Poverty reduction in low- and middle-income countries is increasingly an urban challenge, and a challenge that continues to be constrained by lack of data, including data on the spatial distribution of poverty within cities. Utilizing existing household survey data in combination with Convolutional Neural Networks (CNN) applied to high-resolution satellite images of cities, this study shows that existing data can generate detailed neighborhood-level maps providing key targeting information for an anti-poverty program. The approach is highly automatic, applicable at scale, and cost-effective. The method also provides direct support for policy development, as illustrated by the case study, where the Government of Mozambique is implementing an urban social safety net program, targeting poor urban neighborhoods, utilizing the estimated poverty maps.