Median predicted property value, 100 m cells
Each view is split into fifths, from the highest to the lowest predicted values. The map shows relative classes only. Cells with fewer than 5 buildings are left out.
How these maps were made
A gradient-boosted decision tree model (LightGBM) learned the relationship between the official property valuations of about 172,000 buildings and satellite and geospatial data: building footprint size and shape, building height, night-time lights, vegetation, and distance to roads, schools, shops and other amenities. It then predicted a value for each of the 2.1 million buildings detected in Google Open Buildings.
Cell and parish values are the median over the buildings they contain.
0.72Rank correlation (cross-validated, buildings)
2.1mBuildings
89,490100 m grid cells
265Parishes
Client: International Growth Centre (IGC) & Kampala Capital City Authority (KCCA). Read the IGC publication.