World Bank Malawi Poverty Map — Testing Whether Satellite AI Can Target Social Cash Transfers

Malawi's anti-poverty programmes need up-to-date, small-area estimates of where poor households live, but household surveys detailed enough for fine-grained targeting are expensive and quickly go out of date.

World Bank researchers (Roy van der Weide, Brian Blankespoor, Chris Elbers and Peter Lanjouw) tested whether satellite imagery processed by machine-learning models could fill that gap.

Researchers combined Malawi's 2010/11 Integrated Household Survey with 2008 population census data to build a conventional small-area poverty estimate, then built a second poverty map using machine-learning predictions derived from satellite-based remote-sensing indicators, comparing how well each map identified Malawi's poorest areas.

The two approaches agreed on the broad geography of poverty across regions and districts but diverged substantially for many specific small areas. The remote-sensing/machine-learning map performed well for ranking larger areas (e.g. allocating a national safety-net budget across districts) but was not reliable enough on its own to decide which specific villages or households should receive benefits.

The authors explicitly caution against relying on remote-sensing poverty maps as a stand-alone replacement for household-survey data when fine-grained targeting decisions are at stake, recommending they complement rather than substitute for traditional survey-based targeting — a finding at odds with more optimistic results reported for similar methods in Nigeria.