The dataset covers 2D building contours, building heights, and simple 3D models (LoD1 level) of 2.75 billion buildings around the world, including areas such as Africa and South America that are often missing in other map data.
The data is extremely accurate, with a fine resolution of 3×3 meters, and can be called up directly in the Geographic Information System (GIS) software or downloaded in full. It can clearly represent the distribution of population and urban expansion, so it can be widely used in urban planning, disaster risk assessment, climate adaptation response, and monitoring of sustainable development goals. At present, this dataset and related code have been open for sharing for scientific research and practical application scenarios.
If you look at the Earth from space, the traces left by human activity are actually very intuitive: lights, roads, and regular or irregular buildings. But it is not easy to systematically organize these “things you see” into data.
GlobalBuildingAtlas is trying to accomplish such a basic but important thing.
The core idea of the project is simple: using high-resolution images taken by satellites, allowing artificial intelligence to determine where buildings exist, and then compile the results into a global dataset of building distribution. In other words, it is not studying “urban life”, but answering a more substantial question – where exactly humans have built buildings.
The reason why such a “building distribution map” is specially made is because the global data in reality is very uneven. In a few developed areas, maps are detailed down to every building; And in more places there may be nothing on the map, but in fact a large number of buildings and settlements already exist. This inconsistency will make many macro studies difficult and even biased.
GlobalBuildingAtlas’s approach is to re-observe the planet in a unified way. It trains deep learning models to allow AI to learn to recognize “house-like structures” from satellite images, and then applies these recognition results to a global scale. The result is not a text report, but an “architectural layer” that can be superimposed on the map.
It should be emphasized that this project is very restrained. It does not care about the use and value of the house, nor does it involve demographic, property rights, or social attributes. It does only one thing: to determine if there are man-made buildings here and how much space the buildings take up approximately. Because it is basic enough, this type of data can be used repeatedly in many different fields.
For example, when studying urban sprawl, we can directly compare the changes in the distribution of buildings in different periods; When doing a disaster assessment, you can quickly estimate how many buildings a flood or earthquake may affect; In environmental research, it can also be used to determine the extent to which natural land is occupied by human activities. These seemingly “macroscopic” analyses are often built on such an inconspicuous but critical base map.
Unlike the navigation maps we use every day, GlobalBuildingAtlas is not there to give directions or serve life. It is more like a “base of the earth” for researchers and analysts, emphasizing consistency and comparability on a global scale rather than richness of detail.
This project was released by the Zhu-XLab team and is a typical scientific research result that combines remote sensing technology and artificial intelligence. From an engineering perspective, it reflects large-scale data processing and model reasoning capabilities; From a research perspective, it provides a more equitable and unified global perspective.
Github：https://github.com/zhu-xlab/GlobalBuildingAtlas
Tubing: