Poverty Prediction Using Satellite Imagery

“End poverty in all its forms everywhere” is the first of seventeen goals of the United Nations for 2030 (United Nations, 2015). To achieve this, there is a need for a large amount of reliable world poverty data that could be used to design and implement social programs. However, there is a lack of such data, especially in developing countries. While traditional tools of data collection such as socioeconomic surveys remain costly, there is a modern inexpensive approach based on machine learning and publicly available satellite imagery data that allows us to predict a poverty measure. This research implies the use of a convolutional neural network to accurately estimate consumption expenditure in Ethiopia and answer the question of whether this method is efficient or not. It is anticipated that this technique can identify poor regions, and it is worth spreading across other countries. Keywords: a convolutional neural network (CNN), machine learning, poverty, satellite imagery, transfer learning.