contribution: 3×3 meters is the GBA.Height resolution, 30 times finer than previous 90 m global products, with RMSEs of 38.0–580.0 m3/100 m2 across continents.The height maps support building-volume analysis at local and global scales.
  • Contribution: 2.68 billion building instances comprise GBA.LoD1, with predicted heights and RMSEs ranging from 1.5 m to 8.9 m across continents.This is presented as the first complete global LoD1 building model.

2 Related Work

Existing building products trade coverage against spatial detail, quality, completeness, and updateability. GlobalBuildingAtlas combines satellite-derived height maps with quality-based polygon fusion to produce globally consistent LoD1 building representations.

  • Raster-based products: Raster-based products aggregate building heights or volumes into regular grids, commonly at tens-to-hundreds-of-meters resolution.Examples include WSF 3D at 90 m, GHSL products at 250 m, and global building heights at 150 m.
  • Satellite-derived products: Optical imagery has enabled 4-meter building presence and height data from Sentinel-2, but that dataset is not globally available.Its availability is primarily limited to regions in the Global South.
  • Raster-based products: Raster products can provide broad or global coverage, but lower quality and resolution restrict instance-level applications.Their coverage-versus-granularity trade-off limits detailed building analysis.
  • Instance-level products: Instance-level products delineate individual buildings and assign corresponding 3D features, but existing datasets lack complete or globally consistent sources.Their represented building counts therefore fall short of the United Nations estimate of approximately 4 billion buildings worldwide.
  • Detailed 3D models: High-quality LoD2-or-higher models exist for selected cities or countries, but operational costs and computational demands limit global availability.Examples include LiDAR- or aerial-photography-based models in the Netherlands, Helsinki, Adelaide, Philadelphia, and Japan.
  • Data sources: VGI can offer positional accuracy but typically has limited completeness, particularly for building heights.Mainstream products therefore rely primarily on algorithmic approaches using aerial or satellite remote sensing.
  • GlobalBuildingAtlas: GlobalBuildingAtlas addresses these gaps with PlanetScope-only processing, 3 m height maps, quality-based polygon fusion, and per-building height assignment.The pipeline produces a globally consistent footprint dataset and a complete global LoD1 model.

3 Data Sources

GlobalBuildingAtlas combines PlanetScope optical imagery, multiple open footprint sources, and LiDAR-derived height references to construct global building data. PlanetScope supports both polygon mapping and height estimation, while source fusion addresses incompleteness in existing footprints.

  • Satellite optical imagery: PlanetScope Surface Reflectance imagery supports both building polygon mapping and height estimation at approximately 3-meter spatial resolution.The imagery contains RGB and near-infrared bands and offers frequent revisits for monitoring dynamic urban environments.
  • Building footprints: The dataset uses OSM, Google Open Buildings, Microsoft Building Footprints, and CLSM as building-footprint sources.Because none of these sources is complete, the authors also generate additional polygons from an updated GlobalBuildingMap pipeline.
  • Data fusion: Quality-based fusion integrates all available footprint sources, including newly generated polygons, into the final GBA.Polygon product.The strategy is designed to combine complementary sources because existing large-scale footprint datasets differ in completeness.
  • Height references: LiDAR point clouds from government releases are processed into normalized digital surface models used as reference data for height-model training.LiDAR coverage is concentrated in developed countries, and no such data are available in Africa because of high operational costs.

4 Methodology

The pipeline acquires and preprocesses global PlanetScope imagery, generates building polygons through segmentation, regularization, polygonization, and filtering, and estimates heights with monocular deep learning. It then combines polygons and heights into LoD1 models, while acknowledging regional data limitations.

  • Overall workflow: The workflow has four stages: global data acquisition, building polygon generation, building height estimation, and LoD1 model generation.These stages are illustrated in the proposed workflow.
  • Polygon generation: A UPerNet-based encoder–decoder maps PlanetScope images to binary building masks, followed by a regularization network that refines noisy and merged structures.The regularization model learns to denoise perturbed annotation masks and improve building delineation.
  • Polygon generation: Regularized masks are converted to dense vectors, simplified into polygons, and filtered against a dilated World Cover built-up mask to remove false positives.The filter targets errors associated with cloud, snow, water, forests, and similar land-cover features.
  • Height estimation: A HTC-DC Net estimates building heights from single images using a classification–regression paradigm and training samples derived from LiDAR nDSMs.The height-training dataset contains samples from 168 city-scale regions and 231,656 cropped patches.
  • Integration and limitations: The resulting GBA.Polygon combines source datasets, while GBA.Height and LoD1 generation provide building-level three-dimensional outputs.The approach uses optical imagery exclusively for broad accessibility and scalable acquisition.

