Urban–Rural Performance Disparities in Satellite-Based Asset Wealth Estimation Under Geographic Distribution Shift

Satellite imagery and machine learning provide a scalable approach for estimating socioeconomic conditions in regions where household survey data are limited, but strong aggregate predictive performance may conceal differences in accuracy across geographic populations. This study evaluates urban–rural performance differences in satellite-based asset wealth estimation using the SustainBench Demographic and Health Surveys benchmark, comprising 86,936 georeferenced DHS survey clusters across 48 countries. A modified eight-channel ResNet-18 was trained to predict the DHS asset wealth index from multispectral satellite imagery using the benchmark’s official country-level partitions and was evaluated on 10,543 observations from held-out countries. The model achieved an overall mean absolute error (MAE) of 0.990, root mean squared error (RMSE) of 1.265, and R² of 0.655. However, absolute prediction error was higher for rural observations: rural RMSE was 1.323 compared with 1.167 for urban observations, while rural MAE was 1.047 compared with 0.900 for urban observations. The rural–urban RMSE gap was +0.156 (95% bootstrap CI: 0.117–0.195), and the MAE gap was +0.148 (95% CI: 0.117–0.177). The disparity persisted after approximately matching urban and rural observations by asset wealth, with a rural–urban RMSE gap of +0.219 (95% CI: 0.161–0.277). Rural RMSE was also higher in 10 of 12 held-out countries with sufficient observations from both groups. These findings demonstrate that strong aggregate performance can coexist with meaningful geographic disparities in prediction error and highlight the importance of subgroup-specific evaluation when assessing satellite-based socioeconomic models.