- Author
- Mark Litwintschik
- Published
- 6 September 2025
- Kind
- Personal technical blog post, with commands and query output
- Original
- tech.marksblogg.com/building-footprints-gba.html
- Subject
- Databases and storage
- Themes
- Given away
What it is
Mark Litwintschik is a consultant in big data and mapping. The post describes a dataset from researchers at the Technical University of Munich. It shows how to load the data into QGIS. It also explains how the Parquet copy was built.
What it says
- The post says the dataset estimates 2.75B buildings, against the UN belief of 4B.
- It says the author turned 1.1 TB of GeoJSON into 210 GB of Parquet.
- It counts 2,747,186,815 records, about 208M more than Overture's August release.
- It reports that Google data supplies 1,619,713,039 footprints, the largest single source.
- It says most map hexagons have only one or two sources.
- It finds the footprints more accurate than an earlier AI dataset in Beijing. Flaws remain.
- It says the data lacks timestamps, so old OpenStreetMap buildings may be out of date.
What evidence it gives
The post shows shell scripts, SQL queries and their outputs. It includes source counts, country counts and file sizes. Map screenshots compare datasets in Beijing and one construction site. The accuracy claims rest on visual inspection of a few areas.
Interests and limits
The author converted the data and describes the author's own workflow. The post ends with an offer of consulting services. No formal accuracy test was run. The map images were not viewed, so only the written text was read. The height dataset of 35 TB was not analysed.