Facebook and friends used software to identify buildings in commercially available satellite images, and then estimated population using census data and a few other surveys and programs. Convolutional neural networks powered a model capable of identifying individual buildings in images from across the world.
"There has been a lot of work recently on neural networks that can recognize individual buildings with very high accuracy, but these models are finely tuned on the local characteristics of the region where they are trained," the Connectivity Lab's Tobias Tiecke writes. "We found that these models do not perform well at a global scale with realistic amounts of training data. Therefore, we developed a model that is trained only on binary labeled data -- if an image does or does not contain a building -- which allows for fast acquisition of the training data."
Facebook's preliminary findings suggest that of the 23 countries it studied, 99 percent of the population lives within 63 km of the nearest city, with 44 percent of people living in cities.
"Hence, if we are able to develop communication technologies that can bridge 63 km with sufficiently high data rates, we should be able to connect 99 percent of the population in these 23 countries," Tiecke says.
Facebook has been toying with a handful of options for rural communication technologies, including the shoebox-size OpenCellular platform and a satellite initiative. The population maps will help Facebook streamline its designs.
"We would love to hear ideas (preferably also in an open manner) on how to bridge this distance in an economically viable fashion for line of sight and non-line of sight use cases, and we hope that by sharing these datasets we will stimulate further research," Tiecke says.