Reducing Measurement Error through IPW-Correction under Covariate Shift: Poverty Estimation using Satellite Imagery from Rural and Urban Settlements.

The aim of this study is to reduce predictive error in satellite-based poverty estimation by focusing on image re-weighting rather than through expanding data collection. Model reliability in rural regions is among the most challenging in computer vision, with standard methodologies frequently underperforming when transitioned from urban training environments. Covariate shift, visual feature disparity, and biased sampling exacerbate these issues. Addressing them requires integrating econometric methods, specifically Inverse Probability Weighting (IPW), to mitigate the impact of domain misalignment. Using a pre-trained ResNet-18 and a density ratio estimator, the study investigates how importance sampling affects predictive accuracy. The findings suggest that traditional training strategies achieve limited success in target domains, but incorporating weighted bridge samples improves outcomes. L2 Regression results present a 7.22% reduction in Mean Squared Error on rural test data, though overall accuracy remains low due to inherent data limitations. They also show that weighting urban samples with rural visual characteristics more, is associated with modest improvements in model robustness. This study shows the importance of simple solutions in reducing algorithmic bias in socio-economic AI.