The World's 2.75B Buildings (tech.marksblogg.com)

Researchers at Technical University of Munich released the GlobalBuildingAtlas (GBA), a planet‑scale building footprint and height collection that revises previous estimates by putting the global building count at ~2.75 billion (vs the UN’s ~4 billion). GBA is distributed as two complementary products: a 1.1 TB LoD1 vector layer (922 uncompressed GeoJSON tiles) that the author converted to a 210 GB GeoParquet set now hosted on AWS S3, and a much larger 35 TB raster “Height” product of derived elevation maps. Parts of the atlas were generated by running open‑source deep‑learning code on Planet Labs’ daily satellite imagery constellations (code only; imagery/weights not included). For the AI/ML and geospatial community this is significant: it provides near‑global, structured building footprints with heights at scale, enabling large‑scale model training, validation, urban analytics, disaster response simulation, and transfer‑learning across diverse geographies. The dataset is already practical to use—readable in DuckDB, visualizable in QGIS (2.5D styling) and ingestible into ArcGIS—thanks to conversion to Parquet and tooling like GeoParquet Downloader. Key caveats: the Height rasters are enormous (35 TB) and will be costly/time‑consuming to pull; data quality, coverage heterogeneity and sampling biases (satellite revisit, model accuracy across regions) must be evaluated before downstream use. Overall, GBA is a major open resource that lowers the barrier for global geospatial ML while raising new challenges around storage, provenance and fairness.