Night lights imagery is the most seductive dataset in open Earth observation. A single global composite shows coastlines picked out in white, dense corridors between cities, dark interiors, and it looks immediately like a map of where human activity is. That intuitive readability is exactly why it gets misused.
The instrument behind most current night lights products is VIIRS, a radiometer flying on polar orbiting weather satellites operated jointly by American civilian agencies. It carries a channel designed for extremely low light, usually called the Day Night Band, sensitive enough to record moonlit cloud tops and, at night over land, the light escaping upward from the surface.
What follows is what that measurement physically is, why it became a proxy for economic activity, the distortions that make naive comparisons wrong, and where the line sits between a defensible use and an indefensible one.
What is actually being measured
The sensor measures upwelling radiance in a broad visible and near infrared range: the quantity of light per unit area per unit solid angle arriving at the satellite from a patch of ground below. It is a physical measurement with physical units, not an index and not a count of anything.
Several things follow from that definition and they are worth stating separately, because most misinterpretation comes from skipping them.
It measures light leaving the ground upward, not light useful to people. A well designed street lamp directs light downward onto the pavement and wastes very little upward. A poorly designed one throws light into the sky. Two neighborhoods with identical illumination on the ground can differ substantially at the satellite depending only on fixture design, and the worse engineered one looks brighter.
It measures the top of whatever is between the ground and the satellite. Cloud blocks it. Fog scatters it. Tree canopy hides lights beneath it, so a lit street under mature trees registers dimly, and the seasonal loss of leaves changes the reading without anything on the ground changing at all.
It has a spectral response, and light sources have spectra. Older sodium lamps emit strongly in a range the sensor responds to well. Solid state lighting emits a different distribution, with more energy at the blue end where the band's sensitivity is lower. A city that replaces its lighting stock over several years can show a declining trend that reflects procurement policy rather than decline. This is one of the most important confounders in any multi year comparison and it is invisible in the data itself.
It is an instantaneous sample at a specific local time. A polar orbiting satellite crosses in the small hours. What it captures is the state of lighting at that hour, which is not the same as activity across the evening. Places that shut off lighting late, or that have activity concentrated earlier, are measured differently for reasons unrelated to how much happens there.
Why brightness became a proxy for activity
The appeal is that light correlates with things that are otherwise hard to observe: electrification, roads, settlement, commerce, industry. In places where official statistics are sparse, delayed, or subject to political pressure, an independent physical observation that correlates with development is valuable precisely because it is not produced by the entity being measured.
The correlation is real and the mechanism is not mysterious. Electric lighting requires generation and distribution. It requires something worth lighting. New settlement, new industry, and new roads add light in a way that is hard to fake and easy to see. Conversely, conflict, blackout, disaster, and abandonment remove it quickly and unmistakably.
There are uses where this is close to unassailable. Detecting a power outage after a storm is straightforward, since the lights were there last week and are gone now, and the cause is not ambiguous. Watching a new settlement appear at the edge of a city works because the ground was dark and now is not.
The trouble starts when the correlation is treated as a measurement rather than as a correlate. Light per person varies enormously by climate, by lighting culture, by fixture technology, by the mix of industry and residence, by whether the economy runs at night. An economy dominated by agriculture and one dominated by night shift manufacturing produce very different brightness for similar output. Nothing in the radiance value distinguishes them.
The distortions, and why they matter more than the signal
Every serious use of this data is really an exercise in managing a specific list of artifacts.
Blooming and the point spread. A bright source does not affect only the pixel it sits in. Optical scattering inside the instrument, plus atmospheric scattering, spreads energy into surrounding pixels, so bright cities appear larger than they are and their apparent footprint grows with their brightness. Anyone measuring "lit area" as a proxy for urban extent is partly measuring intensity, and a city that got brighter without expanding will appear to have expanded.
Saturation and dynamic range. Very bright cores can exceed what the instrument records faithfully, compressing differences at the top of the scale. The consequence is that the brightest places are the least well differentiated: two dense downtown areas that differ substantially in activity can return similar values. Historic night lights data from the earlier generation of sensors suffered from this severely, and much of the received wisdom about night lights was formed on data with worse saturation and coarser quantization than current products.
Gas flares and industrial sources. Flaring at oil and gas facilities produces some of the brightest and hottest features visible at night, often in otherwise unpopulated terrain. Left in a composite, a flare field can dominate a region's total brightness and produce nonsense in any per capita interpretation. Flares are separable, because their spectral signature differs from electric lighting, and serious products either mask them or provide the means to. A pipeline that does not handle them will report an empty desert as an economic center.
Fishing fleets and other mobile lights. Lit fishing vessels are bright, they move, and in some seas they form dense clusters. They belong to an economy, but not to the land, and any coastal aggregation that includes them without care will attribute offshore activity to shore.
Moonlight. The band is sensitive enough that lunar illumination reflecting off the surface is a significant contribution. Bright moonlight over bright ground can rival or exceed the light from settlements. Products handle this either by using only observations near the new moon or by modeling and removing the lunar contribution. Raw granules do not.
Snow, ice, and surface albedo. Fresh snow reflects whatever light is available, including moonlight and city light, so a snow covered town measures brighter than the same town in summer. Seasonal snow therefore introduces an annual cycle that has nothing to do with activity, and it is worst exactly where winters are long.
