A team at Facebook has detailed how it used artificial intelligence to analyse billions of images to create extremely accurate population maps which could be used bring internet connectivity to those currently off the net.
Ten percent of the world's population live in parts of the world where no connectivity is available, but working out which technology can be used to get them online requires accurate information on where people live.
Understanding how communities are connected to each other is one key element: villages along a river or road could be connected by terrestrial point-to-point links, but scattered settlements might be better served by drones or satellites.
To help solve this problem, Facebook's Connectivity Lab has built population maps for 20 countries, using techniques from computer vision to analyse high-resolution satellite imagery.
First the lab used a conventional image-processing procedure to select areas that might contain human structures, discarding images with vast bodies of desert, forest, and water. Next, it used Facebook's image-recognition engine -- based on a deep convolutional neural network that provides a fixed dimensional feature embedding for all images -- and trained this to detect whether a satellite image contained a building.
They then used a "weakly supervised" neural network to identify the outlines of buildings. The researchers said neural networks typically need to be trained on large volumes of images to obtain sufficient accuracy but by using this approach they were able to reduce the number of training images to about 8,000 binary satellite shots from within one country.
The researchers analysed 20 countries - 21.6 million square kilometers and 350 TB of imagery. One pass of their analysis saw their convolutional neural nets process 14.6 billion images, typically running on thousands of servers simultaneously.
The Connectivity Lab will release the data publicly later this year and said: "We believe this data has many more impactful applications, such as socio-economic research and risk assessment for natural disasters."