Population distribution modelling from built-environment data
Dasymetric mapping disaggregates census totals onto fine spatial grids by weighting them against satellite-derived proxies: building footprints, roof area, settlement extent, and night-light intensity. The method is powerful and the uncertainty is compounding. Both facts matter.
Sensors
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 5-day revisit at the equator with both satellites. Feeds settlement-extent mapping and impervious-surface classification. Free and open; global archive from 2015. Cannot resolve individual buildings in dense urban cores or distinguish floor counts.
- Maxar WorldView-3: 0.31 m panchromatic, 1.24 m multispectral. Sufficient to delineate individual building footprints and estimate roof area with high confidence. Tasked commercially; archive depth varies by location. Stereo pairs allow building-height estimation, which partially addresses the floor-count problem.
- VIIRS Day/Night Band (DNB): ~742 m spatial resolution, nightly global coverage. Radiance values correlate with economic activity and residential density at neighbourhood scale. Saturation over city centres is a known limit; the DNB is most informative at peri-urban and rural fringes where building-based proxies are sparse.
- Planet SuperDove: 3 m resolution, 8 spectral bands, near-daily revisit globally. Useful for tracking settlement expansion between WorldView tasking epochs and for time-series analysis of construction activity. Not fine enough for individual footprint extraction in dense areas.
What dasymetric mapping actually does
A census polygon, whether a district or a ward, reports a single number: total population. That number tells you nothing about where within the polygon people live. Dasymetric mapping solves this by redistributing the count across sub-polygon grid cells in proportion to a proxy variable that correlates with human presence. Building footprint density is the most defensible proxy available from satellite data, because people occupy buildings. Roof area, settlement extent, and night-light radiance are secondary proxies, each carrying different information and different error.
The output is a gridded population surface, typically at 100 m or finer. WorldPop publishes global surfaces at 100 m; the Global Human Settlement Layer (GHSL) from the European Commission's Joint Research Centre works at 100 m and 1 km. Both are openly available and both are honest about their uncertainty bounds, which is worth noting before you treat any cell value as a ground truth.
The proxy chain and where it breaks
Each link in the proxy chain introduces error. Sentinel-2 at 10 m can delineate settlement extent reliably, but individual building footprints at that resolution blur together in dense urban fabric. WorldView-3 at 0.31 m resolves footprints cleanly, but coverage is patchy and tasking is expensive. Planet SuperDove at 3 m sits in between: good enough for settlement mapping, not quite enough for footprint-level precision in tightly packed blocks.
Night-light intensity from VIIRS DNB adds a different dimension: it captures economic activity and electrification rather than physical structure. In informal settlements with low electrification, VIIRS systematically underestimates presence. In city centres, it saturates. The DNB is best used as a correction layer or a weight in ensemble models, not as a standalone proxy.
The deeper problem is floor count. A nadir optical sensor sees a roof. It cannot see whether that roof covers one storey or twenty. A ten-storey apartment block and a single-storey warehouse have identical footprints from above. Stereo imagery from WorldView-3 can recover building height, and the GHSL Height layer uses a combination of stereo optical and SAR data to estimate this globally, but height-to-floor-count conversion still requires assumptions about local construction norms. In cities undergoing rapid vertical densification, those assumptions age quickly.
Compounding uncertainty: when both inputs carry error
The census total is not a clean denominator. In many countries, census undercounting in informal or rapidly growing areas is well documented. When you distribute an already-uncertain total using an imperfect satellite-derived weight, the errors multiply rather than cancel. A 10 percent undercount in the census combined with a 15 percent misclassification in the building-footprint layer can produce cell-level errors well above 20 percent, depending on how the errors are spatially correlated.
This is not a reason to abandon the method. It is a reason to publish uncertainty surfaces alongside population estimates, to be explicit about which census year and satellite vintage were used, and to treat the output as a planning input rather than a registration record. The WorldPop project publishes top-down and bottom-up uncertainty estimates; the GHSL documentation describes its validation approach against independent population data. Buyers of population grids should read those sections.
High-rise cities: a specific and underappreciated difficulty
Dasymetric methods developed primarily on low-rise settlement patterns, where footprint area is a reasonable proxy for floor area and therefore for occupancy. That assumption fails in Hong Kong, Singapore, Lagos Island, or any city where vertical density is the dominant spatial form. A 100 m grid cell containing a cluster of 40-storey towers may house ten times the population of an adjacent cell with the same footprint coverage but only four storeys.
