Nightlights Find Factories but Cannot Count Them: A Ground-Truth Benchmark Using 6.3 Million Chinese Manufacturing Firms

Satellite nightlights are widely used as a proxy for economic activity at fine spatial resolution, but they are usually validated against aggregate output for large administrative units. I build a cell-level benchmark from 6,311,795 georeferenced Chinese manufacturing firms in 13,849,352 grid cells of 30 arcseconds, observed in 2015 and 2019 and classified into eight subsectors, and merge it with all 24 monthly NPP-VIIRS composites for those years and with the 100 m Global Human Settlement Layer. Using spatially blocked cross-validation, I test whether nightlights predict the level, composition, and growth of manufacturing, and whether they improve on a free daytime settlement layer. At 0.93 km, nightlights detect whether a cell contains manufacturing with an area under the ROC curve of 0.951, but among occupied cells they explain 0.6% of the variance in how many firms a cell holds; this intensive-margin signal appears only above roughly 7 km. The settlement layer outperforms nightlights at every scale, and nightlights add only 0.017 to 0.039 to out-of-sample R^2 once built-up surface and population are known. Composition is close to unpredictable, and no subsector is unusually visible: predictability tracks subsector size (r = 0.96). Relative to the settlement layer, however, light contributes about ten times more for growth than for levels. Nightlights are good at locating manufacturing, but daytime imagery is better at measuring how much of it there is.