We trained a building detector model using zoom 18 imagery from Digital Globe and a standard convolutional neural net from the TensorFlow deep learning library. This algorithm attempts to define a bounding box for each separate building in training data obtained from OSM. These bounding boxes are simply summed to estimate the total building count. Qualitatively, this approach seemed to better detect buildings than the Skynet model though it does not provide a per-pixel mask needed to compute total building footprint.

Conclusions

This approach gave fairly good results when trained on DG’s Vivid imagery and building footprints obtained from OSM. We developed this model on a very small training data set (~100 images), so we expect that there is room for improvement given more time to gather training data and further tune the model to work in many regions of the world. In terms of compute power, We believe that this model can scale up to multiple zones fairly easily assuming there are some light cloud computing resources (e.g., AWS EC2) to support it. One challenge we encountered, however, is that buildings are sometimes constructed physically up against each other making the task distinguishing separate structures using remote sensing data difficult by any algorithm (or human). Therefore, we expect that building counts will represent building complexes, not necessarily the total number of individual buildings. Another potential obstacle is distinguishing between residential and industrial buildings, but we would need to discuss how important this distinction would be moving forward in Phase 2.