Conflict-driven displacement and refugee camp monitoring
Optical time-series and structure-counting methods track the emergence and growth of refugee and IDP camps, turning shelter density and roof-shadow geometry into population proxies. Honest about cloud cover, resolution limits, and the ethical weight of imaging civilian populations.
Sensors
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 5-day revisit at the equator (2-3 days with both satellites). Free and open. Sufficient to detect camp-level extent changes and gross shelter-cluster growth, but individual shelters at 3-5 m width sit near the resolution floor. Best used for change-detection time-series over months.
- Planet Dove: 3-5 m resolution, daily revisit globally. Adequate for counting large shelter blocks and tracking rapid expansion events. Radiometric consistency across the constellation has improved but remains a calibration consideration when comparing scenes from different Dove generations.
- Maxar WorldView-3: 0.31 m panchromatic, 1.24 m multispectral. At this resolution individual shelter footprints, shadow lengths, and roof materials are clearly distinguishable. Tasked commercially; revisit is opportunity-dependent and cost is significant at scale, so WorldView-3 is typically used for spot validation of lower-resolution counts rather than area-wide monitoring.
- Airbus Pléiades Neo: 0.30 m panchromatic, 0.75 m multispectral, stereo-capable. Similar niche to WorldView-3. Stereo pairs allow digital surface model generation, which can resolve ambiguities in shadow-based height estimation and improve roof-area calculations in densely packed camps.
What a shelter grid tells you before anyone counts heads
Refugee and IDP camps are among the most legible human settlements in satellite imagery. Shelters appear rapidly, are often uniform in material (UNHCR-standard plastic sheeting is highly reflective in visible bands), and are arranged in patterns that differ sharply from surrounding land cover. That contrast is analytically useful. UNOSAT has published camp-mapping products for South Sudan, Cox's Bazar, and multiple Sahel sites using exactly this logic: detect the spectral and textural signature of dense, reflective, low-height structures on previously open ground, then track how that signature grows.
The core population-proxy method rests on two measurable quantities. Shelter density counts the number of discrete roofed structures per unit area, derived from object-based image analysis or convolutional neural network detectors trained on labelled high-resolution imagery. Roof-area estimation uses the planimetric shadow cast by each shelter at a known sun elevation angle to infer structure height and, combined with footprint area, to estimate covered living space. Neither method gives you a headcount. Both give you a defensible lower bound on population that field teams can calibrate against periodic registration data.
The public record: what UNOSAT and UNHCR have actually published
UNOSAT's published analyses of the Kutupalong-Balukhali expansion in Cox's Bazar (Bangladesh) documented camp growth from roughly 10 km² to over 33 km² between August and December 2017 using a combination of Pléiades and Planet imagery. Structure counts in those reports ran into the hundreds of thousands of individual shelter objects. For South Sudan, UNOSAT mapped sites including Bentiu and Malakal using Sentinel-2 and WorldView imagery, tracking population-proxy changes across seasonal flooding cycles that periodically inundate camp perimeters.
In the Sahel, particularly in camps along the Mali-Burkina Faso-Niger corridor, the challenge is different: camps are smaller, more dispersed, and often embedded in existing settlements, making spectral separation harder. Published UNOSAT work in that region has leaned more heavily on change detection against baseline imagery rather than absolute structure counting, flagging areas of new construction rather than attempting full inventories. That is an honest reflection of what the method can and cannot resolve at Sentinel-2 resolution.
Resolution floors, cloud, and the limits of the count
Structure counting breaks down in predictable ways. At Sentinel-2's 10 m resolution, individual shelters measuring 3-4 m across are sub-pixel objects. The method works at that scale only by detecting texture and spectral anomaly at the cluster level, not by resolving individual roofs. Planet Dove at 3-5 m gets you to the point where larger shelters are marginally resolvable, but counting accuracy degrades sharply in densely packed areas where shelter edges merge. Reliable individual-structure counts require sub-1.5 m imagery, which means WorldView-3 or Pléiades Neo and the cost and tasking constraints that come with them.
Cloud cover is a persistent problem in the humid tropics. Cox's Bazar sits in one of the world's highest-rainfall zones; during the monsoon season (June to October), optical revisit effectively drops to near zero for extended periods. SAR can see through cloud, but the texture signatures of tarpaulin shelters are less distinctive in C-band backscatter than in optical imagery, and individual structure counting from SAR alone is not yet operationally proven at the scales UNOSAT works with. Analysts working these sites routinely carry multi-week data gaps in their time-series and interpolate between clear acquisitions. That interpolation introduces uncertainty that should be stated explicitly in any delivered product.
Imaging civilians: the ethical and access framework
High-resolution imaging of refugee populations raises genuine ethical questions that the humanitarian community has not fully resolved. Individuals are identifiable in sub-0.5 m imagery. Camp layouts reveal latrine locations, clinic positions, and movement patterns. In conflict settings, that information has dual-use potential. The International Charter on Space and Major Disasters provides a framework for emergency activation and data sharing, but it is oriented toward natural disasters and does not address the sustained, long-term monitoring that displacement situations often require.
