How AI Reads Satellite Images to Count Displaced People (2026)

How AI Reads Satellite Images to Count Displaced People

How AI reads satellite images to count displaced people is one of the most important technical developments in humanitarian data in the last five years. By 2026, machine learning analysis of satellite imagery has become a standard complement to ground-based assessments, producing displacement estimates for regions that are too dangerous, too remote, or too politically restricted for human enumerators to reach. UNOSAT, the Yale Humanitarian Research Lab, and multiple commercial analytics providers now publish regular satellite-based assessments of informal settlements, damage to residential areas, and changes in agricultural land use that indicate displacement.

The technology is not magic. It is pattern recognition at scale. Machine learning models are trained on thousands of labeled satellite images that show what informal settlements, tent clusters, destroyed buildings, and abandoned agricultural fields look like from above. Once trained, the models can scan new imagery and flag areas that match the patterns they have learned. The result is a map of probable displacement that can be produced in hours rather than the weeks or months that traditional assessments require.

For context on how satellite methods compare with traditional counting approaches, see AI vs Traditional Methods.

The Technical Pipeline

The satellite-based displacement estimation pipeline has four stages: image acquisition, preprocessing, model inference, and validation.

Image acquisition draws on multiple satellite constellations. Public missions like Sentinel-2 and Landsat provide free, regular coverage at moderate resolution. Commercial providers like Maxar, Planet, and Airbus offer higher-resolution imagery that can identify individual structures but at significant cost. The choice of source depends on the task: commercial imagery is typically used for settlement mapping and damage assessment, while public imagery supports broader land-use change detection.

Preprocessing corrects for atmospheric conditions, cloud cover, and variations in lighting and angle. A model trained on summer imagery may misclassify winter imagery if preprocessing does not normalize for seasonal vegetation changes. Geometric correction ensures that features in the image align with real-world coordinates, which is essential for combining satellite findings with ground data.

Model inference is where the AI does its work. Convolutional neural networks and vision transformers, architectures originally developed for consumer image recognition, are adapted to recognize humanitarian-relevant features in satellite data. The models output segmentation masks that label each pixel as "tent," "building," "vegetation," "water," or other categories. From these masks, analysts can count structures, measure settlement area, and track changes over time.

Validation compares model outputs against ground truth data where available. In some cases, drone imagery or field assessments provide independent counts that can be used to measure model accuracy. In other cases, where no ground truth exists, analysts use consistency checks: does the model output make sense given what is known about the crisis? Do multiple models agree? Are the trends consistent with other data sources?

What AI Can Detect from Space

Machine learning models can now reliably detect several displacement-relevant features. Informal settlement expansion is the most common use case. When new tent clusters appear on the edge of an existing camp or in open land near a conflict zone, models flag the change within days of new imagery becoming available. In the early weeks of the Sudan conflict, satellite-based settlement detection provided some of the first credible estimates of cross-border arrivals in Chad.

Building damage assessment is another established application. By comparing before-and-after imagery, models can classify buildings as destroyed, damaged, or intact. The method has been used extensively in Gaza, Ukraine, and Syria, producing damage assessments that ground teams could not have generated due to access restrictions.

Agricultural abandonment detection identifies farmland that has been left untended, often indicating that the population has fled. The method is particularly useful in rural conflict zones where displacement does not produce visible settlement patterns but does produce detectable land-use changes.

Limitations and Ethical Considerations

Satellite-based displacement estimates have important limitations. Resolution constraints mean that small structures, individual tents, and dense urban damage can be difficult to distinguish. Cloud cover and seasonal vegetation changes can obscure or mimic displacement signals. And models trained on one geographic context may not generalize to another: a model trained on Sahelian settlements may perform poorly in Southeast Asian rice paddies.

The ethical considerations are equally significant. Satellite imagery of displacement sites reveals the location of vulnerable populations. In conflict contexts, that information could be weaponized. Most humanitarian satellite analysis programs operate under data responsibility guidelines that restrict what is published, how it is shared, and who can access it. But the tension between transparency and protection is unresolved.

Consent is another open question. Displaced people do not choose to be visible from space. The argument that satellite imagery is already publicly available does not fully resolve the ethical concern that machine learning makes it possible to monitor populations at scale in ways that were previously impossible. For a broader discussion of data privacy in humanitarian AI, see Data Privacy for Displaced People.

Sources and Further Reading

• UNOSAT satellite analysis: https://unosat.org/
• Yale Humanitarian Research Lab: https://hrl.fas.yale.edu/
• Sentinel-2 satellite mission: https://sentinel.esa.int/web/sentinel/missions/sentinel-2
• Planet Labs commercial imagery: https://www.planet.com/
• Maxar Open Data Program: https://www.maxar.com/open-data
• REACH Initiative remote sensing: https://www.reach-initiative.org/
• ACAPS satellite analysis briefings: https://www.acaps.org/

Capabilities and limitations described above reflect publicly documented validation work by these organizations through mid 2026.