Nobody Agrees How Many People Are in the Camp

Humanitarian response planning depends on accurate population estimates to size food distributions, water systems, health services, and shelter. But in displacement settings, three data sources — UNHCR registration, satellite imagery analysis, and ground sampling surveys — routinely produce divergent estimates with 15–30% gaps. Registration undercounts people who avoid formal registration; satellites overcount by misidentifying abandoned structures as occupied shelters; surveys have sampling biases in fluid populations. No methodology integrates these sources into a reconciled estimate with quantified uncertainty.

A 15–30% population estimation error translates directly into under-provisioned or wasted food, water, and medical supplies. In the Sudan crisis (2024), conflicting population estimates across agencies led to contested resource allocations. In Gaza (2023–2024), population movement data was so uncertain that humanitarian corridors could not be reliably planned. The difference between 15% undercount and 15% overcount can mean the difference between famine and waste.

UNHCR partnered with WorldPop to combine registration data with satellite imagery, achieving 83–93% accuracy for formal camp footprints — but accuracy degrades sharply for urban refugees, spontaneous settlements, and populations in movement. ML models count tent-like structures from satellite images but cannot distinguish occupied from abandoned, or determine household size. IOM's Displacement Tracking Matrix conducts ground-level tracking but depends on enumerator access denied in conflict zones. Mobile phone CDR data offers movement tracking but is unavailable in many displacement contexts (low phone ownership, destroyed towers). The OCHA Centre for Humanitarian Data centralizes available data but cannot resolve contradictions between sources.

A formal data fusion framework that integrates satellite structure counts, registration records, phone mobility data (where available), and ground-truth sampling into a Bayesian or ensemble estimate with explicit uncertainty quantification. This requires better models of systematic biases in each data source — e.g., satellite overcounts by X% in arid vs. vegetated settings, registration undercounts by Y% under specific access conditions. Small-area estimation techniques adapted for non-stationary (moving) populations.

A team could build a Bayesian population estimation model that integrates satellite-derived structure counts with registration data for a single well-documented displacement setting, calibrating the bias correction factors and reporting uncertainty bounds. Alternatively, a team could analyze satellite imagery of a known refugee camp to develop occupied-vs-abandoned structure classification methods. Statistics, remote sensing, and humanitarian data science skills apply.