UNHCR's own data-protection policies and the ICRC's guidance on digital risks in humanitarian action both flag satellite imagery as a category requiring careful handling. In practice, most published humanitarian imagery products are delivered at resolutions that obscure individual faces and avoid showing identifiable personal data, but the underlying high-resolution scenes used for analysis exist and are held by commercial vendors whose data-sharing terms vary. Organisations commissioning camp monitoring should have a clear data-minimisation policy: use the lowest resolution that answers the operational question, restrict access to processed outputs rather than raw imagery, and document the legal basis for collection. These are not abstract concerns. They are procurement requirements that affect which sensors and which data-sharing arrangements are appropriate.
Turning a time-series into an operational product
The analytic workflow for camp monitoring typically runs in three layers. The first is a baseline map: a high-resolution scene acquired before displacement, establishing land cover, road network, and any pre-existing settlement. The second is a change-detection stack: Sentinel-2 or Planet imagery at regular intervals, processed to flag new impervious surfaces and high-reflectance objects against the baseline. The third is periodic validation: a high-resolution tasked acquisition (WorldView-3 or Pléiades Neo) used to calibrate the structure-count model and check that the change-detection layer is not producing false positives from agricultural plastic, seasonal vegetation die-off, or construction unrelated to displacement.
Delivery formats for operational clients range from GIS polygon layers (camp perimeter and internal zone boundaries updated on a defined cadence) to tabular population-proxy estimates with confidence intervals. Alert products can flag threshold-crossing events, for example a 15% increase in roof-area extent within a 30-day window, triggering a notification without requiring an analyst to review every acquisition. Satellize applies this kind of threshold-alert architecture in its analytics work, including the crop-estimation programme it runs for the Kingdom of Tonga, and the same pipeline logic transfers directly to camp-monitoring contexts.
One practical note: archive depth matters. Planet's Dove constellation has global daily coverage back to approximately 2016. Sentinel-2 archive runs from 2015. For situations where displacement began before a monitoring programme was commissioned, retrospective analysis is possible and often essential for understanding the rate and pattern of camp formation.
Typical figures
| Typical spatial resolution (monitoring) | 10 m (Sentinel-2), 3-5 m (Planet Dove) |
| Typical spatial resolution (structure counting) | 0.30-0.75 m (WorldView-3, Pléiades Neo) |
| Revisit cadence (open data) | 2-5 days (Sentinel-2 dual satellite); daily (Planet Dove) |
| Revisit cadence (commercial tasking) | Opportunity-dependent; typically 1-4 days at mid-latitudes for WorldView-3 or Pléiades Neo |
| Minimum detectable camp extent | Approximately 0.5-1 ha cluster at Sentinel-2 resolution; individual shelters resolvable above ~1.5 m GSD |
| Cloud-cover limitation | Optical methods fail under persistent cloud; monsoon-season gaps of 4-8 weeks are documented in Cox's Bazar analyses |
| Archive depth | Sentinel-2 from 2015; Planet Dove from 2016; WorldView commercial archive from 2009 |
| Spectral bands used | Visible (RGB), NIR for vegetation/impervious separation; panchromatic for structure sharpening |
| Delivery formats | GeoTIFF change layers, GeoJSON/Shapefile perimeter polygons, CSV population-proxy tables, threshold alert feeds |
| Latency (open-data pipeline) | 24-72 hours from acquisition to processed output for Sentinel-2 and Planet |
Analytics Satellize can run
| Camp perimeter and extent map | Supervised land-cover classification and morphological change detection against pre-displacement baseline | GeoJSON polygon layer, updated on defined cadence (weekly or monthly) |
| Structure count and roof-area estimate | Object-based image analysis or CNN detector applied to sub-1.5 m imagery; shadow-length geometry for height inference | Tabular count with confidence interval, georeferenced centroid layer |
| Population-proxy time-series | Roof-area or shelter-density proxy calibrated against UNHCR registration data where available | Monthly trend chart with upper/lower bounds; CSV export for integration with field data systems |
| Rapid-expansion alert | Threshold detection on normalised difference built-up index or spectral anomaly score in rolling 30-day window | Automated alert with scene thumbnail and extent-change percentage, delivered by email or API push |
| Seasonal flood-risk overlay | DEM-derived inundation modelling combined with camp perimeter layer; Sentinel-1 SAR water-extent mapping during flood events | Risk-zone GIS layer showing shelter clusters within modelled 1-in-5-year flood extent |
| Retrospective camp-formation timeline | Archive time-series analysis using Planet and Sentinel-2 historical imagery; keyframe selection at inflection points | Annotated image sequence with dates and extent measurements, suitable for briefing or legal documentation |
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.