To estimate total building footprint in each SEZ, we also use Skynet, which provides a per-pixel prediction of buildings. Using these building mask predictions, the overall area (in m^2) of buildings for each zone can be calculated. Like in the Transport Infrastructure indicator, this trained model is used to predict building footprints on multiple dates from discrete scenes sourced from Digital Globe's IDAHO image tiles and deployed on cloud computing resources. Skynet works on Web Mercator projected tiles, so predictions corresponding to these tiles are first projected into an equal area using a projection such as the Albers projection. After applying this projection, every pixel corresponds to an identical unit of area on the ground, so the pixels can be simply summed and multiplied by that value to obtain total building footprint.

Conclusions

Skynet’s building predictions are noisier and less accurate than the predicted roads (likely because it wasn’t designed for this task). We found that predictions were inconsistent and missed large buildings with certain characteristics. This is likely because the model was trained on other regions with buildings that have different characteristics. For example, the zones where Skynet performed particularly poorly had white roofs (e.g., Bole Lemi), which wasn’t a feature well-represented in the training data. Again, this is an area we are currently working to improve as we develop the next generation of Skynet for both roads and buildings.