Applying machine vision models to satellite images provide a tantalizing opportunity to “see” from space features and geographical areas that correspond to higher and lower levels of welfare and poverty. To date, encouraging results have been identified using night-time light emissions, as increased levels of urbanization, electrification and infrastructure correlate with wealthier geographic segments. But these approaches cover city-wide areas or bigger and do not yield the granular information that policy makers and service providers need to understand localized markets, customer demographics and how interventions impact individual beneficiaries.
This study incorporates day-time satellite imaging and machine learning methods to predict poverty at neighborhood-levels in urban and rural contexts in Uganda and Ghana. The results are encouraging but are best interpreted in terms of directional assessments of poverty, assessing neighborhoods as poor vs not-poor according to income benchmarks. The research explores strategies for measuring the impact and reach of Digital Financial Services by layering call detail records from telecom providers, mobile money data and poverty estimates. This report discusses challenges and lessons learned and offers insights for how future work can continue to advance this tantalizing research area.
