Energy policy promotes electricity use, but the grid is unstable during disasters, so people turn to backup generators, which pollute the air. Yet no public dataset records where they are. — yet no public registry records where these generators are or whether they actually work. We ask whether NASA's VIIRS Black Marble nighttime-light satellite (500 m resolution, daily revisit) can detect the brightness signature of backup power running during major outages.
Across 25 disasters in 17 U.S. states, Puerto Rico, and Türkiye (2016 – 2023), we extract a panel of ~33,700 pixels, treat critical infrastructure (hospitals, airports, fire stations, police, power plants) as proxy labels, and train a tree-based model on purely temporal nighttime-light features — no spatial proximity to facilities — so its predictions reflect satellite signal alone. Across all events the model achieves LOEO AUC = 0.704; against actual generator-permit records in Miami-Dade, 83 % of commercial installations score above the event-wide median probability (vs. 14 % of residential), confirming the resolution boundary. At zip-code scale, model probability is positively associated with real outage severity (standardized β ≈ 0.08, p = 0.012).
