Researchers from the Facebook Connectivity Lab and MIT are working on a project that can generate street addresses from satellite images by learning and labeling regions, roads, and blocks. The reason why such projects are important from a geospatial point of view is because they can enhance the precise physical presence of an area, which will eventually increase the connectivity of all around the world.
The project – Robocodes – uses an algorithm that extracts information from satellite imagery by utilizing deep learning. The system uses satellite images for predicting roads. Then the road predictions are processed to extract road segments. The segments are clustered to obtain regions. The roads are named according to regions and ordering, and the houses are named according to road-oriented axes.
The remote sensing approach of Robocodes algorithm is sensitive to unpaved roads, and less structured urban spaces. The project is also trained to deal with existing map cases where even the ground truth segmentation is fundamentally wrong or disconnected.
According to Chris Sheldrick, CEO, what3words, 75% of the world lacks street addresses, this means that globally, four billion people are, quite literally, not on the map. Hence, the relevance of this project becomes all the more crucial to bring those 4 billion people on maps. But to make the project successful, the system requires training to read satellite images for predicting roads and other things.
Now to understand the use of this project, we will have to first understand that if you go by the presently available methods, you will have to first identify an unknown address on map that involves geocoding for which latitude and longitude information is required. But because sometimes it gets difficult to read some of these settlements that are adjoining to some roads and junctions, information extraction becomes challenging. To overcome such problems, the project employs deep learning on satellites images to have a clear view on the subject.
Now, for example, across the world, the naturally occurring addresses are usually the result of cultural dynamics, politics, economies, and other long term processes adopted by urban authorities. To mimic this organic process, while maintaining a unified representation that is independent of the human factors, an appropriate scheme was needed. To find a solution for this, the team conducted a research on the current addressing methods in many countries and came up with a design that can ease the understanding of both humans and machines.
Under Robocodes, the process of generating maps involves a five step process in which, the system first uses satellite images to highlight the roads from those images, then it divides the map into regions and highlights the roads and regions and compare it with data from OpenStreetMap of the same area.
So, if the project succeeds, it will eventually lead to more coverage of addresses, connecting the invisible population to the world, and will increase their contribution to humanity in developing countries. Connecting the unconnected will increase economic, juridical, and life-sustaining involvement of people of all around the world.