Do Satellites See Commuters? A Critical Benchmark of Vision Foundation Models
Researchers compared four different satellite vision models—language-supervised, self-supervised, geographically grounded, and multisensor—to generate commuting origin-destination data from census data. Findings showed language-based features performed best for known areas, but geographically grounded models transferred better to new locations like the UK. Crucially, no model could effectively predict travel between global cities.