School and health facility access mapping from built-environment data
Satellite-derived building footprints, road networks and gridded population data combine to show which urban residents fall outside realistic catchment distances of schools and health facilities, and which planning assumptions have been quietly wrong for years.
Sensors
- Sentinel-2 MSI: 13-band multispectral at 10 m (visible and NIR) and 20 m (red-edge and SWIR), 5-day revisit at the equator with two satellites. Used here for land-cover classification, impervious-surface delineation and change detection to update settlement extents. Free and openly archived.
- Maxar WorldView-3: Panchromatic at 0.31 m, multispectral at 1.24 m, 8 SWIR bands at 3.7 m. Sub-daily revisit over any target on request. At this resolution, individual building footprints, compound walls, gate openings and narrow alleys are directly measurable, which matters when road-topology extraction relies on passable width.
- WorldPop gridded population: 100 m and 1 km gridded population estimates, produced by the University of Southampton using census microdata, satellite-derived settlement layers and random-forest dasymetric disaggregation. Covers most low- and middle-income countries. Uncertainty is highest in rapidly growing peri-urban areas, which is precisely where access gaps tend to be largest.
- OpenStreetMap road network (satellite-validated): OSM provides the routable graph used for network-distance catchments. Coverage is uneven: dense in city centres, sparse or absent in informal peripheries. Satellite imagery is used to audit OSM completeness and add missing links, but the result is still a snapshot, not a live network.
What the catchment calculation actually measures, and what it does not
A service-area catchment is a deceptively simple idea: draw a boundary around every school or clinic within which a person could reach the facility within some threshold distance or travel time. In practice, two quite different thresholds are in common use. Euclidean buffers, the circles you see on most planning maps, assume straight-line travel. Network-distance buffers route along the road graph and produce irregular, realistic shapes that shrink wherever the street pattern is sparse. Both methods are only as good as the inputs they consume.
The critical point is that neither method captures barriers that exist on the ground but not in the data. A river without a bridge, a compound wall that forces a 400-metre detour, an area avoided after dark because of crime: none of these appear in a road topology derived from satellite imagery or OSM. The catchment polygon says the clinic is 600 metres away. The resident who must walk 1.4 kilometres around a flooded drainage channel knows otherwise. Honest analysis names this gap explicitly rather than presenting a catchment map as settled truth.
How satellite data enters the workflow
Three distinct satellite-derived layers feed the analysis. First, building footprints extracted from WorldView-3 imagery at 0.31 m panchromatic resolution define where people live and where facilities are located. At this resolution, a school's perimeter wall, its gate and the road it faces are all distinguishable. Sentinel-2 at 10 m is too coarse for individual footprint extraction in dense urban fabric, but it contributes the land-cover classification that separates residential from commercial and industrial zones, and it provides the change-detection layer that flags new settlement growth since the last high-resolution acquisition.
Second, the road network is extracted or validated against satellite imagery. OSM completeness audits using WorldView-3 imagery routinely find that informal settlement areas have 30 to 60 per cent of their passable paths missing from the graph, based on published comparisons in sub-Saharan African cities. Missing paths mean the network-distance catchment is artificially large: the algorithm thinks travel is harder than it is, or routes people along roads that do not exist at the required width.
Third, WorldPop's 100 m grid distributes population into the building footprint layer using dasymetric reallocation, concentrating estimated residents into pixels that satellite imagery confirms as residential. The output is a population-weighted demand surface. Facilities with large populations inside their catchment but small facility capacity are flagged as pressure points, not just access gaps.
Where WorldPop's uncertainty bites hardest
WorldPop estimates are derived from census data disaggregated by machine-learning models trained on settlement proxies visible from space: building density, roof area, nighttime light intensity. In stable, well-surveyed cities the estimates are reasonably consistent with ground counts. In rapidly growing peri-urban zones, where new structures appear between census cycles and informal tenure means households are undercounted, the uncertainty can be substantial. The WorldPop team publishes uncertainty surfaces alongside their estimates, and any responsible catchment analysis should propagate that uncertainty into the final access-gap statistics rather than reporting a single number.
This matters for planning decisions. If the model says 12,000 people live outside the 1-kilometre catchment of any primary school, but the population uncertainty range for that zone spans 8,000 to 17,000, the policy implication is different from a precise count. Presenting the range is not a weakness in the analysis; it is the analysis.
Calibrating the threshold: what counts as 'accessible'?
The WHO and various national planning standards suggest different distance thresholds for different facility types. Primary health care is often benchmarked at 5 kilometres in rural settings, but urban planning guidance in many countries uses 1 to 2 kilometres for clinics and 500 metres to 1 kilometre for primary schools, recognising that urban residents travel on foot and that children's safe walking range is shorter than adults'. The choice of threshold is a policy decision, not a technical one, and the satellite analysis should be run at multiple thresholds so decision-makers can see the sensitivity.
