Health facility catchment and physical access mapping using terrain and road data

Satellite-derived elevation, road extraction and gridded population data combine to produce travel-time surfaces that show which communities genuinely cannot reach a clinic within an hour. Static maps lie; this method quantifies the gap.

Sensors

Copernicus DEM GLO-30: Global digital surface model at 30 m horizontal posting derived from TanDEM-X radar data. Vertical accuracy typically better than 4 m RMSE over non-forested terrain. Used to compute slope, aspect and terrain ruggedness index for friction surface generation.
SRTM (NASA/NGA): Shuttle Radar Topography Mission DEM at 30 m (1 arc-second) for latitudes 56°S to 60°N. Vertical accuracy around 9 m absolute at 90th percentile. Widely used as the baseline elevation input where GLO-30 is unavailable or for historical comparisons.
Sentinel-2 MSI: 10 m multispectral imagery (13 bands, 400–2200 nm) with a 5-day revisit at the equator. Used to extract or validate road and settlement extents via supervised classification or deep-learning segmentation, particularly in areas where OpenStreetMap coverage is sparse.
WorldPop gridded population (University of Southampton / RCMRD): 100 m resolution population count and density grids, disaggregated by age and sex for most low- and middle-income countries. Produced using dasymetric modelling anchored to national census data. Accuracy varies by country; uncertainty is higher where census data are old or incomplete.

What a travel-time surface actually measures

A Euclidean buffer drawn around a clinic at 5 km radius tells you almost nothing about whether a patient can reach it in under an hour. It ignores hills, river crossings, the difference between a paved road and a footpath, and the fact that most rural patients in low-income settings walk rather than drive. A travel-time surface replaces the circle with a friction map: every pixel in the landscape is assigned a cost in minutes per metre, derived from slope, surface type and assumed transport mode. Least-cost path algorithms then accumulate that cost outward from each facility location to produce an isochrone surface.

The standard friction inputs are slope (from the DEM), road class (from OpenStreetMap or extracted from imagery), and land cover (to distinguish open ground from forest or wetland). The Malaria Atlas Project published a widely cited global 2015 travel-time-to-cities layer using exactly this approach, and subsequent work extended it to health facilities specifically. The method is well-validated; the uncertainties lie in the input data, not the algorithm.

Elevation data: GLO-30 versus SRTM, and why it matters

Slope is the dominant cost driver for walking populations. A 10-degree slope roughly doubles walking time relative to flat ground under Tobler's hiking function, the empirical speed-slope relationship most friction models use. Getting slope wrong by even a few degrees in steep terrain shifts isochrone boundaries by hundreds of metres. Copernicus GLO-30, derived from TanDEM-X interferometry, is the current best freely available global DEM for this purpose. Its 30 m posting and sub-4 m vertical RMSE over open terrain outperform SRTM in most comparative studies, particularly in high-relief areas where SRTM radar shadowing introduced voids that were filled with interpolation.

Neither DEM captures vegetation height or building mass accurately. Both are digital surface models in forested areas, which means canopy returns inflate apparent elevation. For health-access mapping this matters less than for flood modelling, because the effect is spatially consistent and the slope calculation is largely unaffected. The honest caveat: in dense equatorial forest, canopy-induced slope errors can reach 3 to 5 degrees, which is non-trivial for walking-speed estimates.

Road networks and the limits of static data

OpenStreetMap is the standard road input for low-income country health-access work, and its coverage has improved substantially since 2015. In many African and South-East Asian countries, primary and secondary roads are now well-mapped. Tertiary tracks and footpaths are not, and those are exactly the routes that rural populations use. Sentinel-2 imagery at 10 m can detect unpaved tracks as narrow as 5 to 10 m wide in open terrain, though detection degrades under tree cover and requires training data or manual validation.

The deeper problem is passability. A road that exists in OpenStreetMap and is visible in a dry-season Sentinel-2 image may be impassable during the wet season, after a flood, or in a conflict zone. Static friction surfaces built from these inputs will overestimate access for populations who live downstream of a seasonal ford or in an area where bridges have been destroyed. This is not a failure of the satellite data; it is a structural limit of any analysis that treats infrastructure as time-invariant. SAR-derived flood inundation layers can partially address the seasonal passability problem, but conflict-related road denial requires ground intelligence that no satellite currently provides.

Population grids and who is actually counted

WorldPop 100 m grids are the standard population input for catchment analysis in low- and middle-income countries. They are produced by redistributing census population counts into gridded cells using building footprints, land cover and nighttime light as ancillary variables. The approach is sound, but the output is only as good as the underlying census. Countries with censuses more than ten years old, or with significant internal displacement, will have population grids that misrepresent where people actually are.

Age and sex disaggregation matters for health-access analysis. A facility offering antenatal care is most relevant to women of reproductive age; a vaccination campaign targets children under five. WorldPop provides age-sex disaggregated grids for most countries, which allows analysts to compute not just total population beyond the one-hour isochrone but the specific demographic groups most affected. This is the level of specificity that a ministry of health needs to prioritise facility construction or mobile outreach routes.

