Povertyreductionis increasingly an urbanchallenge, and a challenge that continues to be hampered by lack of data. One such example is the urban social safety net program implemented by the Government of Mozambique, that is spatial in nature, but works without any data on the within-city spatial distribution of poverty. The lack of detailed dataon poverty is commonin many developing as well as middle-income countries. This study appliesConvolutional Neural Networks onhigh-resolutionsatelliteimagesof cities in Mozambique,andcombines theiroutputs with householdlevel geo-referenced survey data. The results show that readily available data sources can generate detailed neighborhood-levelpoverty maps, providing key operational guidance for implementation of the urban social safety net.Importantly,the approach ishighly automatic, applicable at scale, and cost-effective. It is thus a key step forward in the application of remote sensing image recognition for urban poverty reduction.