By operating on the ground in Kenya, the HOT team brought the human touch. The actual data collection activity was entirely led by locals, from introducing the project to flying the drones, and refugees in the camp helped manually identify features in the area. HOT teams flew drones to capture high-resolution aerial imagery and conduct field validation with refugee mappers. But most importantly, they trained and empowered members of the refugee camp—turning data collection into community engagement.
Refugees became mappers, interpreters, and stewards of their own environment, creating ground truth data. Their local knowledge added depth and accuracy, giving them ownership of the process. HOT’s team manually tagged 10 sq miles (16 km²) of imagery, creating a rich training dataset for AI development that can be kept current as the camp evolves.
“AI was primarily used for pattern matching and time saving. It helped us find signals in the data that would be hard to spot manually,” says Dr. Gupta.
Drawing on the rich, community-tagged imagery collected in the refugee camp, Microsoft’s AI for Good Lab developed advanced machine learning models using Azure cloud services. These models were trained to accurately identify a wide range of features—buildings, sanitation blocks, solar panels of streetlights and rooftops, and elements of the power network like poles and lines—reflecting the camp’s diverse and irregular landscape.
By leveraging both local expertise and AI, the team was able to overcome the challenges posed by the refugee camp’s unique structures, enabling rapid analysis and pattern recognition that would be difficult to achieve manually. All models and datasets were released as open source on GitHub, empowering developers, researchers, and humanitarian organizations worldwide to build on this work and adapt it for other communities in need.