Refugee and IDP camp growth monitoring from space
Very-high-resolution optical satellites can detect new shelters, access tracks, and latrine pits within days of their appearance, giving UNHCR and UNOSAT population proxies where ground access is denied or unsafe.
Sensors
- Maxar WorldView Legion: Panchromatic resolution of approximately 29 cm at nadir, multispectral at around 1.2 m. Up to 15 revisits per day over priority sites once the full six-satellite constellation is operational, making it the highest-cadence very-high-resolution option for rapidly expanding camps.
- Airbus Pléiades Neo: 30 cm panchromatic, 1.2 m multispectral, four-satellite constellation offering daily revisit at mid-latitudes. The four-band plus red-edge configuration supports vegetation-clearance detection around camp perimeters alongside structure mapping.
- Planet SkySat: 50 cm panchromatic, roughly 1 m multispectral, with a 21-satellite fleet capable of multiple daily tasking passes. Lower radiometric depth than Pléiades Neo but strong for change detection cadence and cost-effective area coverage.
- Sentinel-2 MSI: 10 m multispectral at 10 m in visible and near-infrared bands, five-day revisit globally and free to access. Insufficient for individual shelter detection but useful for tracking camp footprint extent, bare-soil expansion, and vegetation loss at the settlement boundary over months.
What a shelter roof gives away
A tarpaulin or corrugated-iron roof has a spectral signature quite different from bare soil or scrub. At 30 cm to 50 cm resolution, individual shelters are directly resolvable: their footprints, ridge orientations, and cast shadows are measurable features. UNOSAT has published camp-mapping workflows since at least 2013 that count shelter objects and multiply by an assumed occupancy rate, typically drawn from UNHCR field surveys, to derive population estimates. The occupancy assumption is the largest single source of error in the method, not the imagery.
Texture matters as much as spectral response. Dense, regular shelter grids produce a characteristic high-frequency texture in the grey-level co-occurrence matrix that distinguishes them from agricultural plots or urban fabric. Deep-learning segmentation models trained on labelled camp imagery, such as those published by UNOSAT and several academic groups working with WorldView and Pléiades archives, now achieve intersection-over-union scores above 0.75 on held-out test camps. Performance degrades on camps where shelters share the spectral tone of surrounding soil, a real problem in parts of the Sahel.
Access roads, latrine pits, and the other proxies
Shelter count is not the only signal. Access-road formation appears in imagery as linear bare-soil features widening over successive acquisitions. Road width and branching density correlate with population pressure and humanitarian logistics capacity. Latrine pits, visible as small dark rectangular excavations typically 1 to 2 metres across, are a sanitation-density proxy: SPHERE standards specify one latrine per 20 persons, so pit counts offer an independent cross-check on shelter-derived estimates. Both features require sub-metre imagery to be reliably detected.
Water-point queues and market areas appear at very-high resolution as aggregated human presence, though extracting meaningful counts from static imagery is unreliable. Thermal infrared from ECOSTRESS or Landsat 8/9 TIRS can indicate cooking-fire density at coarser scales, but that is a separate analytical chain. The optical proxies described here are the ones operationally used by UNHCR and UNOSAT for population estimation without ground access.
Change detection cadence and the growth-rate problem
A camp can double in area within two weeks after a displacement event. The Rohingya crisis in Cox's Bazar between August and October 2017 is the most-documented example in the public record: satellite analysis showed the Kutupalong-Balukhali expansion from roughly 10 km² to over 40 km² in approximately ten weeks, tracked by UNOSAT using WorldView imagery. That growth rate demands acquisitions at least weekly, ideally every two to three days, to produce a useful time series for humanitarian planners.
Cloud cover is the principal operational constraint. Cox's Bazar sits in a monsoon climate; sustained cloud periods of five to ten days are common. Planet SkySat's multiple daily passes increase the probability of a clear acquisition on any given day, but they do not remove cloud. Synthetic aperture radar is the cloud-independent complement for coarse structural change, though SAR cannot resolve individual shelters at the scale needed for shelter counting. The two sensor families are complementary, not interchangeable.
Archive depth matters for baseline construction. Maxar and Airbus hold commercial archives going back to the early 2000s for many conflict-affected areas, allowing analysts to establish pre-displacement land cover and detect even subtle early-phase encroachment before a camp is formally recognised.
The privacy and protection problem
Publishing precise shelter-level geospatial data about a refugee camp creates real protection risks. A georeferenced map showing which compounds belong to which ethnic or religious community, or identifying the location of safe houses and protection spaces, can be weaponised. UNHCR's data-protection guidelines explicitly restrict the publication of individual-level location data. UNOSAT's public products are typically released at block or sector level, with precise coordinates withheld or deliberately blurred for sensitive sites.
Analysts working in this space face a genuine tension. The same spatial precision that makes satellite imagery useful for population estimation also makes it dangerous in the wrong hands. Responsible practice involves tiered access: aggregated statistics and sector-level maps for public release, full-resolution products restricted to vetted humanitarian partners under data-sharing agreements. This is not a technical problem that better imagery will solve. It is a governance problem that requires deliberate institutional design.
