Seven hundred thousand roofs

Google Open Buildings v3 is the first inventory of buildings that covers the whole of Gabon. A model reads satellite imagery and draws a contour around every roof it recognises, with its ground area and a confidence score. This paper describes what that dataset contains for Gabon, so that whoever uses it knows what they are counting. We cut it to the country, 738,615 footprints kept out of 9,480,086 read, 29 refused and accounted for, and answer four questions: where the roofs are, how large they are, how sure the model is of them, and what a confidence threshold removes.

The roofs are concentrated to an extreme degree. They occupy a twentieth of the territory, and within it 10% of the occupied cells hold 89% of the roofs. The Libreville agglomeration alone holds 55%; the ten largest built patches hold 85%; fifteen towns hold 75%.

The roofs are small. The median footprint is 71 m²; 14% are under 20 m², a size at which the model may be seeing a shed or a shadow rather than a dwelling. All the roofs together cover 76 km².

Confidence is lowest where roofs are isolated and where they are packed. Nationally, 29% of footprints score below 0.75. In the scattered rural cells, 54% do. In the densest urban blocks, 36% do in Libreville and 44% in Port-Gentil. Everywhere in between, 27 to 29%.

A confidence threshold therefore removes the hamlets and the informal quarters first, not noise at random. A cut at 0.75 keeps 71% of the country’s roofs but 46% of the scattered ones; a cut at 0.80 keeps 54%. Whoever applies a threshold should publish, next to it, what it removes in those two fabrics.

Keywordsbuilding footprints · Open Buildings · settlement · confidence · Gabon

The cut. Four S2 tiles, 9,480,086 footprints, are read line by line; a footprint is kept if its centroid falls inside one of the nine provinces. Every line is counted as kept, outside the country, or refused with a reason (28 multi-ring geometries, 1 malformed), and the producer refuses to write if the three do not sum to the lines read. Two people repeated the cut with two methods; the provincial counts coincide.

Cells, classes, patches, towns. Footprints are counted in cells of 0.01° (about 1.1 km). Cells form five settlement classes by their count, scattered (1–9), hamlet (10–49), village (50–199), town (200–999) and city (1,000 and more); occupied cells that touch form built patches; fifteen towns are described within 11 km of their centre.

Confidence against density. The score is measured against local density at two scales: nationally in the 1.1 km cells, and inside five cities in blocks of one hectare. For each density bin: mean confidence, share below 0.75, median footprint area.

Read withDetected is not verified: a footprint is the model’s guess, qualified only by its score. The imagery has no published date, a roof is not a household, and under forest canopy the satellite does not see.

The GIA platform, which uses the cut dataset →

1. W. Sirko, S. Kashubin, M. Ritter, A. Annkah, Y. S. E. Bouchareb, Y. Dauphin, D. Keysers, M. Neumann, M. Cissé and J. Quinn. Continental-scale building detection from high resolution satellite imagery. arXiv:2107.12283, 2021.
2. Google Research. Open Buildings v3, polygons dataset, CC BY 4.0. https://sites.research.google/open-buildings/