For those of us who live in places where driving directions are available at our fingertips, it might be surprising to learn that millions of miles of roads around the world have yet to be mapped. For more than 10 years, volunteers with the OpenStreetMap (OSM) project have worked to address that gap by meticulously adding data on the ground and reviewing public satellite images by hand and annotating features like roads, highways, and bridges. It’s a painstaking manual task. But, thanks to AI, there is now an easier way to cover more areas in less time.

With assistance from Map With AI (a new service that Facebook AI researchers and engineers created) a team of Facebook mappers has recently cataloged all the missing roads in Thailand and more than 90 percent of missing roads in Indonesia. Map With AI enabled them to map more than 300,000 miles of roads in Thailand in only 18 months, going from a road network that covered 280,000 miles before they began to 600,000 miles after. Doing it the traditional way — without AI — would have taken another three to five years, estimates Xiaoming Gao, a Facebook research scientist who helped lead the project.

“We were really excited about this achievement because it has proven Map With AI works at a large scale,” Gao says.

Starting today, anyone will be able to use the Map With AI service, which includes access to AI-generated road mappings in Afghanistan, Bangladesh, Indonesia, Mexico, Nigeria, Tanzania, and Uganda, with more countries rolling out over time. As part of Map With AI, Facebook is releasing our AI-powered mapping tool, called RapiD, to the OSM community. RapiD is an enhanced version of the popular OSM editing tool iD. RapiD is designed to make adding and editing roads quick and simple for anyone to use; it also includes data integrity checks to ensure that new map edits are consistent and accurate. You can find out more about RapiD at mapwith.ai.

The power of a good map

“Many rural parts of the world are difficult to map on the ground. As I experienced in my previous work with the Red Cross, the challenges include remote locations, lack of power and internet access, and complicated economic and political environments,” says Drishtie Patel, program manager, Maps & Location Infrastructure at Facebook. “Map data gaps can affect everything, including disaster response, community planning, and helping the local economy.”

Earlier this year, Facebook data scientists and AI researchers also released AI-powered population density maps to help humanitarian organizations. Over the past year we have been getting requests from many organizations asking for AI-generated roads and populations density maps, so we have been testing our tools with them to make sure it meets their needs.

“The RapiD tool was developed in conjunction with those in the mapping community who have been working in this area for many years. Because this tool was built with their input, it is already having an impact,” says Tyler Radford, the executive director of the Humanitarian OSM Team (HOT), which aims to make sure OSM represents all parts of the world.

“RapiD is a big step forward toward meeting this goal,” Radford says. “By augmenting what was previously an entirely manual process — tracing of roads from satellite imagery — RapiD combines the best of machine learning with the best of human expertise. It supercharges mappers.”

OSM data is used across all sectors, including humanitarian organizations, government agencies, small businesses, and community groups. One of Facebook’s core goals is to connect people and ensure that everyone is represented on the map. Like other technology companies, we also use OSM in products and services. In fact, OSM is the basis for maps across the Facebook family of apps. So expanding the coverage of these maps allows us to better serve people in more places around the globe.

The Map With AI team is collaborating with HOT to add more features to RapiD. For one step in that process, they’ve integrated RapiD into a development branch of HOT Tasking Manager, which pairs volunteer mappers with specific areas to map, as an early experiment.

“We became close collaborators with the OSM community during our work in Thailand and have received many requests for sharing our AI-generated roads to help the community more broadly,” says Gao. “That’s one big motivation behind the launch of Map With AI: to team up with communities and map the world together.”

Using AI to help experts map faster and better

Map With AI uses a subfield of artificial intelligence called computer vision, whereby machines learn to spot complex patterns in images so they can analyze the same type of satellite imagery that OSM volunteers have worked with for years. The AI system has been trained to identify possible roads and highlight them in the mapping tool. The OSM volunteers use their expertise to review and then confirm or modify the AI system’s suggestions.

