Mapping Schools with Artificial Intelligence

Mark Altaweel


There are initiatives that are focused on using artificial intelligence (AI) methods to tackle pressing global problems. One of these is getting as many children on the planet to be in school. In a recent project using Google and open source tools, researchers are using AI to map schools in countries where many schools are undocumented so as to connect children with schools.[1]

The United Nations International Children’s Emergency Fund (UNICEF)’s Giga initiative has created a challenge to map all schools in the world. This is particularly important for lower income countries to have resources so as to know how and where to connect children to schooling. Under this initiative, over 23,000 schools from Kenya, Rwanda, Sierra Leone, Niger, Honduras, Ghana, Kazakhstan, and Uzbekistan have now been mapped.

Satellite imagery from Maxar, a company providing high resolution tiled images, was processed using an AI workflow that trains and experiments with Kubeflow, a tool that is used on Google’s Cloud Kubernetes Engine. The application also uses a tool called ML-Enabler, which provides AI-powered data inference. The use of these hardware and software tools has enabled the large set of satellite imagery covering wide regions to be searched for buildings identified as schools.

A section of Rwanda showing previously unmapped (yellow dots) schools.  Source: Development Seed.
Screenshot of an interactive map showing a section of Rwanda showing previously unmapped (yellow dots) schools. Source: Development Seed.

The AI convolutional neural network (CNN) models developed had to be adapted to different regions and even countries. The schools were classified using tile-based classifiers with six country-specific trained models, two regional models, and a more general global model.

Free weekly newsletter

Fill out your e-mail address to receive our newsletter!

So far, the mapping models have been used in eight countries in Asia, Africa, and South America. The algorithms have searched 71 million map tiles with 60 cm data resolution. The regional models outperformed the country-specific models, indicating models trained with more varied data are likely to be more accurate.

The shape, size, and general appearance of schools helped in classifying imagery. Data from existing schools that are known are used for training and validating the results. Supertiles are more high resolution data, which appear to perform the best for training data. The researchers have published their results and accuracy of their models, with the model to be released in an open format shortly. 

Examples of unmapped schools identified on satellite imagery. Source: Development Seed.
Examples of unmapped schools identified on satellite imagery. Source: Development Seed.

Challenges in Mapping Schools Using Artificial Intelligence and Satellite Imagery

The challenge is now to map in more countries and scale the approach to varied areas. A significant challenge is on the training data and validation side, where people need to be employed to provide and check training data and validate results, significantly slowing the process.

Additionally, it is clear that not all buildings used as schools appear as schools, at least from space. It will now be critical to identify the range and variety of school buildings used in regions across the planet, including informal sites, in order to map all schools on the planet, which is the ultimate challenge. 

Increasingly, we are seeing the use of AI for mapping, including such areas as helping to improve navigation in places where map data are often limited. In this case, only using satellite imagery, road features and buildings can be tagged and then mapped and classified using known, machine-trained features. In effect, the model used for mapping schools is comparable, although it has the challenge of trying to identify schools that may not be easily identifiable using imagery alone.

For researchers, the challenge is less computational but more on the data side.[2] Mapping efforts trying to map undocumented features or sites are challenging because it requires human resources to identify items for training and validating features.

The next wave of AI development may enable more general AI, which promises to create methods that require far less training and AI that is adaptable to different problems as machines can identify features simply by stating the problem rather than having to obtain a lot of training data.[3] However, researchers have claimed that we have made very little progress here and most AI efforts can be defined today as artificial narrow intelligence (ANI) rather than the more ambitious artificial general intelligence (AGI). The problems experienced by the school mapping efforts highlight the challenge.

Automated mapping tools are becoming common and many of these tools are helping to solve important social problems such as knowing how to connect schools with children who may not have schools accessible to them. While this may give a good public impression, researchers are also working on better ways to solve problems that require less resources, particularly in having to obtain a lot of training data and validate these results using human resources. We are not there yet but for now we can create specific efforts that target problems with the power of cloud computing, AI platforms, and human expertise. 


[1]    For more on the school mapping project supported by UNICEF and applying AI, see:

[2]    For more on using AI to map places with poor map data, see:

[3]    For more on general versus more narrow AI, see:  Fjelland, R. (2020). Why general artificial intelligence will not be realized. Humanities and Social Sciences Communications7(1), 10.


Photo of author
About the author
Mark Altaweel
Mark Altaweel is a Reader in Near Eastern Archaeology at the Institute of Archaeology, University College London, having held previous appointments and joint appointments at the University of Chicago, University of Alaska, and Argonne National Laboratory. Mark has an undergraduate degree in Anthropology and Masters and PhD degrees from the University of Chicago’s Department of Near Eastern Languages and Civilizations.