Predicting sprayable structures using machine learning
June, 2017

While incredibly detailed maps showing roads, buildings and rivers are available for most well resourced countries, the same is not true for the less well resourced countries affected by malaria. OpenStreetMap, via the Humanitarian OpenStreetMap ( and Missing Maps projects (, is helping to map some of the poorest places on Earth. In many countries, such as Swaziland, Botswana and Zimbabwe, there is nearly complete information on roads and buildings, providing a fantastic set of data to help plan malaria household intervention campaigns such as indoor residual spraying with insecticide, distribution of mosquito nets and mass drug administration. However, complete information relating to the building type is rare, making it impossible to distinguish residential from non-residential structures. Yet, this information is crucial for helping to plan and execute intervention campaigns targeted at households.

The DiSARM team is working to apply powerful machine learning to predict whether a building is sprayable or not based on features such as its size, shape, distance to road and roof type. Early results such that we can predict ~90% of structures correctly. By building this information into DiSARM, the platform allows a better estimate of the scale of operations required for a given campaign and helps to improve estimates of coverage.

OSM predication portrait

Maps showing the observed building classifications (above) and independent predictions from machine learning (below) in Mbabane, Swaziland

Back to Updates