How the damage was assessed and validated

fAIr detects building footprints on pre-event imagery; an AI damage model compares pre- and post-event imagery. fAIr, Microsoft AI for Good Lab and an OSU/CUNY Sentinel-1 radar product each flag damage on the same H3 grid. Turn any combination on: each source draws as its own layer, and where sources overlap the colour darkens, so the more sources that agree the stronger a cell reads. MapSwipe volunteers then check AI-flagged areas: 4 to 7 people review each area and mark it damaged or not sure. Areas the crowd confirms as damaged (majority agree) become confirmed damage; everything else, where the crowd does not confirm damage, reads as uncertain. Per-building outputs are advisory; the H3 cells and confirmed areas are the screening unit.

Full data, schemas and MapSwipe project links: HuggingFace dataset.