The first 3D world map of 2.75 billion buildings: how GlobalBuildingAtlas redraws the planet

Home » Python » The first 3D world map of 2.75 billion buildings: how GlobalBuildingAtlas redraws the planet

The planet has quietly gained an extraordinary new layer of information: a global 3D map of 2.75 billion buildings, covering roughly 97% of all human-made structures on Earth. From sprawling skyscraper districts in China to scattered rural homes in the Sahel, almost every roof has been turned into a measurable object in three dimensions.

Behind this feat is GlobalBuildingAtlas, a massive open dataset that reconstructs the height, footprint and volume of virtually every building using satellite imagery and machine learning. Far from being just a shiny visualization, the project is being positioned as a foundational tool for urban planning, climate and energy modelling, disaster risk assessment and social research on a truly planetary scale.

What exactly is the 3D map of 2.75 billion buildings?

At its core, GlobalBuildingAtlas is a worldwide inventory of buildings in 3D, produced at a spatial resolution of 3 × 3 metres. Every mapped structure is represented both as a 2D footprint on the ground and as a simplified 3D block with an estimated height, enabling researchers to calculate floor area, volume and how dense or sparse built-up areas really are.

According to the team, the atlas contains 2.75 billion building polygons. For about 2.68 billion of them (around 97%), the data reaches what is known as Level of Detail 1 (LoD1): geometrically simplified solid blocks that capture the basic shape and elevation of each building. This is not architectural-level detail, but it is accurate enough to feed numerical models, simulations and AI systems that need consistent global coverage.

Compared with earlier global building datasets, which topped out at around 1.7 billion structures, the new map adds over a billion additional buildings and provides much finer granularity. The spatial detail is described as up to 30 times higher than some of the most widely used previous inventories, particularly in regions that were barely represented before.

That increased coverage matters because formerly under-mapped areas such as large parts of Africa, South America and rural Asia now appear with a level of detail similar to that traditionally reserved for Europe or North America. In other words, the atlas is not just bigger; it is also more geographically balanced.

Six to seven years of work: how the global 3D map was built

Creating a 3D model of almost every building on the planet was not a matter of a single algorithm run. The project took about six to seven years of development, combining satellite remote sensing, deep learning and a patchwork of reference datasets provided by various mapping initiatives and public agencies.

The backbone of the project is imagery from PlanetScope Surface Reflectance, a constellation of commercial satellites that image the Earth at roughly 3 metres per pixel. For this atlas, the researchers collected and processed around 800,000 satellite scenes, mainly from 2019 and supplemented in a few cases with 2018 data, carefully selecting images that were largely free of clouds and atmospheric disturbances.

These scenes were not simply stacked. They were orthorectified and atmospherically corrected so that each pixel corresponds to a precise patch of ground and reflects the properties of the surface, rather than haze or lighting artefacts. The team then stitched this enormous archive into a global mosaic, picking, pixel by pixel, the cleanest observation for each location.

To focus processing power where humans actually build, the researchers used a prior product, the Global Urban Footprint, to identify tiles likely to contain settlements. Only those segments were passed through the subsequent building-detection pipeline, cutting down computation while still capturing isolated settlements and small towns.

One of the trickiest challenges was to distinguish real buildings from other bright or structured objects seen from space—such as roads, cliffs, industrial infrastructure or tree canopies. The group developed a multi-step workflow to detect, refine and finally convert potential buildings into usable vector footprints.

From satellite pixels to individual buildings

The first step in turning imagery into a building map involved training a deep neural network to recognise where buildings are present. For this, the team cut the satellite mosaic into smaller patches and paired them with existing building footprints from sources such as OpenStreetMap and a large annotated dataset from China.

Those vector footprints were rasterised to match the three-metre grid of the PlanetScope imagery, producing training data where each pixel was labelled as “building” or “not building”. A encoder-decoder style neural network then learned to output a “building mask”: an image where bright pixels indicate predicted building locations.

However, the raw output of this model tended to merge neighbouring buildings into continuous blobs, especially in dense urban cores. To address this, the team built a second regularisation network to clean up the masks, split fused shapes and sharpen boundaries before turning them into polygons. Algorithms for contour detection, polygon simplification and small-object filtering were applied to convert these binary masks into vector footprints.

Even then, not all detected objects were genuine structures. The researchers cross-checked the results against a global land-cover map (WorldCover), removing elements clearly located over water bodies, glaciers, compact forest or other land types where buildings are extremely unlikely. This additional filtering step proved essential for limiting false positives in remote areas.

Because no single footprint dataset is complete or consistent at global scale, the project uses a quality-guided fusion strategy. In each administrative region, the team picked the most reliable source—often OpenStreetMap, but also Google’s Open Buildings for parts of Africa and South America, Microsoft building data or a regional dataset for East Asia (CLSM)—as the primary layer and then enriched it using secondary sources where gaps existed.

