Welfare estimations from imagery. A test of domain experts ability to rate poverty from visual inspection of satellite imagery

Human experts showed limited accuracy in assessing poverty from satellite imagery alone: Only 115 out of 608 clusters (19%) received identical wealth ratings from domain experts compared to the official 2015 Tanzania Demographic and Health Survey, suggesting that visual inspection has significant limitations for welfare estimation.

Visible infrastructure features are reliable poverty indicators: The study identified specific, observable characteristics that correlate with wealth levels—including modern roofing materials (slate or tile), road quality and coverage, building density, and farm sizes—which could potentially be used to train more transparent machine learning models for poverty prediction.

Expert bias toward underestimation exists: In 59% of clusters where experts' ratings differed from DHS data, they rated wealth as lower than official surveys indicated, suggesting a systematic tendency to underestimate welfare based on visual cues alone.

Satellite imagery analysis could improve transparency in AI poverty mapping: By using visually identifiable features rather than opaque algorithmic patterns, machine learning models trained on these observable infrastructure indicators could provide more interpretable and auditable poverty assessments compared to black-box approaches.

Visual poverty assessment requires caution in real-world applications: The moderate performance of domain experts highlights the risks of relying on satellite imagery analysis for high-stakes decisions affecting vulnerable populations, suggesting such tools should complement rather than replace traditional survey methods.