Satellite AI and Refugee Camps: How Computer Vision Maps Displacement From Space

From the Rwanda camps of 1994 to the daily revisit of 2026

Satellite imagery has been used to map refugee camps since at least the Rwandan crisis of 1994, when the HIU at the US State Department and what became UNOSAT began producing manual settlement maps from commercial imagery. What has changed in the last five years is throughput. The Copernicus Sentinel constellation provides free 10-metre optical and C-band SAR data with revisit measured in days; Planet Dove and SkySat fleets push that to daily and sub-daily; Maxar WorldView offers 30 cm tasked imagery on demand. Computer vision models trained on labelled camp footprints turn this firehose into structured layers: shelter counts, camp area, road network growth, water-point detection, agricultural change in the surrounding catchment.

The model architectures doing the work

Two architecture families dominate. U-Net and its descendants (DeepLab, SegFormer) segment imagery into shelter, road, vegetation, and bare-ground classes; they are the workhorses behind UNOSAT's rapid mapping products and most NGO-run pipelines. Object detectors (YOLO variants, DETR) count individual shelters and vehicles when sub-metre imagery is available. For change detection across time, siamese networks comparing image pairs are standard. Training data scarcity remains the binding constraint: labelled refugee-camp imagery is dominated by a few well-studied sites (Za'atari, Cox's Bazar, Dadaab) and models trained on these generalise poorly to spontaneous urban displacement or to camps with non-standard shelter typologies.

What the satellite layer can and cannot tell you

It can tell you with high confidence: how many shelters are present, how the camp footprint has grown, where new roads have been compacted, whether water points and latrines are at planning density, whether nearby cropland is being cleared. It cannot tell you: who is in the camp, where they came from, why they left, or whether they are receiving services. Every credible 2026 product pairs the satellite-derived count with a registration figure from UNHCR proGres or IOM DTM and flags the gap explicitly. Reporters who quote a satellite shelter count as a population figure are overclaiming; the multiplier from shelters to people varies from roughly 4 to 7 depending on context and is itself a researched parameter.

Cloud, smoke, and the SAR fallback

Optical sensors are blind under cloud, at night, and in heavy smoke. Synthetic aperture radar (SAR) from Sentinel-1 and the ICEYE and Capella commercial constellations provides all-weather imaging at the cost of more difficult interpretation. SAR-based damage assessment, in particular, has matured rapidly since 2022 and is now the standard tool for confirming destroyed neighbourhoods in Ukraine, Gaza, and Sudan within 24-72 hours of an event. The UNOSAT rapid mapping service publishes these products under open licences.

Ethics and protection

Public dissemination of high-resolution camp imagery raises real protection concerns. Identifying individual shelters, vehicles, or people in a camp serving a persecuted population can directly enable targeting. Mainstream practice in 2026 is to publish derived layers (counts, footprints, change polygons) rather than raw imagery for active protection-sensitive sites, and to apply spatial aggregation before release. The Signal Code and the IASC data responsibility guidance are the operative references.

How to read a satellite-derived camp estimate

Ask five questions: what sensor, what date, what model and version, what was the validation accuracy on a held-out site, and what registration figure does it triangulate against? An estimate that answers all five is publishable. One that answers fewer is a starting point for reporting, not a finding.

Further reading and primary sources

• UNOSAT: https://unosat.org/
• Copernicus Open Access Hub: https://dataspace.copernicus.eu/
• Planet: https://www.planet.com/
• Maxar: https://www.maxar.com/
• HIU: https://hiu.state.gov/
• Signal Code: https://signalcodeorg.wordpress.com/
• IASC data responsibility guidance: https://interagencystandingcommittee.org/