In practice, this means that in each region the atlas retains all buildings from the best source available, adds non-overlapping buildings from the second-best source and relies on its own automatically generated footprints to fill remaining voids. The outcome is a single, harmonised layer of building polygons that, according to the authors, is more complete than any of its components in isolation.

How the team estimated height and volume

Turning 2D building outlines into 3D objects required another major step: estimating how tall each structure is. To do this, the group assembled a large collection of aerial LiDAR data covering 168 cities, mostly in Europe, North America and Oceania, where airborne laser scanning has been deployed at scale.

From these LiDAR sources they derived normalised digital surface models (nDSM), where every grid cell indicates how many metres that point rises above the ground surface. These nDSMs served as the “ground truth” for training a separate neural network that could infer building height directly from a single optical satellite image.

Once trained, this monocular height estimation model was run over the global PlanetScope mosaic, sliding across the surface with overlapping windows to cover every pixel. For each 3 × 3 metre cell, the network produced a predicted height value. To quantify reliability, the system generated multiple slightly perturbed predictions and measured how much they varied, assigning an uncertainty estimate to each location.

The final step was to combine the refined building footprints with this height grid. For every individual building polygon in the atlas, the system sampled the height layer and typically assigned the maximum height value found within that footprint as the representative building height, along with the associated uncertainty metric. From that height and footprint area, total volume for each building can be calculated.

Although the LoD1 models are visually simple—more like carefully stacked boxes than fully detailed architecture—they capture enough of the built form to support robust analysis. Tests in cities across North America, South America, Europe, Asia and Oceania show that, while errors vary by region and urban form, the global dataset performs at least on par with, and often better than, existing large-scale building-height products.

What the 3D building atlas reveals about the world

With the technical pipeline in place, the atlas can be used to draw a sort of numerical X-ray of the built environment. Across all continents, the dataset totals around 506,640 million square metres of building footprint and approximately 2.85 trillion cubic metres of constructed volume.

One immediate finding is that previous global estimates of building counts appear to have been too high. A common figure circulating in United Nations reports suggested there might be roughly 4 billion buildings worldwide. The 2.75 billion identified here—combined with the systematic way they were detected—suggests that the earlier number likely overstated the true total.

Regional comparisons bring more nuance. Asia emerges as the undisputed heavyweight in terms of both building numbers and total volume. The atlas counts roughly 1.22 billion buildings on the continent, alongside about 1.27 trillion cubic metres of built volume. These figures reflect the rapid urban expansion and high population densities of countries like China, India and those in Southeast Asia.

Africa holds the second-largest number of buildings, at around 540 million structures, but with far less accumulated volume—on the order of 117 billion cubic metres. This mismatch between building count and volume underscores the prevalence of low-rise, small-footprint dwellings, particularly in informal settlements and rural communities.

In Europe and North America, the atlas finds fewer buildings than in Africa, but the average volume per structure is substantially higher. Urban areas often combine mid-rise and high-rise blocks, warehouses and larger detached houses, all of which push up the typical building volume even when the number of buildings is lower.

South America, meanwhile, stands out in the analysis for having some of the largest errors in height and volume estimation. The team links this to complex mixtures of high-rise cores and informal, densely packed neighbourhoods that are more challenging for the model to interpret consistently, highlighting where future improvements and more local reference data would be most useful.

A new metric: built volume per person

Perhaps the most provocative aspect of the project is the introduction of a new indicator: built volume per capita. Instead of just measuring how much land is urbanised, this metric looks at the total constructed volume relative to the number of people living in a given area.

The research team argues that this approach captures inequalities that flat 2D maps tend to hide. Two neighbourhoods might cover the same surface area on a traditional map, but their vertical profiles—and the living conditions they offer—can be dramatically different.

Using the new 3D dataset, they highlight cases such as Finland and Greece. Finland turns out to have roughly six times more built surface area per person than Greece, indicating more space per inhabitant and different urban and housing patterns. On the other end of the scale, Niger appears with a built area per capita that is about 27 times lower than the global average, pointing to severe deficits in infrastructure and housing.

These disparities are not limited to Europe or Africa. Across continents, the atlas reveals that wealthier regions usually enjoy more volume per person, wider streets and larger buildings, whereas poorer districts often combine cramped, low-rise housing with limited public infrastructure. The contrast becomes stark when comparing, for example, affluent districts of major cities with nearby informal settlements.

For the project’s lead scientist, Professor Xiaoxiang Zhu of the Technical University of Munich, this shift is fundamental. She and her colleagues argue that cities should be treated as three-dimensional objects when assessing progress towards the UN Sustainable Development Goal 11, which focuses on sustainable cities and communities, rather than relying solely on how much land is classified as “urban”.

In their view, the volume of buildings per inhabitant offers a more direct, although still imperfect, lens on living standards, the availability of infrastructure and the intensity of land use than maps that merely outline the edges of built-up areas.

