What nighttime lights can—and cannot—tell us
A reflection on using remote-sensing signals as evidence for events on the ground.
Nighttime-light imagery is compelling because it makes change visible. A city brightens, a corridor dims, a boundary appears. The picture feels close to an explanation.
Working with VIIRS data taught me to slow down at exactly that moment. Light is a measurement. The human condition we care about is usually something else.
The appeal of a visible proxy
Satellite observations offer consistent, repeated coverage across places where other data may be delayed, incomplete, or unavailable. Nighttime radiance can help reveal changes in electrification, activity, infrastructure, and disruption.
But “can help reveal” is not the same as “directly measures.” A drop in radiance may be consistent with an outage, population movement, damaged infrastructure, seasonal change, cloud contamination, or processing choices. Several mechanisms can point in the same visual direction.
The map becomes evidence only after those alternatives enter the analysis.
Comparisons carry assumptions
Many remote-sensing questions are really comparison questions: before versus after, affected versus less affected, one region versus another.
Each comparison asks for a counterfactual—what would this place have looked like without the event? Calendar effects, sensor artifacts, weather, land cover, and long-term trends can all weaken a simple comparison.
This shifted how I think about a “clean” visualization. A clean figure is not one that removes complexity. It is one that makes the comparison and its uncertainty legible.
Combine signals without pretending they agree perfectly
Using vegetation indices alongside nighttime lights can add context because the signals respond to different processes. That does not automatically make the conclusion stronger. The spatial resolution, observation schedule, noise, and interpretation of each measure still differ.
The useful question is not whether two maps look similar. It is whether the mechanisms that generate them support the same explanation—and what disagreement between them might teach us.
A lesson beyond remote sensing
The same pattern appears in machine learning evaluation. We often choose a convenient observable—accuracy, preference, agreement, benchmark success—as a proxy for a broader quality.
Proxies are unavoidable. The mistake is allowing the proxy and the target to become interchangeable in our language.
The practice I want to carry forward is simple: name the target, name the measurement, and keep the gap between them visible.