The quantification of markers of economic development from time-series satellite imagery using deep learning

Abstract

Archival satellite imagery contains massive quantities of largely untapped, objective data documenting the development of nations over time. A key obstacle to leveraging these images for the purposes of advancing population science research has been the lack of systematic methods for quantifying the visually observable changes in a manner that scales efficiently. This paper succeeds in quantifying economically relevant features pertaining to building development, such as square footage and geographical location, from satellite images spanning the years 2005-2009 with an F1-score of 0.8081 in a particularly challenging classification setting. We implement a principled image preprocessing pipeline and a version of the SegNet convolutional neural network architecture described by Badrinarayanan et al. (2016).