A new benchmark dataset and segmentation model tackle one of urban AI's least glamorous problems: mapping the parts of cities that official maps often miss.
Researchers built SLUM-i, a semi-supervised framework for identifying informal settlements in satellite imagery, and paired it with a new benchmark covering roughly 900 square kilometers across Lahore, Karachi, and Mumbai, plus four cities in Sub-Saharan Africa and Latin America. The technical hurdle isn't just label scarcity — informal and formal housing often look nearly identical in aerial imagery, and existing annotations frequently contain noise. SLUM-i addresses this with two components: a dynamic confidence threshold that prevents the model from ignoring underrepresented classes, and a filter based on Meta's DINOv2 model that strips out-of-distribution training tiles before training begins. Tested across seven cities and five random seeds, the framework beat state-of-the-art semi-supervised baselines by up to 5.9 percentage points on a standard segmentation metric, with both components adding no overhead at inference time.
Informal settlements house hundreds of millions of people globally, yet accurate maps of them remain scarce — partly because producing them requires expensive manual annotation of satellite imagery at scale. A framework that can work with sparse labels and generalize across three continents lowers the practical barrier for governments and NGOs trying to allocate services or plan infrastructure in areas that formal urban datasets routinely ignore.
The architecture-agnostic design is a deliberate hedge: teams can slot these components into existing pipelines without rebuilding from scratch. The less tidy question is whether any of this exits the research literature and reaches the planners who actually need it.