Population-distribution demand mapping for network dimensioning
High-resolution population grids derived from building footprints, built-up-area classification and nighttime-light intensity let operators dimension macro-cells and build spectrum-licence business cases from evidence rather than extrapolated census figures.
Sensors
- VIIRS Day/Night Band (DNB): Suomi-NPP and NOAA-20 VIIRS DNB provides ~750 m ground sample distance nighttime radiance composites. Monthly cloud-free composites from NOAA/NGDC distinguish electrified settlements from unlit informal housing. Detection floor is roughly 2 × 10⁻¹⁰ W cm⁻² sr⁻¹, sufficient to separate kerosene-lamp glow from grid-connected areas in most conditions.
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands (B02–B08) with a 5-day revisit at the equator under cloud-free conditions. Used for built-up-area classification via spectral indices (NDBI, NDVI) and for detecting impervious surface expansion in peri-urban zones. Free and open; global archive from 2015.
- Landsat 8/9 OLI: 30 m multispectral resolution with a 16-day single-satellite revisit, extendable to 8 days with both satellites in tandem. Longer archive (Landsat 7 from 1999, Landsat 8 from 2013) supports multi-year built-up-area change detection. Band 6 (SWIR) is particularly useful for distinguishing bare soil from low-density built fabric in arid peri-urban settings.
- WorldView-3: 31 cm panchromatic, 1.24 m multispectral resolution. Used for building-footprint delineation in high-priority licence areas where 10 m Sentinel-2 pixels conflate multiple structures. Tasked commercially; cost and cloud risk must be budgeted. Not suitable as the sole source for national-scale surveys.
What a rooftop count does and does not tell you
Building-footprint extraction from high-resolution imagery is the most tractable proxy for population distribution in areas where census data is stale or spatially coarse. The logic is straightforward: people occupy structures, structures have rooftops, rooftops are visible from orbit. Automated detection using convolutional neural networks trained on labelled urban imagery can achieve footprint delineation at roughly 1 m accuracy when WorldView-3 or similar very-high-resolution imagery is available. At Sentinel-2's 10 m resolution, individual small structures merge into mixed pixels, so the output shifts from footprint counts to built-up-area fraction per pixel.
The problem is that footprint area and occupancy density are poorly correlated in informal settlements. A 40 m² concrete-block structure in a formal suburb may house four people. The same footprint in a peri-urban informal area may house twelve, across multiple internal partitions invisible from above. WorldPop's dasymetric modelling acknowledges this explicitly: it uses census microdata to calibrate per-structure occupancy priors, but those priors carry confidence intervals that can exceed ±40% in rapidly growing informal zones. For network dimensioning, that uncertainty propagates directly into subscriber-count forecasts and, by extension, into cell-capacity calculations.
Where nighttime light earns its place in the stack
VIIRS DNB monthly composites add a dimension that building counts alone cannot supply: electrification status. An electrified household is almost certainly a smartphone-capable household. A dark rooftop cluster may represent genuine low-demand rural population or it may represent a dense informal settlement running on mobile generators and solar lanterns, where 4G data demand is high but grid connection is absent. The DNB cannot resolve this ambiguity alone, but it constrains the prior substantially.
The practical workflow combines DNB radiance with Sentinel-2 built-up fraction. Pixels that are spectrally built-up but radiometrically dark flag the peri-urban informal zones that most challenge demand forecasting. These are precisely the areas where operators face the largest uncertainty in spectrum-licence business cases and where field survey investment is most justified before committing to macro-cell site acquisition. DNB's 750 m footprint means it cannot distinguish a single bright commercial building from a dense residential cluster of equivalent luminosity, so it must always be used as a stratification layer rather than a standalone demand signal.
WorldPop and GHSL: two methodologies, two honest limitations
WorldPop (University of Southampton) produces gridded population estimates at 100 m and 1 km globally, using random-forest models trained on census data, settlement extents, land cover, road networks and nighttime light. The outputs are probabilistic: each pixel carries a mean estimate and a credible interval. In countries with recent, high-quality census microdata, the 100 m product performs well. In countries where the most recent census is more than a decade old, or where internal migration has been rapid, the model inherits the age of its training data. Operators should treat WorldPop as a structured prior, not a ground truth.
The Global Human Settlement Layer (GHSL), produced by the EU Joint Research Centre, takes a different route. It classifies settlement type (urban centre, urban cluster, rural) from Landsat and Sentinel-2 time series, then applies population grids derived from census disaggregation. GHSL's Degree of Urbanisation framework is now used by the UN for international comparisons, which gives it political credibility. Its spatial resolution for the built-up layer is 10 m (from Sentinel-2, in recent epochs), but the population grid remains at 100 m. Neither WorldPop nor GHSL is clearly superior for all geographies; the sensible approach is to run both, compare their outputs in the licence area, and flag cells where they disagree by more than a defined threshold as requiring validation.
Translating population grids into traffic demand
A population grid becomes a traffic-demand surface through a chain of assumptions, each of which deserves explicit documentation in a spectrum-licence business case. Penetration rate (what fraction of the population holds an active SIM) varies from below 30% in parts of sub-Saharan Africa to above 100% in many markets where multi-SIM ownership is common. Data-capable device penetration is a further discount. Average revenue per user and average data consumption per user per day are operator-specific inputs that satellite data cannot supply.