Aurora and airglow. At high latitudes the atmosphere itself emits light. Auroral activity contaminates high latitude observations badly enough that high latitude night lights analysis requires specific handling.
Seasonal vegetation, atmosphere, and viewing geometry. Canopy blocks light in summer. Haze, humidity, and aerosol scatter it. The satellite views a given point from different angles on different passes, and light escapes upward with an angular distribution, so off nadir views of the same place differ from views directly overhead.
The list matters because it is not a set of small corrections around a clean signal. For many questions the artifacts are larger than the difference being investigated.
Raw radiance and calibrated composites are not the same product
A single night's observation is a granule of calibrated radiance with all of the above baked in: cloud, moon, aurora, whatever transient light happened to be there.
A composite is a different object. Composites are built by taking many observations over a period, typically a month or longer, filtering out cloudy and moonlit views, removing or flagging transient and non electric sources, correcting for viewing geometry, and combining what remains into a per pixel value intended to represent the stable background lighting of that place. Some products go further and produce an explicit stable lights layer, separating persistent illumination from ephemeral events.
The practical differences are sharp.
- A composite is comparable across time in a way a granule is not, because the filtering and correction are applied consistently.
- A composite hides events. If the question is when the power went out, monthly composites are the wrong instrument, because a three day blackout is averaged into near invisibility.
- A composite is not available immediately. Filtering requires accumulating observations, so there is an inherent lag between the period and the published composite.
- Composites differ between producers and between versions. Two composites of the same month from different processing chains are not interchangeable, and a version change mid series can create a step in a time series that looks exactly like a real change.
The rule that follows: never mix products or versions inside one trend without checking, and always state which product, which version, and which period. A night lights chart without that provenance cannot be evaluated by anyone, including its author six months later.
What this data supports, and what it does not
The defensible use is change over time in one place, measured with a consistent product, over an area large enough to be robust to blooming and geolocation jitter.
This works because most of the distortions are approximately constant for a fixed location and a fixed method. Fixture technology changes slowly, canopy is similar year to year at the same season, viewing geometry averages out across many passes, and terrain and albedo do not change. When comparing the same place to itself at the same time of year with the same product, most confounders subtract out, and what remains is more likely to be a real change in lighting.
The indefensible use is comparing absolute brightness between countries and inferring relative wealth or output.
The reason is that everything that differs between two countries also differs in ways that affect brightness: lighting technology and its regulation, whether outdoor lighting is culturally normal, energy prices and subsidy structures, the presence of flaring, climate and snow cover, canopy cover, latitude and its effect on both aurora and the length of night, and the sheer geometry of how settlement is arranged. A number that bundles all of these is not a measure of economic output. It is a measure of how much light escapes upward, which is a genuinely different quantity.
Between those poles are uses that are defensible with care. Comparing two cities in the same country, in the same climate zone, with similar lighting infrastructure, is far more reasonable than comparing across continents. Using brightness change as one input among several, rather than as the answer, is reasonable. Using brightness to detect the direction of change rather than its magnitude is reasonable.
A practical procedure
For a question of the form "has activity in this area changed over the last few years", a workable method looks like this.
Choose one composite product and one version, and hold it fixed for the whole series. Define the area of interest, then buffer it outward, because blooming means light from inside spills outside and a tight boundary will systematically undercount. Compare the same calendar months across years rather than consecutive months, so that seasonal snow, canopy, and night length are held roughly constant. Check explicitly for flares and other non settlement sources inside the area and mask them. Look at the distribution across pixels, not only the sum, because a sum can rise from one new bright industrial site while the residential area dims, and those are different stories. Then, before drawing any conclusion, establish whether the change is larger than the year to year variation seen in a nearby comparison area that is expected to be stable.
That last step is the one usually skipped and the one that does most of the work. Without a control, there is no way to distinguish a change in the place from a change in the sensor, the processing, the season, or the sky.
The ethics of inferring wealth from brightness
A dark area is not a poor area. It may be an area with underground infrastructure, dense canopy, effective lighting regulation, a culture of early nights, or simply a different settlement pattern. Treating darkness as deprivation encodes an assumption that a particular style of energy consumption is what development looks like.
This matters because these inferences are used. Night lights products feed into resource allocation, into risk models, into commercial site selection, and into decisions about who gets counted. When brightness stands in for prosperity, places that are dark for reasons unrelated to poverty can be systematically deprioritized, and the error is invisible to everyone downstream because the input looked objective.
The failure has a specific shape. Aggregate statistics derived from lights tend to be least reliable exactly where alternative data is weakest, meaning rural, poor, and less governed areas. That is precisely where a night lights estimate is most likely to be used unchallenged, because there is nothing to check it against. The dataset is most trusted where it is least trustworthy.
There is also a surveillance dimension worth naming. At the resolution these products offer, individual buildings are not identifiable, and that is a genuine privacy property rather than a limitation to be engineered away. Pressure to push toward finer resolution night observation should be treated as a question about consent and purpose, not only about capability.
The practical obligation is modest and concrete. State what was measured, which is upward radiance. State what was inferred, which is anything else. Keep the two separated in the output, so a reader can accept the observation and reject the inference. Anyone who cannot articulate why a place might be dark for reasons other than poverty should not be publishing conclusions about that place from its brightness.