A research team from TUM1 led by Prof. Xiao Xiang Zhu has published a 3D dataset comprising 2.8 billion buildings world-wide. From the announcement:
How many buildings are there on Earth – and what do they look like in 3D? The research team led by Prof. Xiaoxiang Zhu, holder of the Chair of Data Science in Earth Observation at TUM, has answered these fundamental questions in this project funded by an ERC Starting Grant.
The GlobalBuildingAtlas comprises 2.75 billion building models, covering all structures captured in satellite imagery from the year 2019. This makes it the most comprehensive collection of its kind. For comparison: the largest previous global dataset contained about 1.7 billion buildings. The 3D models with a resolution of 3×3 meters are 30 times finer than data from comparable databases.
The GlobalBuildingAtlas data comes in two types: a digital surface model (DSM) at 3 meters resolution. Secondly, almost all buildings (97%) are also provided as vector data at LOD12.
The TUM team used a ML3-based approach to extract building footprint polygons and building heights4 from PlanetScope satellite data5. Data fusion with existing openly available building polygons6 was also applied. RMSE7 of building height is reported to range from 1.5 meters to 8.9 meters.
The code used to produce the dataset is available under an MIT License with Commons Clause Restriction. The data can be obtained from TUM. A Data description paper in Earth System Science Data offers more detail about the data and the underlying methodoloy.
Footnotes
Technical University Munich, Germany↩︎
Level-of-detail 1, meaning the model captures the shape of the building footprint and the building height, but not, for example, roof shape.↩︎
Machine-learning.↩︎
Apparently, using a method called “monocular height estimation model”, that is the model estimates building height from a single image, not stereometrically.↩︎
PlanetScope is an ESA (European Space Agency) constellation of over 400 small satellites (“cube sats”) with three or four bands (optical to near-infrared) and roughly 3 meters spatial resolution orbiting roughly 500 km above the Earth surface offering a daily revisit time at nadir.↩︎
The Data description paper lists OpenStreetMap (OSM), Google Open Buildings, Microsoft Building Footprints, and CLSM as data sources.↩︎
Root mean square error.↩︎