It appears that the best population density models from Facebook are also the most accurate ones ever made. The maps created by the major company’s machine learning system can identify structures built by humans with a resolution of five meters. This amazing feature could be used for improving the responses in case of natural disasters, but also for research and for expanding rural Internet access.
The density model from Facebook is one of the AI machine learning technologies that have been used in the past to filter email, translate various languages and recognize objects in pictures. However, the major company has found a way to use its image-analysis technology to create maps showing where people live, but also determine the factors that impede them from connecting to the Internet.
Until now, twenty countries have been successfully mapped, as the technology that usually powers tagging features on the social media platform is also capable of recognising buildings from remote or rural regions.
This endeavour is part of the Connectivity Labs effort from Facebook, whose purpose is to monitor technical aspects of the Aquila drones and Internet.org. In order to achieve this result, a team of researchers had to work with others from machine learning, data science and artificial intelligence. Together, they analysed about 350 TB of images covering more than 21.6 million square km.
The technology will surely prove useful in the attempt of providing Internet to more people from developing countries, be it through cell or wireless tower connections or through lasers, satellites or drones.
The maps were created through a complex process that involved discarding pictures with non-inhabited areas like water, forests or deserts. The next step was to detect the buildings in the images. Subsequently, the technology had to learn to identify those man-made structures without filtering desolate areas like before.
Among the countries mapped by Facebook are Mexico, Turkey, Ukraine, Egypt, South Africa, India, Algeria, Kenya, Ivory Coast, Sri Lanka, Uzbekistan, Ethiopia, Madagascar, Uganda, Cameroon, Tanzania, Burkina Faso, Ghana, Mozambique and Nigeria.
According to Descartes Labs chief executive Mark Johnson human populations data is both slow and very costly to obtain, and thus the machine learning system will prove to be of immense value.
The best population density models from Facebook are also expected to serve to a multitude of future purposes, including risk assessments for many natural disasters, aiding socioeconomic research and validating census data.
Image Source: The Verge