How refugees helped Microsoft use AI to map one of the world’s largest camps

Microsoft has been working with humanitarian partners and residents of the Kakuma refugee camp in Kenya to upgrade an AI-powered mapping system, in a project that puts people living in the camp at the centre of the technology.

Kakuma is home to more than 300,000 displaced people from over 20 countries and covers roughly 15 square miles. Keeping maps up to date in a place that large, and that dynamic, is a constant challenge. Microsoft said the project was designed to improve how services are planned and delivered by combining local knowledge with machine learning.

The work brought together several groups with very different roles. UNHCR’s Hive team helped define what the camp actually needed from better maps. The Humanitarian OpenStreetMap Team, or HOT, handled on-the-ground data collection and community engagement. Crucially, much of that work was led by people who live in Kakuma themselves.

Local residents helped introduce the project, flew drones over parts of the camp and manually identified buildings and infrastructure in the imagery. “I think collaboration was key because each person brought something unique to the table,” said Dr Simone Fobi Nsutezo, an applied research scientist at the Microsoft AI for Good Lab.

Those manual efforts mattered. HOT teams tagged around 10 square miles, or 16 square kilometres, of drone imagery by hand. That data became the “ground truth” used to train AI models, making sure the system reflected the reality on the ground rather than assumptions made from afar.

Microsoft’s AI for Good Lab then used that community-labelled imagery alongside Azure cloud services to train machine learning models. The models can now detect things like buildings, sanitation blocks, solar panels, streetlights, rooftops and parts of the camp’s power network.

“Once you’ve trained a model on a small amount of data, it’s very fast to get predictions on new areas,” said Dr Amrita Gupta, an applied research scientist at the Microsoft AI for Good Research Lab.

Importantly, Microsoft said it has released all of the models and datasets as open source on GitHub. The idea is to let developers, researchers and humanitarian organisations reuse the code, improve it and keep the data up to date as Kakuma continues to change.

The Recap

• Residents led data collection to upgrade Kakuma’s AI maps.
• Teams manually tagged 10 sq miles (16 km²) of imagery.
• Models and datasets released open source on GitHub.