#building segmentation

[alphatree]

Building Footprint Detection/Segmentation in Africa

Building footprints for Africa are mainly available via the Google Open Buildings and Microsoft Building Footprints datasets. Both are known to be of limited accuracy and hard to refresh due to the very high cost associated with VHR imagery at a continental scale. The African Building Foot-Print (ABFP) layer of Atlas AI is a low-cost alternative computed from high-resolution multispectral base-map imagery refreshed on a monthly basis.

The images showcase the results on model inference over Nairobi, Kenya. The colored heat map shows model confidence, and white polygons infused, show the corresponding findings from the Google Open Buildings dataset. Both results are superimposed on Airbus VHR imagery over the targeted area.

Note the frequency of regions in the ABFP heat map for which no white polygons are present.

Disclaimer: the ABFP is not a commercial product. It is a layer produced for research purposes only.

[alphatree]

Building footprint map of the city of Brest, France. The layer is generated using the PROTOGEN task of the Geo-spatial Big Data Platform (GBDX) @ DigitalGlobe. The input image is a WorldView-2 multi-spectral acquisition of total coverage of 15km x 16km @ 1.4m spatial resolution. The output layer has an overall accuracy of > 90% evaluated by random sample selection against reference data. The core module concludes in approximately 1min 20sec. 

[alphatree]

Automatic building segmentation from DSM data. The method maps the nested iso-height regions on Max-Tree data structures and performs attribute filtering form target refinement with safe & pre-defined attributes and thresholds.