How AI applied to HighRes Satellite Imagery can help OSM Mappers

At Facebook, our mission is to make the world more open and connected. Through collaboration with the OpenStreetMap community, we believe we can enable greater transparency, deliver the best local experiences, and facilitate innovation at scale.

OpenStreetMap is the largest open map of the world with more than two million contributors, and it has been integral to popular Facebook product experiences like live maps (https://www.facebook.com/livemap/), as well as check-ins and place pages for countries such as Japan and South Korea. However, there are still parts of the world in which the map quality varies. Frequent road development and changes can also make mapping challenging, even for developed countries.

At Facebook, we’ve utilized crowdsourcing and AI to investigate and research new mapping techniques. While there have certainly been meaningful advances in computer vision research on extracting roads from aerial imagery, mapping the developing world still poses significant challenges due to factors such as poor road conditions and a wide range of road types. As illustrated in the images below, a large variance in roads types and conditions can be observed between Egypt, India, and Rwanda. In the figure of Egypt on the left, the roads are unpaved, grey, and blended-in with buildings. India’s roads on the top right are mixed with greenery, and can only be identified by fences. In Rwanda on the lower right, there is a full range of dirt roads comprised of major roads and narrow trails for pedestrians. Thus, it’s unclear as to which roads should appear on the map, as well as what type.

Figure1

Figure 1. There’s wide variance in the appearance of roads in developing countries, making it very difficult, even for humans, to draw the map correctly. Town in Egypt on left, village in India on top right, and town in Rwanda on lower right.

In partnership with DigitalGlobe, we are currently researching how to solve this problem by using a high resolution satellite imagery (up to 30cm per pixel). Combined with the use of advanced deep neural net models, we have been able to train models accurate enough to detect roads automatically in these countries. The figure below shows road masks generated by our models in Egypt and Thailand.

Figure2

Figure 2. Road detection near a town in Thailand (top), and for a village in Egypt (bottom).

Beyond generating road masks with deep neural net models, we have applied a novel algorithm to extract OSM-compatible road vectors from the raster mask. We first apply a threshold and extract road segments using a Voronoi diagram, and then connect road segments using the shortest path algorithm. Once that is completed, we trim any remaining disconnected road segments.

Prior to being uploaded to OSM, the machine learned roads are merged with the existing OSM roads in order to avoid duplicates. During this process, a team of editors manually verifies each road segment to make sure that they comply with the OSM guidelines and conventions. At that point, the OSM communities feedback has proven to be critical in helping us submit correct edits and helped us catch errors.

Figure3

Figure 3. Upload AI generated roads through OSM iD tool, modified to highlight generated roads in green to help editors

For small geographical areas, this technique has allowed our team to contribute additional secondary and residential roads to OSM, offering a noticeable improvement in the level of details of the map. Below is an example of before/after images of road geometries with OSM in Egypt.

Figure4

Figure 4. One area in Egypt where we submitted roads (left: before, right: after)

While mapping road locations is part of the challenge, road naming is another. We are in the midst of building an integration with our crowdsourcing platform that allows people to name roads near them based on their local knowledge. Our proposal is to start by generating candidates of road names from POIs that are on currently unnamed roads and then ask people to vote for the correct name. We have not yet submitted any crowdsourced road names since we are still analyzing these results, but we estimate that this approach could for instance improve road name coverage in Egypt by 20% - 30%.

We hope to continue developing this partnership with OpenStreetMap, and learn more about how we can use our platform to help improve the map. Over the last few months we’ve been learning a lot and are thankful for the feedback from the community. Our team will be at the State of The Map US conference July 21st - 23rd (http://stateofthemap.us/program/) to share more about the research we are doing around how AI can help us make maps.

  • Pierre, from Facebook