Network-distance thresholds should also be adjusted for road quality where that information is available. A 1-kilometre path along a paved street is not the same as 1 kilometre across unpaved, seasonally flooded ground. Satellite-derived surface classification using Sentinel-2 spectral indices can distinguish paved from unpaved surfaces at city scale, and that classification can be used to apply a friction coefficient to the road graph, producing a travel-time surface rather than a raw distance surface.
Honest limits of the method
High-resolution imagery captures physical infrastructure. It does not capture whether a school is open, staffed, or has functioning toilets. A building that looks like a clinic from 0.31 m resolution may be derelict, private, or operating at a fraction of its nominal capacity. Ground-truth validation against administrative facility registers is not optional; it is the step that converts a satellite-derived map into a usable planning input.
Cloud cover is a persistent constraint on WorldView-3 acquisition in tropical cities, where the rainy season can delay tasking by weeks. Sentinel-2 mitigates this through its 5-day revisit and its ability to composite cloud-free mosaics over longer periods, but at 10 m it cannot substitute for WorldView-3 in footprint extraction tasks. Archive depth for WorldView-3 varies by location; some cities have consistent coverage back to 2014 when the satellite launched, others have sparse historical records.
Finally, the road graph extracted from satellite imagery represents the network at the time of acquisition. Informal cities change quickly. A path that was open in the imagery used for extraction may be blocked by a new structure six months later. Periodic re-acquisition and network refresh are part of the maintenance cost of keeping the analysis current. Satellize incorporates this kind of scheduled re-analysis into programme design, as it did when structuring the data pipeline for the Kingdom of Tonga crop-estimation programme, where temporal consistency of inputs was equally critical.
What a planner receives at the end
The deliverable is not a single map. It is a set of GIS layers: facility catchment polygons at each threshold, a population-weighted access-gap surface showing estimated residents outside catchment, a road-network completeness audit flagging segments added or corrected against imagery, and a facility-pressure index ranking existing schools and clinics by the ratio of population served to estimated capacity. Each layer carries metadata documenting the imagery date, the population grid version and the threshold assumptions used.
Where the analysis identifies a candidate location for a new facility, it can also run a simple optimisation: given a budget for one or two new sites, which locations would reduce the total population outside catchment by the largest margin? This is a standard p-median or coverage-maximisation problem, well documented in the operations research literature, applied to a satellite-derived demand surface. The satellite data does not make the decision. It makes the decision harder to get wrong.
Typical figures
| Building footprint resolution | 0.31 m panchromatic (WorldView-3); individual structures and compound walls distinguishable |
| Land-cover classification resolution | 10 m (Sentinel-2 visible and NIR bands); 20 m for red-edge and SWIR indices |
| Population grid resolution | 100 m (WorldPop constrained model); 1 km also available for regional overviews |
| Sentinel-2 revisit | 5 days at equator (two-satellite constellation); cloud-free composites typically achievable within 30 days in tropical regions |
| WorldView-3 revisit | Sub-daily over any target on commercial tasking; archive depth varies by city, generally from 2014 onwards |
| Network catchment thresholds | Configurable; typical runs at 500 m, 1 km and 2 km for schools; 1 km, 2 km and 5 km for health facilities |
| Minimum detectable road width | Approximately 1.5 m at WorldView-3 panchromatic resolution; paths narrower than this require field validation |
| Archive depth (Sentinel-2) | 2017 onwards globally; earlier acquisitions available for some regions from 2015 |
| Delivery formats | GeoPackage, GeoJSON, Shapefile, Cloud-Optimised GeoTIFF; facility pressure index as CSV or dashboard-ready feed |
Analytics Satellize can run
| Building footprint layer with facility identification | Object-based image analysis and deep-learning segmentation on WorldView-3 panchromatic and multispectral imagery; facility type inferred from morphology and spectral signature | GIS polygon layer of residential and facility footprints with attribute table |
| Road network completeness audit | Automated comparison of OSM graph against satellite-derived road centrelines; missing segments flagged and added to routable graph | Corrected road network GeoPackage with audit log of additions and modifications |
| Euclidean and network-distance catchment polygons | Straight-line buffer and network-service-area analysis using corrected road graph; run at multiple configurable thresholds | Catchment polygon layers per facility and per threshold, in GeoJSON and Shapefile |
| Population-weighted access-gap surface | Dasymetric reallocation of WorldPop 100 m grid to building footprints; subtraction of population within catchment; uncertainty range propagated from WorldPop uncertainty surface | Raster access-gap surface and summary statistics table with uncertainty ranges by administrative zone |
| Facility pressure index | Ratio of dasymetrically allocated population within catchment to nominal facility capacity (from administrative register); ranked list of overloaded facilities | CSV and dashboard-ready ranked facility list with supporting map layer |
| Candidate new-facility site optimisation | P-median or maximal coverage location problem applied to population-weighted demand surface; solved for one to five candidate sites | Ranked candidate site report with estimated access-gap reduction per scenario |
| Travel-time surface with surface-quality friction | Sentinel-2 spectral index classification of paved versus unpaved roads; friction coefficients applied to road graph; least-cost-path travel-time raster generated | Travel-time raster (minutes to nearest facility) at 10 m resolution, Cloud-Optimised GeoTIFF |
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.