In this case, the computer vision system is what’s called a deep neural network (DNN) segmentation model. The actual output from the tool is an enhanced satellite image giving the probability that each pixel is part of a road. Bright magenta means there is high confidence of the pixel being a road, and transparent means there is low confidence.

Outside of some limited proofs of concept, this AI-powered mapping approach had never been deployed at scale before. Roads in different places vary greatly in width and color. Roads in satellite imagery can be partially obscured by trees, and there are features that might look like roads but aren’t, such as dry riverbeds or field walls.

To handle these challenges, the Facebook AI team built a 34-layer DNN model, which can recognize roads on satellite images around the globe with a resolution of roughly 2 square feet per pixel.

“Technology of this scale, complexity, and precision became available only in recent years,” says Danil Kirsanov, a Facebook engineering manager who helped build Map With AI. “This level of detail means it can spot unpaved roads, as well as alleys and even pedestrian pathways, and distinguish them from visually similar riverbeds or walls. We still have some false positives, of course, and this is where expert human judgment is needed.”

To build the computer vision model for detecting roads, the Facebook team took Maxar satellite imagery, traced roads by hand, and then used these hand-drawn images to teach the machine what was and was not a road.

After repeated rounds of training and evaluation, the system was ready for actual mapping. Based on the lessons learned by mapping Thailand, the Map With AI team was able to adjust their training strategy to generate a model that could work well at global scale. (More technical details are available in this Facebook AI blog post.)

How computer vision meets human expertise

After converting the output of the model into a format that mapping volunteers can use, the AI system’s predictions then serve as the basis for the mapping process. The AI does most of the work for the person creating the maps, so she or he only needs to fill in any gaps, double-check the accuracy, and select the appropriate road type. An extensive set of validation tools has been built into RapiD to help users catch and fix data issues in real time, thus improving the quality of the final submissions to OSM.

“The tool strikes a good balance between suggesting machine-generated features and manual mapping,” says Martijn van Exel, a longtime leader in the open mapping community. “It gives mappers the final say in what ends up in the map but helps just enough to both be useful and draw attention to undermapped places.”

The result has been a better way for people with mapping expertise to do their work.

“Automatic data validation checks [allow] me easily to spot necessary changes,” says Dennis Irorere, a research fellow at YouthMappers in Nigeria. “This has gone a long way to not only enable me to map faster but also improve the quality of road map data that I submit after mapping.”

Rendering a suggested road is very quick, he notes, which makes the Map With AI tool easier to use in areas with limited connectivity.

“In my own case, the tool has helped in such a way that the majority of the roads in a selected task are mapped and are mapped faster,” says Samuel Aiyeoribe, associate manager, Technical GIS, at eHealth Africa. “Once the AI predicts a road feature, this can be quickly and easily tagged into appropriate road features accordingly.”

Delivering the benefits of accurate maps to everyone

Now that Map With AI and RapiD are widely available, Irorere and others in the mapping community say they hope it will dramatically accelerate their work around the globe.

“I think [it] will go a long way to make mapping and creation of important geospatial data easier and faster,” Aiyeoribe says.

“The World Bank applauds the use of AI-assisted feature extraction to help create worldwide geospatial data. We look forward to the expanded use of open geospatial data to enable data-based policy in developing countries,” says Walker Kosmidou-Bradley, Geographer at the World Bank.

RapiD is already making an impact. “This is definitely going to be a key part of the future of OSM. We can never map the world, and keep it mapped, without assistance from machines,” van Exel says. “The trick is to find the sweet spot. OSM is a people project, and the map is a reflection of mappers' interests, skills, biases, etc. That core tenet can never be lost, but it can and must travel along with new horizons in mapping.”

We'd like to thank Anil Batra, Charmaine Bonifacio, Benjamin Clark, Nicholas Holroyd, Danil Kirsanov, Zack LaVergne, Patty Lawler, Yunzhi Lin, Peter Ouyang, Zvone Sparovec, and Jason Sundram for their contributions to this project.