Catchment boundaries versus service availability

A travel-time surface tells you who can physically reach a facility. It says nothing about whether that facility has a doctor present, whether it stocks the relevant medicines, or whether it is open on the day a patient arrives. Catchment analysis has a well-documented tendency to overstate effective access by conflating physical proximity with functional service delivery. The satellite-derived layer is necessary but not sufficient.

The correct framing for any client is that the travel-time surface identifies the population that is geographically excluded from care. Populations within the one-hour walking isochrone may still face barriers related to cost, language, gender norms or facility quality. Those barriers require survey data and community health records to quantify. Satellize's contribution to this analysis sits firmly on the geographic side: producing the friction surface, the catchment boundaries and the population aggregates. The service-availability layer has to come from the health system itself.

For outbreak response or environmental health emergencies, the travel-time surface becomes a prioritisation tool. If a cholera outbreak is confirmed at a cluster of villages, the friction surface immediately identifies which of those villages is more than two hours from the nearest oral rehydration facility, allowing logistics planners to pre-position supplies or route mobile teams. That is the operational use case where the method earns its cost.

Delivering this analysis at country scale

A full national health-access analysis, covering facility locations, terrain friction, road network and population aggregation, can be produced in a few days of compute time for a country the size of Uganda or Mozambique. The outputs are GIS layers: a travel-time raster, facility catchment polygons, and a population table showing how many people (by age and sex) fall outside specified time thresholds (typically 30, 60 and 120 minutes). These integrate directly into health management information systems or humanitarian coordination platforms.

Satellize runs this class of analysis on open data stacks, adding Sentinel-2 road extraction where OpenStreetMap coverage is thin. The same analytical pipeline that underpins the Tonga crop-estimation programme, combining open satellite archives with structured ground-truth validation, applies here: the geometry changes, but the discipline of knowing exactly what the data can and cannot tell you does not. If you are a ministry of health or a humanitarian organisation preparing for an outbreak response, the practical next step is a scoping call to define facility locations, the target time thresholds and the population denominator you need.

Typical figures

DEM spatial resolution	30 m (Copernicus GLO-30 and SRTM 1 arc-second); 90 m SRTM also available for rapid prototyping
DEM vertical accuracy	GLO-30: typically <4 m RMSE over open terrain; SRTM: ~9 m absolute at 90th percentile
Road extraction imagery resolution	10 m (Sentinel-2 MSI); tracks narrower than ~5 m are below reliable detection threshold
Sentinel-2 revisit	5 days at equator (combined Sentinel-2A and 2B); cloud cover may extend effective revisit in tropical regions
Population grid resolution	100 m (WorldPop); age-sex disaggregated grids available for most low- and middle-income countries
Travel-time output resolution	Typically 100 m or matched to population grid; finer outputs possible but not meaningful below DEM resolution
Archive depth	SRTM: single 2000 acquisition; GLO-30: TanDEM-X 2010–2015 baseline; Sentinel-2: 2015 to present; WorldPop: annual grids from 2000
Delivery formats	GeoTIFF friction and travel-time rasters, GeoPackage or Shapefile catchment polygons, CSV population aggregates by facility and time threshold
Minimum detectable road width (Sentinel-2)	Approximately 5–10 m in open terrain; degrades significantly under tree canopy

Analytics Satellize can run

National travel-time surface	Least-cost path accumulation over DEM-derived and road-class friction surface, using Tobler's hiking function for walking speed and published road-speed assumptions for vehicle travel	GeoTIFF raster: minutes to nearest health facility, at 100 m resolution, covering full country extent
Facility catchment polygons	Voronoi partitioning of travel-time surface, clipped to specified time thresholds (30, 60, 120 minutes), with facility attribution	GeoPackage or Shapefile of catchment boundaries, attributed with facility name, type and estimated catchment population
Underserved population table	Zonal statistics aggregating WorldPop age-sex grids over areas beyond specified travel-time thresholds	CSV and summary report: population counts by age group and sex for populations beyond 60- and 120-minute isochrones, by administrative unit
Road network extraction layer	Supervised classification or convolutional neural network segmentation of Sentinel-2 10 m imagery to detect unpaved tracks absent from OpenStreetMap	GeoPackage road centrelines with surface-type attribution, ready for integration into friction surface
Seasonal passability risk flag	Overlay of static road network with SAR-derived flood inundation extents (Sentinel-1) to identify road segments at risk of wet-season inundation	Road network GeoPackage with passability-risk attribute; separate wet-season and dry-season travel-time rasters
Outbreak prioritisation layer	Intersection of confirmed case cluster locations with travel-time surface and population grid to rank villages by access deficit and exposed population size	Ranked facility-gap report and GIS layer for logistics planning, updated as case data are provided

Sources

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.