Open Source Population Density Models

When Facebook needed population density data in order to best position their satellites, drones and terrestrial networks for optimum connectivity, the first thing they learned was how inadequate the existing data was. So they set out to transform their image recognition tools – which had been developed for people, pets and other picture subjects – and trained them to recognize man-made structures such as houses and other buildings as evidence of human population. After analyzing 14.6 billion images covering 21.6 million square kilometers, and cross-referencing the results with existing census counts, what they’ve created in a population map with five-meter resolution for 20 countries.

Now, Facebook is working with Columbia University to make the data publicly available for the myriad of government agencies, mobile operators, researchers and other organizations that might want access to it. The information could prove to be invaluable for such things as optimizing humanitarian relief efforts after a natural disaster. As satellite imagery continues to improve, the technology can also be used to look backwards at earlier satellite data and monitor worldwide trends.