How maps built with Facebook AI can help with COVID-19 vaccine delivery

April 15, 2021

As work continues to vaccinate most of the world’s nearly 8 billion people against COVID-19, accurate population maps are urgently needed. This need is especially great in low-income communities, where population data may be poor or out of date. Today, we’re releasing updates to our population density maps using new methods, starting with Pakistan and Scandinavia, countries where the original methodology proved insufficient in correctly identifying structures.

Two years ago, we launched the highest-resolution population density maps for nearly every country and territory in the world. Our maps have since been downloaded millions of times by nonprofits, researchers, and governments to inform the delivery of critical services around the world. For example:

  • After Cyclone Kenneth struck Mozambique in 2019, Direct Relief, the World Health Organization, NetHope, and Harvard School of Public Health used our high-resolution population density map to determine the number of doses of cholera vaccinations that would be needed in areas affected by flooding in Mozambique. Our maps helped inform resource allocation for the vaccination of more than 250,000 children.

  • Saha Global, a nonprofit that provides access to safe drinking water to rural communities in the Northern Region of Ghana, used our population density map to pinpoint villages that were without access to water and sanitation, and extended its services to an additional 7,000 people.

  • In Kenya, our nonprofit partners Cadasta and Pamoja Trust used the Facebook population density map to determine the number of people who were at risk of eviction. This information facilitated the decision to halt construction that would have displaced over 70,000 people.

Accurate population density data is critical when performing vaccination campaigns, and our newest population density maps will be particularly useful for the delivery COVID-19 vaccinations in the global south. Previously, the American Red Cross and the Ministry of Health in Malawi embarked on a measles vaccination campaign designed to cover 95 percent of households within a limited time frame. Using Facebook’s population density map, they quickly discovered that 97 percent of the landmass of Malawi was uninhabited and could be removed from any house-to-house messaging plans. As a result of consulting our maps, Red Cross volunteers were able to complete 100,000 home visits in three days. We anticipate that these same mobilization approaches can be leveraged when working to vaccine populations at scale in the global south and are working with our partners around the world to assist in this effort.

Facebook’s high-resolution population density maps use a mixture of machine learning techniques, satellite imagery, and population data to map hundreds of millions of structures distributed across vast areas and leverage this information to extrapolate the local population counts. The satellite maps used in this project are generated using commercially available satellite images from Maxar and national census data for each country, which comes from Columbia University’s Center for International Earth Science Information Network (CIESIN), which collaborates with Facebook on this project.

Normally, we use a convolutional neural network (CNN) classifier and determine a range from zero to one, above which we consider an image to contain a building, and below which an image very likely does not contain a building. However, if many false positives get labeled as having structures, this classification threshold falls to such low levels that false positives occur everywhere in the region, putting the accuracy of the entire map into question.

Map of Pakistan with false positives from the original model in blue, and fewer false positives from the new digital surface model in green.

In the case of Pakistan, in certain regions, we had difficulties distinguishing actual structures from false positives because of the unique characteristics of the terrain. In Scandinavia, the imagery itself was distorted, leading to issues with the output population density calculations. Additionally, in the original methodology, our classifier also produced false positives in Nordic boreal conifer forests, where the trees’ growth characteristics were being mistaken for human-built structures.

To improve our algorithms in these settings, we added an additional classifier that relies on a digital surface model called Advanced Land Observing Satellite (ALOS) as an input, which was generated based on synthetic aperture radar (SAR) data. In contrast to the satellite imagery of the visible spectrum, represented in red-green-blue (RGB), that we use for our established building classification, radar is not impeded by the presence of clouds and the data does not suffer from the same distortions as RGB imagery. We train this new surface classifier with extracts at known building locations from OpenStreetMap (OSM) as positive examples, as well as extracts at locations with very low CNN classification scores as negative examples. OpenStreetMap, licensed under the Open Database License, uses a volunteer mapping force to confirm the presence of structures in its data set, so we rely on their platform to ensure that the new classifier performs well in identifying true households.

Visualization from Karachi using the Advanced Land Observing Satellite digital surface model:

Settlement example of the ALOS digital surface model (structures) in Pakistan.
Mixed landscape example of the ALOS digital surface model (no structures) in Pakistan.

Tiles in which Facebook model predicts a structure in green, combined with tiles in which OpenStreetMap predicts a structure, which are subsequently added to the Facebook Population Density Map.

The input values for this new classifier are fast Fourier transform matrices that we feed into a simple CNN that identifies configurations of height differences in the surface model. We are able to predict that there are structures in areas with blocky images that have predictable height differences, compared with more random height differences of areas that contain hills or trees. As the surface model has a resolution of about 30 meters per pixel, we use extracts of about 450x450 meters, represented by a 15x15 matrix of elevation values.

With this added classifier, we can work with reduced CNN thresholds and weed out the false positives in the original population density model. In addition, we add the buildings already identified by users of OpenStreetMap to prevent the CNN model from missing structures. The result is a significantly more accurate map of where structures truly are and where people truly live. We hope that these new maps can assist in the delivery of life-saving services to populations in need in Pakistan, Scandinavia, and around the world.

Written By

Dr. Andreas (Andi) Gros

Technical Lead

Prateik Mahendra

Data Engineer

Rebecca Raezada

Data Engineer

Laura McGorman

Public Policy Manager, Data for Good