How AI-powered maps help improve vaccination campaigns and rural electrification

To generate these maps, a team of data scientists and AI researchers in the Facebook Boston office took a creative approach. They used publicly available population data and commercially available high-resolution satellite imagery, and then applied machine learning techniques to map hundreds of millions of structures distributed across vast areas.

Digital volunteers with the Missing Maps project — an initiative co-founded by the American Red Cross — used these AI-powered maps to filter out the 97 percent of the terrain in Malawi that is entirely uninhabited. Volunteers with Missing Maps, which is a collaborative effort by humanitarian organizations to map parts of the world that are most exposed to natural disasters and other crises — could then concentrate their efforts on mapping the remaining 3 percent, knowing that they weren't overlooking any small, remote communities. With guidance from the maps, Red Cross volunteers were able to locate communities so they could answer questions about the vaccination process and address the concerns of those who otherwise might not have brought their children to be immunized.

“Facebook’s high-resolution population maps have supported Humanitarian OpenStreetMap and Missing Maps’ mission of putting the world’s most at-risk places on the map,” says Tyler Radford, executive director of the Humanitarian OpenStreetMap Team, which participates in the Missing Maps project. “The maps from Facebook ensure we focus our volunteers’ time and resources on the places they're most needed, improving the efficacy of our programs.”

Knowing exactly where help is needed

Laura McGorman, a public policy manager at Facebook, says her previous work with international development efforts convinced her that the mapping project could have a significant impact in building partnerships with NGOs and relief organizations.

“Having started my career at USAID working on malaria control, I have witnessed firsthand the critical role that accurate data play in the effectiveness of humanitarian efforts,” she says. “What’s exciting about working on projects like these at Facebook is that it provides an opportunity for our company to contribute to these efforts through our expertise in data and machine learning, opening space for our partners to focus on service delivery and impact.”

The Boston-based team uses advanced computer vision and machine learning techniques to combine satellite imagery from Digital Globe with publicly available population data to create detailed population density maps of Africa.

Facebook started developing population density maps in 2016 to provide better tools to support connectivity efforts around the world. However, the team realized these maps could also function as a powerful public good by helping with public health, rural electrification, and disaster response. The project now aims to keep adding new continents and countries.

When former American Red Cross employee and Missing Maps co-founder Drishtie Patel joined Facebook in 2016 to work on this project, the two organizations realized that Facebook could offer significant help to improve the efficiency of volunteer mapping for programs in the humanitarian sector.

In Tanzania, Facebook’s AI-powered maps helped kick-start efforts to bring renewable electrification to rural areas. Fewer than one in three people in Tanzania has access to electricity. There are considerable gaps in energy supply to the rural population, making it difficult to supply power from one centralized source. As part of the mini-grid component of the Scaling-Up Renewable Energy Program (SREP) funded program in Tanzania, the International Finance Corporation (IFC), in close collaboration with the Rural Energy Agency (REA) of Tanzania, commissioned a consortium including Humanitarian OpenStreetMap Team (HOT), Reiner Lemoine Institute (RLI) and INTEGRATION Environment and Energy on a project to better understand which locations would benefit from decentralized energy solutions. This was achieved by combining our high-density population maps with detailed data on settlement locations and structures from OpenStreetMap.

Humanitarian OpenStreetMap team personnel then traveled to villages identified as high priority and conducted surveys aimed at understanding the local populations’ electricity needs. The results of these surveys were provided to agencies involved in rural electrification, helping mini-grid operators choose the most appropriate locations to begin the work.

A project that’s perfect for AI

High-resolution satellite imagery already exists for much of the world. However, prior to Facebook’s mapping project, it required countless hours of digital volunteers' time to comb through millions of square miles of pictures to identify which contained a tiny town or remote village.

The Facebook team used AI to solve the problem. They took those commercially available high-resolution satellite images and split them into millions of roughly 100-foot-by-100-foot squares — each just a little bigger than a baseball diamond. They then worked with a team of experts to label examples that could be used to teach our computer vision system to accurately recognize all the structures in those images. Because building styles can vary greatly from region to region, the labelers created separate training examples for each country in the project. They also reviewed the results to ensure accuracy.

Given the huge scale, it was important to create a system that could work efficiently. For Africa alone, for example, the system examined 11.5 billion individual images to determine whether they contained a structure. Their approach found approximately 110 million structure locations in just a few days.

Once the researchers had mapped the structures, they could then use public census data in order to accurately extrapolate the local population density.

“With just the census data, the best you can do is assume that people live everywhere in the district – buildings, fields, and forests alike,” says James Gill, a Facebook engineer who worked on the project. “But once you know the building locations, you can skip the fields and forests and only allocate the population to the buildings. This gives you very detailed 30 meter by 30 meter population maps.

“The satellite data for the world is 1.5 petabytes,” says Gill. That’s enough to fill more than 1,000 hard drives on a typical laptop. “How do you store that, how do you move that around, how do you run models on that? Algorithms designed for single machines — or even small numbers of machines — don't work at this scale. We had to devise parallelized versions of these algorithms. This allowed us to utilize Facebook's infrastructure for storage, transport, and compute. There are only a small number of companies that can do this at this scale.”

Responsible sharing

Building tools from satellite imagery and national census data shared with Columbia University's CIESEN allows Facebook researchers to share their data science and compute power with the world. Sharing these products openly and responsibly creates opportunities for the humanitarian sector to improve the health and safety of communities around the world. The methodology for these maps was developed in close coordination with experts in geographic and demographic data at Columbia University’s CIESEN, and we’ve worked closely with partners like Humanitarian OpenStreetMap to make sure the maps are used to help humanitarian projects.

“When I talked with humanitarian organizations about their technical challenges, it became clear that obstacles that were insurmountable to them could be done on Facebook infrastructure in a month,” says Gill. “By leveraging our skills and resources, we can provide groups like Humanitarian OpenStreetMap with the information that lets them focus on what they do best: helping the world.”

For more on the technical work done to create these population density maps, read our companion blog post on the Facebook AI blog here.