From climate modelling to disaster response

Beyond describing global patterns, the 3D building map is designed to be practically useful for a wide range of applications. Because every building has an associated footprint, height and location, the atlas can feed directly into models that need detailed representations of the built environment.

One clear area is climate and energy analysis. Buildings are estimated to account for around 40% of global CO₂ emissions, largely through heating, cooling and electricity use. Having consistent 3D data on building stock worldwide allows researchers to better estimate energy demand, simulate different retrofit scenarios and quantify potential emission reductions from changes in construction, insulation or urban design.

Another immediate use is disaster risk reduction. Institutions like the German Aerospace Center, which is involved in the International Charter: Space and Major Disasters, are already exploring how the atlas can help assess which structures and populations are exposed to floods, earthquakes, landslides or storms. With three-dimensional data, it is easier to estimate how many people might be affected at different floors, or how much built volume lies within a floodplain.

For urban planners and local authorities, having a consistent 3D baseline opens up possibilities to simulate interventions before they are built. City governments can, for instance, identify neighbourhoods where housing supply is far below population needs, locate potential sites for new schools or health centres and test how adding green spaces or changing street layouts might affect heat exposure or accessibility.

The open nature of the dataset is key. The atlas is available online through an interactive map that functions in a way many users will find familiar: one can pan, zoom, choose different background layers such as standard maps or satellite views, and search for specific places by name or address. Users can toggle between volume visualisations and LoD1 3D block representations to explore their own city or remote regions.

For those needing deeper access, the underlying data and code can be downloaded from GitHub. This allows researchers, public agencies and even private companies to run their own analyses, integrate the atlases into existing systems or adopt managed graph databases to represent complex relationships.

Monitoring urbanisation in near real time

One of the most appealing promises of GlobalBuildingAtlas is that it need not remain a one-off snapshot of the year 2019. Because the pipeline is based on regularly acquired satellite data and trained models, it can, in principle, be rerun periodically to produce updated views of the world’s building stock.

Urban planning scholar Dorina Pojani, from the University of Queensland, has emphasised that this could allow researchers and policymakers to track the evolution of cities over the next five to ten years, rather than relying on infrequent censuses or local datasets that are rarely harmonised across countries.

In regions where planning information is scarce or outdated, such as fast-growing secondary cities in Africa or Asia, this could provide the first reliable, up-to-date baseline of the built environment. Planners would be able to see, for instance, how informal settlements expand, where industrial zones encroach on agricultural land or which peri-urban areas are filling up with new construction.

For demographic and socio-economic studies, such temporal updates could be crossed with population estimates to observe how built volume per person changes over time. Are certain areas catching up in terms of housing and infrastructure, or are disparities widening? Which policies are associated with more balanced growth in built volume and population?

From a technical perspective, the potential for more frequent updates will depend on factors like satellite data availability, computing resources and the ability to refine models with new reference datasets, especially in underrepresented regions. Still, the pipeline demonstrated for the 2019 map offers a template for future “snapshots” of the world’s buildings.

Transparency, governance and even corruption research

Beyond physical planning and climate studies, the atlas may also have implications for governance and transparency. Because it enables systematic linking between the physical presence of buildings and other datasets, some researchers see it as a tool to investigate how power and money shape the built environment.

Urban planning expert Dorina Pojani has pointed out that, in principle, one could use the building-level data to associate specific projects with developers, corporations or political actors. By overlaying land registries, company records or public procurement data, analysts could start asking whether certain networks of individuals have a disproportionately large presence in high-value or strategically located projects, supported by Amazon Neptune.

Such analyses could contribute to studies of urban corruption, land speculation and capture of planning processes. They might help identify patterns where building booms coincide with policy changes, or where certain neighbourhoods receive repeated high-end development while others remain systematically neglected.

Another expert, Liton Kamruzzaman from Monash University, has underlined that the atlas offers particular value in countries that currently lack reliable planning information. In such contexts, where even basic maps of urban expansion may be missing, the availability of a global 3D building layer could support more transparent debate about how cities are growing, which communities receive infrastructure and how risks and amenities are distributed.

Of course, the atlas does not provide a full picture of ownership, tenure rights or social dynamics. However, by making the physical side of the story far more visible and measurable, it can serve as a starting point for more informed discussions about equity, justice and accountability in urban development.

Looking ahead, the fact that the dataset is public and reproducible means journalists, civil society and researchers can independently test claims about construction patterns, infrastructure provision or the outcomes of major development programmes, rather than relying solely on official statistics.

Across all these domains—urban studies, climate science, risk management and governance—the new 3D map of 2.75 billion buildings marks a step change in how the world’s built environment can be observed and analysed. By replacing a flat, patchy picture with a three-dimensional, nearly global inventory, GlobalBuildingAtlas offers a common reference frame for understanding where and how people live, what has been built for them and how unevenly that built volume is shared.