Connecting the world with better maps

This post details the development of a high-resolution (5-meter) population distribution dataset for 20 countries by applying computer vision techniques to satellite imagery. The process involved identifying human-built structures as proxies for population density, using a combination of image preselection, a modified Facebook image-recognition engine (deep convolutional neural network), and a weakly supervised neural network. The resulting data was used to redistribute census counts, providing granular insights for planning short-range access networks and backhaul solutions for internet connectivity in underserved areas. The scale of the operation involved processing 14.6 billion images on thousands of servers.