5 Results

GlobalBuildingAtlas contains 2.75 billion buildings with continental variation in counts, area, volume, and estimation accuracy. GBA.Polygon generally performs best for footprints, while GBA.Height and GBA.LoD1 provide highly complete height products with regional error differences.

  • Continental statistics: Asia contains 1.22 billion buildings, followed by Africa with approximately 540 million, Europe with 403 million, and North America with 295 million.South America contributes around 264 million buildings, while Oceania has 14 million.
  • Continental statistics: Asia leads building volume at approximately 1.272 trillion m3, whereas Africa contributes 117 billion m3 despite its larger building count.The results associate this disparity with the prevalence of smaller-scale or single-story structures in Africa.
  • Accuracy: 5.5 m is the global average height RMSE for GBA.LoD1, ranging from 1.5 m in Oceania to 8.9 m in South America.Europe, North America, and Asia have RMSEs of 4.1 m, 5.3 m, and 5.9 m, respectively.
  • Comparative evaluation: GBA.Polygon achieves the best performance across most footprint metrics, except in South American cities, while GBA.Height and GBA.LoD1 achieve the highest completeness scores.GBA.LoD1 outperforms compared methods for volume estimation across every continent except South America and is strongest for height accuracy in North America and Oceania.
  • Validation of global count: The global building-count estimate based on GBA.Polygon is 2.71 billion, with estimated bounds spanning 2.64 billion to 2.97 billion buildings.The estimate provides an alternative assessment of the UN building-count estimate.

Strengths

GlobalBuildingAtlas offers globally extensive building polygons, high-resolution height maps, and LoD1 models derived primarily from PlanetScope imagery and multi-source footprint fusion. Its coverage and spatial detail are paired with regional accuracy limitations tied to data availability and building morphology.

  • Global footprint coverage: GBA.Polygon comprises approximately 2.75 billion buildings, supported by multi-source fusion and a PlanetScope-based footprint-generation pipeline.The generated pipeline fills previously unavailable footprint gaps, although quality is limited in some regions by the input imagery’s spatial resolution.
  • Height-map resolution: GBA.Height provides a global 3-meter building height map, compared with 90–150-meter resolutions in existing global products.The product is derived exclusively from PlanetScope imagery rather than very high-resolution aerial data.
  • Scalability: PlanetScope’s lower data dependency, lower cost, and frequent revisits support rapid and scalable updates of global building height maps.The approach relies on a single affordable optical satellite source for height estimation.
  • Global LoD1 model: GBA.LoD1 contains approximately 2.68 billion buildings and fills a gap of over 1 billion previously unrepresented structures.Its height RMSE ranges from 1.5 to 8.9 meters across geographic regions.

Limitations

The paper identifies regional limitations in height-estimation training data and systematic underestimation in some high-rise urban areas. These issues constrain generalizability and accuracy, especially in African contexts.

  • Regional training-data limitations: Height data scarcity in Africa meant the model was neither trained nor validated on samples from the region.The authors identify potential domain shift when applying the model to African urban areas.
  • Regional training-data limitations: The model may have limited generalizability and accuracy in African urban contexts because of domain shift.The paper links this boundary directly to the lack of accurate reference data from Africa.
  • Regional training-data limitations: Community efforts are still needed to curate validation data for Africa.Weakly supervised learning is proposed as a possible way to address limited training data, but validation-data curation remains necessary.
  • High-rise underestimation: The height model tends to underestimate buildings in some South American and Asian high-rise cities.The authors attribute this partly to minimizing pixel-level rather than building-instance-level height errors.

6 Applications and Enrichment

The applications use GlobalBuildingAtlas building volumes to examine population alignment and development-related indicators. Results show strong population–volume associations, substantial geographic disparities, and stronger GDP-ranking agreement for volume than area.

