Finding Poverty in Satellite Images

The lack of reliable poverty data in developing countries poses a major challenge for making informed policy decisions and allocating resources effectively in those areas of the world. Unfortunately, it can be prohibitively expensive to frequently conduct comprehensive surveys that track measures of economic progress. A cheap and scalable method of producing poverty maps would greatly facilitate economic progress in these developing countries. In this paper, we present a method for predicting socioeconomic indicators directly from satellite images. We take the output of a convolutional neural network (CNN), a 4,096-element feature vector, and use these image features along with known survey data from certain parts of Uganda and Tanzania to perform linear regression on continuous wealth measures. We then use our models to predict consumption-based wealth measures and asset-based wealth measures for both Uganda and Tanzania. We find that our machine-learning based predictions approach survey accuracy at a fraction of the cost.