Improving Microestimates of Poverty from Satellite Images

Abstract

Accurately mapping the geographic distribution of poverty is pivotal for advancing development, yet this effort is often hampered by sparse, unreliable, and non-granular data. Using publicly available satellite images for nearly 20,000 villages in Africa, this paper demonstrates how self-supervised pre-training can enhance the accuracy and scalability of microestimates of poverty, as measured by the asset wealth index (AWI). This method outperforms a fully supervised machine learning approach by extracting more predictive features from the images, explaining approximately 72% of the survey-measured variation in AWI and surpassing the current state-of-the-art by about 3 percentage points. By offering a more accurate and scalable solution for poverty estimation, this research provides valuable insights for informed policymaking and targeted poverty alleviation.