Machine learning for building stock modelling

A method to map buildings through time with machine learning and aerial photos

Francis Barre

Scientist / PhD Candidate

Nils Dittrich

PhD Candidate – NTNU

Zoé Cord’homme

Research Assistant – NTNU

Building stock modelling often relies on top down approaches, implying assumptions on building lifetime and societal dynamics. However, with the growing accessibility of aerial imagery, as well as the increasing performance of segmentation models, it is now possible to retrieve information on the past building stock from a bottom up approach.

Independent machine generated data
Local  authorities use orthophoto projects for purposes like urban   planning and forestry.

Timeseries with rich history
These orthophotos are available since the advent of cheap   photographic film – reaching back to 1935 in some places in Norway.

Coverage follows human settlements
 Most (historic) projects center around the built environment of municipalities.

A building segmentation model was trained on the Norwegian dataset, using the most recent aerial photos combined with the cadastre data as a label. The model was then used to detect buildings on older photos, to provide the building stock in Norway for the last decades.

The building locations for different years allows for several studies:

  • Analysis of demolition and renovation patterns
  • Impact of natural hazards on the build environment
  • Mapping urban development and land use from buildings
Scroll to Top