The Satellite Can't See the Settlements That Need It

Automated building footprint extraction from satellite imagery works well in North American and European cities — achieving >85% F1 scores — but degrades dramatically (to 40–60% F1) in informal settlements in Sub-Saharan Africa, South Asia, and Latin America. These are precisely the areas where building maps are most needed for disaster preparedness, utility planning, and population estimation. The gap stems from training data bias (most labeled datasets come from Western cities), structural differences (informal buildings are smaller, irregularly shaped, and use non-standard roofing materials), and contextual interference (dense vegetation, narrow alleys, sheet-metal glare).

Over 1 billion people live in informal settlements, and this number is growing. Accurate building footprint maps are essential for disaster risk assessment (earthquake, flood), disease vector mapping, infrastructure planning, and census enumeration. Manual digitization by humanitarian mappers (e.g., Missing Maps project) is slow — mapping a single city can take thousands of volunteer hours. Automated extraction could accelerate this by 100×, but only if it works in the contexts that matter. The Open Cities AI Challenge (DrivenData, 2020) specifically targeted this gap and found that top competition solutions still struggled with generalization across African cities.

Standard semantic segmentation models (U-Net, Mask R-CNN) trained on SpaceNet datasets (primarily US cities) transfer poorly to informal settlement imagery. Fine-tuning on local data helps but requires labor-intensive labeling in each target area, which defeats the automation purpose. Domain adaptation techniques (style transfer, adversarial training) can close part of the gap but introduce artifacts. The fundamental challenges are: (1) building appearance varies drastically between regions (tin roofs in Nairobi vs. thatched roofs in Dar es Salaam vs. concrete in Dhaka), (2) buildings are often partially obscured by vegetation or adjacent structures, and (3) imagery resolution varies (50 cm commercial vs. 30 cm drone), creating scale-dependent detection performance.

Two directions are promising: (1) foundation models pre-trained on globally diverse building imagery — analogous to how large language models generalize across languages — that learn building-ness independent of roof material, shape, or regional style; (2) active learning systems that efficiently select the most informative images for human labeling in each new target area, minimizing annotation effort while maximizing transfer. Coupling either approach with drone imagery (higher resolution, oblique angles) could provide the fine detail that satellite imagery lacks for small informal structures.

A team could benchmark existing building detection models on a curated set of informal settlement images from multiple regions and quantify the specific failure modes (false negatives from unusual roofing, false positives from terrain features). Alternatively, a team could develop a minimal annotation pipeline that uses a pre-trained model's uncertainty to prioritize human labeling, testing how few labeled examples are needed to achieve acceptable accuracy in a new city. Skills: computer vision, geospatial data, machine learning, GIS.