  • Population and building volume: Logarithmic regressions relate population counts and building volume across 1 km × 1 km grids in the EU and its member states.Pearson r and Spearman ρ measure linear and monotonic association, respectively.
  • Population and building volume: Population generally has a strong positive correlation with building volume, while correlation values indicate how evenly volume aligns with population density.The regression slope reflects average building volume per capita; higher correlations indicate more uniform allocation across spatial units.
  • EU disparities: Finland has six times Greece’s per-capita building volume, while only seven EU countries fall significantly below the EU-wide correlation benchmark.Finland and Estonia also show abundant building stock whose spatial distribution is less aligned with population density.
  • Global disparities: 0.36%: Niger’s building volume per capita relative to Finland’s, and 27 times below the global average.Country-level analysis finds strong positive population–volume correlation but much greater cross-country imbalance in volume per capita.
  • Development indicators: 0.85: correlation between GDP per capita and building volume per capita, compared with 0.76 for building area per capita.The analysis uses GDP per capita as a development-status proxy and compares it with both building-based per-capita indicators.
  • Development indicators: 83.5%: ranking agreement using building volume per capita, versus 79.6% using building area per capita.Across 20,301 sampled pairs from 202 countries and territories, volume produced 788 additional ranking agreements.
  • SDG applications: Volume-based indicators are presented as having greater potential than area-based metrics for assessing sustainable urban-development progress.The authors could not directly compute the SDG indicator because time-series data were unavailable.

7 Code and Data Availability

The paper provides public access to its code, dataset, figures, and an interactive preview. The code is distributed under an MIT license with a Commons Clause restricting commercial use.

  • Code: The development code and figure-reproduction scripts are available on GitHub.The repository also provides an interactive dataset-preview portal.
  • Code: The GitHub software uses an MIT license with a Commons Clause that restricts commercial use.This licensing condition defines a boundary on how the released software may be used.
  • Data: The GBA dataset is accessible through mediaTUM.The paper supplies the dataset access location separately from the code repository.

8 Conclusions

GlobalBuildingAtlas addresses the lack of comprehensive individual-building 3D information with a global LoD1 modeling pipeline based on optical satellite imagery. It delivers broad building coverage, height estimates, and applications linking building volume to population and development status.

  • 8 Conclusions: The paper targets the lack of comprehensive individual-building 3D information for large-scale urban environments and detailed assessments.This gap motivates the construction of a global building dataset with polygons, heights, and LoD1 models.
  • 8 Conclusions: The pipeline uses only optical satellite imagery and combines building polygon generation with pixel-wise height estimation.The height maps have 3 m resolution, while polygon generation supplements missing building instances.
  • 8 Conclusions: 2.75 billion: buildings represented, with 97.7% having height estimates.The resulting LoD1 3D building model is described as the most comprehensive and accurate dataset to date.
  • 8 Conclusions: 1.5 m to 8.9 m: RMSE range for height estimation across different continents.Building-volume errors range from 46.8 m3/100 m2 to 586.8 m3/100 m2.
  • 8 Conclusions: Strong population–building-volume correlation coexists with substantial disparities in building volume per capita across continental and global scales.These analyses use the dataset’s volumetric information to characterize where people live and differences in built infrastructure.
  • 8 Conclusions: Building volume-based indicators better reflected development status than building area-based metrics in the reported analysis.The paper proposes these indicators for assessing progress toward sustainable-development goals.

Appendix A: Building Volume Distribution by Country or Territory

China accounts for the largest share of global building volume, followed by the USA. Improved building-data completeness also reveals higher volume proportions for Russia, India, and Brazil.

  • 24.8% of total building volume is attributed to China, followed by 15.4% for the USA.
  • Higher proportions of building volumes are uncovered for Russia, India, and Brazil because of improved building-data completeness.
  • Other countries and territories constitute smaller building-volume percentages than in earlier research, aligning more closely with expected patterns.

Appendix B: Contribution of Different Data Sources in Quality-guided Building Polygon Fusion

GBA.Polygon combines multiple building-footprint sources through quality-guided fusion to improve global completeness and consistency. Source contributions vary by continent, with Google Open Buildings dominant in several regions and complementary datasets addressing regional gaps.

  • Source contributions by continent: 1.62 billion buildings from Google Open Buildings make it the largest contributor in Asia, Africa, and South America.
  • Source contributions by continent: 0.49 billion buildings come from OpenStreetMap, with stronger coverage in Europe and North America than in Africa and South America.
  • Source contributions by continent: 0.43 billion buildings come from Microsoft Building Footprints, which supplement OpenStreetMap across multiple continents, especially Europe.
  • Source contributions by continent: 0.07 and 0.14 billion buildings from CLSM and the authors’ polygons primarily address gaps in Asia.
  • Area contributions and fusion effects: Open Buildings contributes a smaller share of building area than of building count, indicating smaller average building size.
  • Area contributions and fusion effects: The authors’ polygons contribute relatively more area than count, likely because lower-resolution imagery can merge adjacent buildings into single polygons.
  • Area contributions and fusion effects: Integrating additional footprint datasets substantially enhances the completeness and coverage of the final building polygons.
Loading 2506.04106v1…