Partial remedies exist. Building height from stereo WorldView or from SAR interferometry can be combined with footprint area to estimate gross floor area, which is a better proxy than footprint alone. Some urban planning authorities publish floor-area-ratio data that can be integrated as a prior. Where none of this is available, the honest answer is that the model uncertainty in high-rise districts is substantially higher than the headline figures suggest, and local validation data, even a partial household survey, is worth more than additional satellite layers.
Where Satellize fits into this workflow
Satellize processes open constellations (Sentinel-2, VIIRS) and adds commercial tasking through client licence where sub-metre resolution is needed for footprint extraction. The analytics pipeline integrates building-footprint layers, settlement-extent classifications, and DNB radiance into dasymetric weighting schemes aligned with WorldPop and GHSL methodologies. Outputs are gridded GeoTIFF surfaces with accompanying uncertainty layers, delivered at the spatial resolution the input data supports rather than at a notional resolution that flatters the product.
The Tonga crop-estimation programme is the one named analytics engagement Satellize publishes; population-distribution work for government clients follows the same principle of being explicit about what the data can and cannot resolve. If a client's census is five years old and their city has grown rapidly since, that caveat goes into the report, not into a footnote.
What good validation looks like
A population grid without validation is a hypothesis. Validation requires independent population counts at a spatial scale finer than the census polygons used to build the model. Mobile network operator data, where available and properly anonymised, provides one such source. Household survey samples provide another. Comparing gridded estimates against known high-density areas, refugee camp populations with documented headcounts, or university campus populations with registration data are all published approaches.
The GHSL validation reports compare modelled population against independent municipal data across dozens of cities. WorldPop publishes country-level accuracy assessments. Neither claims perfection. The honest position for any practitioner is that validation is not optional, it is the only way to know whether the uncertainty surface is calibrated or merely decorative.
Typical figures
| Spatial resolution (settlement extent) | 10 m (Sentinel-2); 3 m (Planet SuperDove) |
| Spatial resolution (building footprint) | 0.31 m pan / 1.24 m multispectral (WorldView-3) |
| Population grid output resolution | Typically 100 m; finer where sub-metre footprint data available |
| Night-light proxy resolution | ~742 m (VIIRS DNB); nightly global coverage |
| Sentinel-2 revisit | 5 days at equator (both satellites combined) |
| WorldView-3 tasking latency | 1 to 3 days for new collect; archive access near-immediate |
| Archive depth | Sentinel-2 from 2015; VIIRS DNB from 2012; WorldView archive varies by location |
| Key spectral inputs | Visible, NIR, SWIR (Sentinel-2/SuperDove); panchromatic + 8-band MS (WorldView-3); panchromatic low-light (VIIRS DNB 0.5–0.9 µm) |
| Uncertainty surface | Published per-cell standard deviation or confidence interval; mandatory output |
| Delivery formats | GeoTIFF (population grid + uncertainty), GeoPackage, CSV with centroid coordinates |
Analytics Satellize can run
| Dasymetric population grid | Top-down dasymetric mapping using satellite-derived building footprint density and settlement extent as weights against census polygon totals | 100 m GeoTIFF population surface with per-cell uncertainty layer; updated on census refresh or major settlement-change event |
| Building-footprint weight layer | Object-based image analysis on WorldView-3 or deep-learning footprint extraction; validated against Sentinel-2 settlement mask | Vector footprint layer and rasterised density grid at 1 m and 10 m; GeoPackage |
| Night-light correction surface | VIIRS DNB radiance normalisation and saturation correction following published NOAA/NGDC methods; used to adjust weights in peri-urban and rural zones | Annual and monthly DNB composites with saturation flags; GeoTIFF |
| Gross floor area estimate | Building height from WorldView-3 stereo or GHSL Height layer combined with footprint area; floor count inferred from local construction-type priors | Per-building and gridded GFA surface; GeoTIFF with confidence class attribute |
| Population change surface | Multi-epoch dasymetric comparison using Sentinel-2 time series to detect settlement expansion; change attributed to population delta from intercensal estimates | Change GeoTIFF showing gain/loss per cell between two epochs; planning report |
| Validation report | Comparison of gridded estimates against independent point-count data (survey samples, known-density reference sites); RMSE and bias statistics by urban zone type | PDF accuracy assessment with zone-stratified error tables; recommended for any operational use |
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.