Nighttime light intensity mapping for economic activity and data-demand forecasting
VIIRS Day/Night Band radiance time-series proxy economic activity and electrification at sub-kilometre resolution, giving network planners a leading indicator of data demand where census data is thin or stale. The signal is real but confounded; interpreting it correctly is most of the work.
Sensors
- VIIRS Day/Night Band (Suomi-NPP / NOAA-20 / NOAA-21): 750 m native pixel, nightly global revisit, radiometrically calibrated in watts per cm² per steradian. The primary workhorse for radiance time-series. Monthly composites from NOAA's Earth Observation Group remove cloud and stray-light contamination. Archive runs from 2012.
- DMSP-OLS archive (NOAA): Roughly 2.7 km resolution, uncalibrated digital number (0–63 scale), annual composites 1992–2013. Useful only for multi-decade trend analysis; saturation over bright city centres is a documented and significant limitation.
- Landsat 8/9 OLI panchromatic band: 15 m resolution, 16-day revisit per satellite (8-day combined). Not designed for nighttime radiometry but usable for targeted high-resolution snapshots of industrial sites, ports or flaring locations that contaminate VIIRS pixels. Daytime imagery used to characterise land cover confounders.
- Sentinel-3 OLCI: 300 m resolution, daily revisit. Primarily a daytime ocean-colour and land sensor, but its synoptic daily coverage supports land-cover and seasonal-cycle characterisation that helps separate agricultural burning events from genuine electrification signals in VIIRS composites.
Why radiance predicts demand at all
The empirical relationship between nighttime radiance and sub-national GDP per capita has been documented since the DMSP-OLS era and has been refined substantially with VIIRS calibrated radiance. The mechanism is straightforward: economic activity requires energy, and a meaningful fraction of that energy escapes upward as light. At the sub-national level, radiance explains a substantial share of GDP variance in data-sparse countries, though the relationship is non-linear and varies by sector mix. A mining enclave, for instance, produces high radiance and high GDP but almost no residential data demand.
For telecoms planners, the more useful relationship is not the level of radiance but its rate of change. A district where VIIRS monthly composites show radiance growing at several percent per year is, with reasonable confidence, electrifying faster than the national average. That is a leading indicator of handset penetration and, with a lag of roughly two to four years in documented emerging-market rollouts, of mobile data demand. The lag exists because grid connection precedes smartphone adoption, which precedes data-hungry application use. Radiance growth does not tell you when to build; it tells you where to look harder.
What a stable composite actually measures, and what it does not
NOAA's Earth Observation Group produces monthly VIIRS Day/Night Band composites that filter out cloud-contaminated observations, ephemeral fires and aurora. The result is a radiance field that represents persistent artificial light. At 750 m, a single pixel can contain a mix of streetlighting, industrial facilities, fishing vessels anchored offshore, and residential lighting. Disaggregating those contributions requires ancillary data: land-cover classification, known industrial site locations and, where available, gas-flare registers.
Gas flaring is the most disruptive confounder in hydrocarbon-producing regions. A single large flare stack can saturate surrounding pixels and produce radiance signatures that dwarf the surrounding population centres. NOAA's VIIRS Nightfire product, which uses shortwave infrared bands to characterise combustion temperature, allows flares to be identified and masked before economic inference. Without that step, radiance-based demand models in the Niger Delta, the Permian Basin or the Caspian littoral will be systematically wrong.
Seasonal agricultural burning produces transient radiance spikes that monthly compositing mostly removes, but in regions with dense, repeated burning cycles, residual contamination persists. Public lighting policy changes are a subtler problem: a government decision to install LED streetlights across a province will reduce radiance even as economic activity grows, because LED fixtures emit less upward-directed light than the sodium vapour lamps they replace. This is documented in European and Chinese cities and is increasingly relevant in sub-Saharan Africa as LED rollouts accelerate under donor programmes.
Building a demand-forecast layer from radiance time-series
The practical workflow starts with a per-pixel radiance time-series extracted from monthly VIIRS composites, typically over a five-to-ten-year window to capture trend rather than noise. A pixel-level linear or log-linear trend is fitted; pixels with statistically significant positive slopes are candidates for near-term demand growth. Those candidates are then cross-referenced against population-distribution layers, road-network accessibility and existing coverage maps to produce a ranked opportunity surface.
The radiance layer supplements rather than replaces census-based demand models. In countries where the most recent census is more than a decade old, or where internal migration has been rapid, radiance growth provides a more current signal of where economic mass is accumulating. The combination of a stale census baseline corrected by a radiance-derived growth index is a published approach used in development-finance infrastructure planning and is directly applicable to universal-service fund prioritisation.
One honest caveat: radiance-based models perform worst precisely where they are most needed. In very low-radiance rural areas, the VIIRS Day/Night Band approaches its detection floor, which is roughly 2 × 10⁻¹⁰ watts per cm² per steradian for monthly composites under favourable lunar conditions. Sparse electrification in those areas produces signals that are difficult to distinguish from background noise and atmospheric scatter. For those geographies, daytime proxy indicators such as rooftop density from Landsat or building-footprint counts from commercial optical imagery are more reliable demand signals.
