Mapping Refugee Settlements With Open Data and Machine Learning
Refugee camps rarely start with a formal civic plan. They are stood up quickly in response to urgent displacement and often expand organically, which makes delivering even basic services a serious challenge for humanitarian organizations. In protracted crises, the issue becomes more acute: Kenya’s Kalobeyei settlement, for example, hosts more than 300,000 refugees from over 20 countries, and supporting a population of that size requires functioning roads, clinics, schools, and utilities.
The problem starts with a fundamental gap: none of that infrastructure can be mapped or maintained without spatial data. In many refugee camps, there are no street names, no consistent addressing, and no centralized records for where essential resources are located. Without that baseline, planning is reduced to guesswork.
UNHCR, the UN Refugee Agency, approached this as a problem open collaboration could solve. The idea was to pair local knowledge with a global network of technologists, and publish every dataset and tool openly so other humanitarian efforts could build on it.
From Aerial Imagery to Training Data
The effort began locally, with the people who know the settlement best. Refugees and other residents were trained by the Humanitarian OpenStreetMap Team (HOT) to fly drones over the camp and capture large volumes of imagery. A portion of those images were then manually annotated, marking features such as homes, solar panels, clinics, and sanitation sites. That became the reference dataset.
From there, the Microsoft AI for Good Lab trained machine learning models on those annotations to recognize the same classes of features across the rest of the settlement. What would have taken years of manual surveying was compressed into a far shorter window.
Open Repositories as the Collaboration Layer
The data collection was only one part of the project’s design. Equally important was how the outputs were distributed. Every dataset, trained model, and piece of code was made public on GitHub, which means the methods used in Kalobeyei can be replicated by teams working in other camps, disaster zones, or unplanned urban areas.
GitHub Copilot was also used in the workflow to streamline routine formatting and repository cleanup, which shortened the time between building the models and publishing a polished project for outside contributors to work with.
That open structure is what turns finished tools into an ongoing effort. Release on GitHub effectively issued an invitation for developers, data scientists, and civic technologists to adapt the code to new contexts or improve on what was already created. As one project lead put it, “GitHub is the beginning for the next phase of this project. Connecting developers to these kinds of problems is our mission.”
A Reusable Blueprint for Displacement Crises
The most durable outcome is not the maps of a single camp but the workflow itself. Refugees contributed the ground truth with training from HOT; researchers and engineers converted that into reusable models and code; and the public repository framework gives other humanitarian teams a starting point to do the same for other settlements.
The result is a cooperative, open source method for turning aerial imagery from a chaotic camp environment into organized plans — while keeping the end product available to anyone who needs it.
