Predicting economic well-being in Africa using temporal satellite imagery and selfsupervised learning

Abstract

Accurate and reliable data on economic livelihoods remain scarce in the developing world, and major development agencies continue to study these outcomes and find the most effective means of assisting the impoverished. To accomplish this, wide and accurate local-level measurements of human well-being are necessary. Satellite imagery in the sense of measuring poverty has been proven a key data resource as it can fill in the resulting data gaps from scarce data. Prior research is based on predicting estimates of survey-based asset wealth from a time series of satellite images using convolutional neural networks. This work does not only implement these but also proposes a way of using self-supervised learning to improve the current state of the art models. Consequently, this work proposes novel training methods that exploit the spatio-temporal structure of remote sensing data. Through pre-training a network using contrastive learning with a MoCo framework and designated pretext tasks, one could increase the overall predictive performance in estimating poverty. The models were trained on surveys from 36 African countries and explained up to 66.4% of the variation in asset wealth at local-level locations, compared to 63.7% of an entirely supervised model.