During the 2019-2020 pilot supported by Microsoft, 18 million building footprints were automatically extracted from satellite imagery for all of Tanzania and Uganda. HOT found that on average, mappers working without AI assistance could map between 1000-1500 buildings per working day. For areas with high-quality AI output, providing mappers with AI-generated building footprint suggestions increased this rate to up to 2500-3000 buildings per day approximately doubling the rate at which building data could be added to OpenStreetMap, which is the crucial link for making it available to the humanitarian information management community.
While doubling mapping efficiencies (100% efficiency gain) are promising, one of the greatest challenges is making sure data is converted from AI/ML to OpenStreetMap in a rapid yet responsible community-centric way (respecting existing data contributions already in OpenStreetMap). This project enabled us to take the ‘next step’ after receiving the building predictions from Microsoft by building better tools for data conflation. This is expected to dramatically reduce manual intervention. By doing this, human mappers can focus their time and skills on value-added activities: ground-truthing and validating predictions and adding local knowledge to the map not visible from satellite imagery (such as place names, location of key lifeline infrastructure, etc.
We’ll describe our mapping workflow and progress updates for Kenya and Nigeria, and the specific challenges met with this dataset extracted through machine learning will be explained. Considering the growing availability of such AI-related datasets, we’ll review common errors and how we adapted to them and to other issues such as imagery offsets, heterogeneity of existing data and other context specific challenges. Eventually we’ll propose recommendations regarding this type of editing and aspects to consider before starting such imports and how to get the community involved.