From pixel to population figure: the honest uncertainty budget
A typical satellite-derived population estimate for a large camp carries an uncertainty of plus or minus 15 to 25 percent, depending on image quality, model training data, and the occupancy assumption used. Shelter detection itself, in good imagery conditions, can achieve completeness rates above 85 percent on well-separated structures. Densely packed shelters with shared walls or overlapping roofs cause under-counting. Temporary shade structures and cooking shelters cause over-counting.
Occupancy rates vary by emergency phase, cultural context, and season. A rate of 4.5 persons per shelter used in one East African context may be wrong by a factor of two in a South Asian or Middle Eastern setting. Satellite-derived estimates are most reliable when calibrated against at least one ground survey and updated as the emergency matures. They are not a replacement for registration data; they are a gap-filler when registration is impossible or incomplete.
Satellize runs shelter-detection and camp-footprint analytics on commercial tasking from Pléiades Neo and SkySat, combining deep-learning segmentation with Sentinel-2 time series for boundary tracking. The workflow is structurally similar to what we apply in the Tonga crop-estimation programme: object detection calibrated against a sparse ground-truth set, with explicit uncertainty bounds reported alongside the headline figure.
What a commissioning organisation should specify
Before ordering imagery or an analytics package, a humanitarian organisation needs to define three things: the minimum detectable change in camp area that is operationally meaningful, the acceptable latency between acquisition and delivered product, and the data-sharing regime that governs who sees what resolution of output.
For most UNHCR or NGO applications, a weekly shelter-count update with a 48-hour processing latency is operationally useful. Faster is better during acute emergencies, but the bottleneck is usually cloud cover and analyst capacity, not satellite availability. Specify the camp boundary polygon in advance so that tasking can be targeted precisely: commercial very-high-resolution imagery is priced per square kilometre, and a well-defined area of interest can reduce costs substantially compared with open-ended tasking. If you want to discuss a specific camp or displacement zone, the right next step is a scoping call with a named analyst, not a brochure.
Typical figures
| Best available spatial resolution (panchromatic) | ~29 cm (Maxar WorldView Legion), 30 cm (Airbus Pléiades Neo), 50 cm (Planet SkySat) |
| Multispectral resolution | 1.2 m (WorldView Legion, Pléiades Neo); ~1 m (SkySat); 10 m (Sentinel-2) |
| Revisit cadence (VHR) | Up to 15×/day (WorldView Legion at full constellation); daily (Pléiades Neo); multiple passes/day (SkySat fleet of 21) |
| Revisit cadence (medium resolution) | 5 days globally (Sentinel-2 twin satellites); free and open access |
| Minimum resolvable structure | Individual shelters ≥2 m footprint detectable at 30–50 cm resolution; latrine pits (1–2 m) detectable under good contrast conditions |
| Cloud-cover limitation | Optical imagery entirely blocked by cloud; monsoon or tropical climates may produce 5–10 consecutive clouded days |
| Typical population-estimate uncertainty | ±15–25% depending on shelter-type homogeneity, occupancy assumption, and image quality |
| Commercial archive depth | Maxar and Airbus archives extend to early 2000s for many sites; Sentinel-2 archive from 2015 |
| Typical product latency (emergency tasking) | 24–72 hours from acquisition to delivered GIS layer, depending on cloud and processing queue |
| Delivery formats | GeoTIFF orthoimage, GeoJSON or Shapefile shelter polygons, PDF sector-level population report, WMS/WMTS feed |
Analytics Satellize can run
| Shelter-count and density map | Deep-learning instance segmentation (e.g. Mask R-CNN class) trained on labelled VHR camp imagery; GLCM texture features as auxiliary input | GeoJSON polygon layer of detected shelters with density-per-hectare statistics by sector; updated on each clear acquisition |
| Camp footprint and expansion time series | Supervised land-cover classification on Sentinel-2 time series combined with VHR boundary delineation; change vector analysis between epochs | Monthly GIS layer showing camp perimeter and area in km², with annotated growth-rate chart in PDF |
| Satellite-derived population estimate | Shelter count multiplied by context-specific occupancy rate drawn from UNHCR field data; uncertainty bounds propagated from both detection error and occupancy variance | Tabular population estimate by sector with explicit ±% confidence interval; suitable for situation-report annexes |
| Access-road network extraction | Linear feature extraction on panchromatic VHR imagery using morphological filtering and road-detection CNNs; validated against OpenStreetMap where available | Shapefile of detected track network with estimated carriageway width; logistics-accessibility score per camp sector |
| Latrine-pit density proxy | Small dark-object detection on VHR multispectral imagery using spectral thresholding and shape filters; cross-checked against SPHERE sanitation ratios | Sanitation-density raster and flag layer identifying sectors below minimum SPHERE threshold |
| Vegetation-clearance alert | NDVI differencing on Sentinel-2 time series at camp perimeter buffer; threshold-based alert when clearance rate exceeds defined weekly area loss | Automated alert with affected area polygon, triggered within 24 hours of a clear Sentinel-2 pass |
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.