Leveraging Satellite Imagery for Childhood Poverty Estimation

Methods using satellite imagery have been increasingly explored for analyzing demographic, health, and development indicators. This paper introduces a new dataset that pairs satellite imagery with high-quality survey data to benchmark state-of-the-art computer vision methods targeting childhood poverty estimation. The dataset includes 33,608 images from 16 countries in Eastern and Southern Africa between 1997 and 2022, integrating six childhood poverty indicators derived from Demographic and Health Surveys (DHS). Baseline approaches based on band statistics and spectral indices are compared against deep learning foundation models (e.g., DINOv2 and SatMAE), demonstrating that classical methods remain strong baselines while deep learning vision models with high-resolution input further improve accuracy. Open-source code is provided to reproduce and extend the pipeline, including dataset construction and model comparisons.