Confounders that will embarrass an uncritical model
Three categories of false signal recur often enough to deserve explicit treatment in any methodology note delivered to a client. First, fishing fleets. In coastal and island geographies, anchored or slow-moving fishing vessels using high-intensity lamps to attract squid and other species can produce radiance signatures comparable to small towns. The South China Sea and waters around Southeast Asian archipelagos are well-documented examples in the published literature. A demand forecast built on those pixels would be nonsense.
Second, refugee and displaced-person settlements. Large camps served by humanitarian generators produce genuine radiance that reflects population density but not the economic activity profile that drives commercial data demand. Third, temporary infrastructure: mining exploration camps, seasonal agricultural processing facilities and large construction sites all produce radiance spikes that decay when the activity ends. A time-series long enough to observe the full cycle, rather than a single composite snapshot, is the minimum acceptable input.
Where this fits in a planning workflow
Radiance-based demand forecasting is most defensible as a screening tool at the regional or national scale, used to rank districts for deeper field investigation or higher-resolution analysis. It is not a substitute for site-level traffic modelling, which requires ground-truth subscriber data and application-mix assumptions that no satellite can supply.
At Satellize, this type of radiance time-series analysis sits alongside the kind of crop-estimation work done for the Kingdom of Tonga: both are cases where open-constellation data, processed carefully, can substitute for ground surveys that are expensive or impractical. The output for a telecoms client is typically a ranked district-level opportunity layer, updated quarterly as new VIIRS monthly composites are released, with explicit flagging of pixels where known confounders reduce confidence. That flagging is not a weakness in the product; it is the product. An uncaveated demand map is a liability.
Typical figures
| Native spatial resolution (VIIRS DNB) | 750 m at nadir; degrades to ~1,600 m at swath edge |
| Revisit (VIIRS) | Nightly global coverage per satellite; Suomi-NPP, NOAA-20 and NOAA-21 provide multiple overpasses per night at mid-to-high latitudes |
| Composite latency | Monthly composites from NOAA Earth Observation Group typically available 2–4 weeks after month end |
| Detection floor (monthly composite) | Approximately 2 × 10⁻¹⁰ W cm⁻² sr⁻¹ under favourable lunar and atmospheric conditions; higher in practice for low-radiance rural areas |
| Spectral band (VIIRS DNB) | Panchromatic, 500–900 nm; sensitive to visible and near-infrared emission from artificial lighting and combustion |
| Archive depth | VIIRS from 2012; DMSP-OLS annual composites from 1992 (uncalibrated, ~2.7 km resolution) |
| DMSP-OLS saturation limit | Digital number caps at 63; bright urban cores are systematically saturated, making inter-city comparisons unreliable |
| Delivery formats | GeoTIFF radiance composites, per-pixel trend rasters, vector district-level summary tables (CSV/GeoJSON), quarterly PDF summary reports |
| Cloud impact | Monthly compositing removes most cloud-contaminated observations; persistent cloud cover in equatorial regions reduces effective sample size and increases composite uncertainty |
Analytics Satellize can run
| Radiance trend surface | Per-pixel log-linear regression on VIIRS monthly composites over a user-specified window (minimum 3 years recommended); trend slope and p-value retained per pixel | GeoTIFF raster of annualised radiance growth rate with statistical significance mask; updated quarterly |
| Electrification frontier map | Threshold-based classification of pixels transitioning from sub-detection to persistent radiance across successive annual composites, following published methods from the World Bank ESMAP electrification tracking literature | Vector polygon layer of newly electrified areas by year, with confidence tier based on number of cloud-free observations |
| Confounder-flagged demand opportunity ranking | Radiance trend surface cross-referenced against VIIRS Nightfire flare register, land-cover classification (Sentinel-3 derived) and known fishing-ground extents; residual signal attributed to economic activity | District-level ranked table with explicit confounder flags; delivered as GeoJSON and CSV for import into network planning tools |
| Radiance-to-GDP index calibration | Regression of sub-national VIIRS radiance against published sub-national GDP or consumption estimates (World Bank, national statistics offices) to produce a country-specific elasticity coefficient for demand scaling | Calibration report with coefficient, confidence interval and documented data sources; one-off deliverable updated when new sub-national economic data is published |
| Temporal anomaly detection for transient infrastructure | Z-score deviation from pixel-level seasonal baseline; flags construction camps, temporary flaring and other non-persistent radiance sources that would bias trend estimates | Monthly alert layer of anomalous pixels with classification (probable flare, probable vessel, probable construction); GeoJSON feed |
| Multi-decadal change archive (DMSP + VIIRS intercalibrated) | Intercalibration of DMSP-OLS and VIIRS using overlap years (2012–2013) following published invariant-target methods; produces a continuous 1992-to-present relative radiance index per district | District-level time-series CSV with documented intercalibration uncertainty bounds; suitable for long-range infrastructure investment planning |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.