What satellite-derived grids do supply is the spatial allocation of demand, which is the input to macro-cell dimensioning. Given a total-market subscriber forecast, the grid distributes those subscribers across candidate cell footprints in proportion to population density, weighted by electrification status from DNB. The output is a per-cell subscriber count, which feeds directly into link-budget and spectrum-efficiency calculations. The grid also identifies where demand is concentrated enough to justify a macro site versus where a shared rural site or a fixed-wireless access node is the more defensible investment.
Honest limits and what field data must still provide
Cloud cover is a persistent constraint in tropical markets, which are often the same markets where population data is weakest. Sentinel-2's 5-day revisit does not guarantee a cloud-free observation within any given month in equatorial zones; annual compositing reduces cloud contamination but loses temporal currency. WorldView-3 tasking in persistently cloudy regions carries a real probability of failed acquisitions.
Vertical structure is invisible to optical sensors. A city block of six-storey apartment buildings and a block of single-storey dwellings can produce identical spectral signatures at 10 m resolution. Building-height extraction from stereo or LiDAR data is a separate analytical step, covered in a sibling page in this library. For population mapping purposes, the practical workaround is to use GHSL's morphological settlement classification, which distinguishes dense urban fabric (implying multi-storey construction) from open low-rise, and apply differentiated occupancy multipliers accordingly. This reduces but does not eliminate the uncertainty.
Satellize runs population-grid analytics on open Sentinel-2, Landsat and VIIRS archives, with optional WorldView-3 tasking for high-priority licence zones. The Tonga crop-estimation programme demonstrated the team's approach to calibrating satellite-derived spatial models against sparse ground-truth data, a methodology that transfers directly to peri-urban population work. For a new licence area, the practical starting point is a grid-comparison report covering WorldPop and GHSL outputs, a DNB electrification stratification layer, and a documented uncertainty map flagging the cells that carry the highest demand-forecast risk.
Typical figures
| Built-up area classification resolution | 10 m (Sentinel-2 MSI); 30 m (Landsat 8/9 OLI) |
| Building footprint resolution (commercial tasking) | 0.31 m pan / 1.24 m multispectral (WorldView-3) |
| Nighttime light ground sample distance | ~750 m (VIIRS DNB, Suomi-NPP / NOAA-20) |
| Population grid output resolution | 100 m (WorldPop, GHSL standard products); 1 km also available |
| Sentinel-2 revisit (equator) | 5 days (two-satellite constellation); cloud-free composites typically monthly in tropics |
| VIIRS DNB monthly composite latency | Approximately 2–4 weeks after month end (NOAA/NGDC release schedule) |
| Sentinel-2 archive depth | From 2015 (Sentinel-2A); Sentinel-2B from 2017 |
| WorldPop uncertainty range (peri-urban informal zones) | Credible intervals commonly ±30–50% at 100 m pixel level in rapidly growing informal settlements |
| VIIRS DNB detection floor | ~2 × 10⁻¹⁰ W cm⁻² sr⁻¹ (sufficient to detect small electrified clusters) |
| Delivery formats | GeoTIFF population grids, GeoPackage / Shapefile settlement classification, CSV per-cell demand table, PDF uncertainty report |
Analytics Satellize can run
| Built-up area fraction grid | Spectral index classification (NDBI, NDVI) on Sentinel-2 or Landsat time-series composites, with morphological filtering to remove false positives from bare soil | GeoTIFF raster at 10 m or 30 m, covering the licence area, with per-pixel built-up probability and a binary classification layer |
| Electrification stratification layer | VIIRS DNB monthly composite analysis; thresholding and spatial clustering to distinguish grid-connected, off-grid-lit and unlit settlement zones | GeoTIFF and vector polygon layer classifying settlement pixels into three electrification tiers, with source composite dates documented |
| Gridded population estimate with uncertainty bounds | Dasymetric disaggregation of census data using WorldPop random-forest model outputs and GHSL settlement classification; dual-model comparison to flag high-disagreement cells | 100 m population grid (mean and 90% credible interval), plus a divergence map highlighting cells where WorldPop and GHSL estimates differ by more than a configurable threshold |
| Per-candidate-cell subscriber forecast surface | Population grid weighted by electrification tier and operator-supplied penetration-rate assumptions; aggregated to candidate macro-cell Voronoi polygons | CSV and GIS layer with per-cell population, electrification-weighted subscriber estimate, and confidence band; ready for import into radio-planning tools |
| Peri-urban informal settlement risk flags | Intersection of high built-up fraction, low DNB radiance and high WorldPop/GHSL divergence; identifies cells where demand uncertainty is highest and field validation is most cost-effective | Vector polygon layer of flagged zones with uncertainty score, recommended field-survey priority ranking, and notes on spectral ambiguity sources |
| Multi-epoch built-up change detection | Bi-temporal or time-series change detection on Sentinel-2 annual composites; identifies newly urbanised pixels for demand-forecast updating ahead of licence renewal | Change raster and summary report showing net built-up area gain per administrative unit per year, with dates of detected change |
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.