Predicting Poverty through Machine Learning and Satellite Images

Abstract

Poverty remains a big challenge in many countries even with the continuous development of world economy. Effectively monitoring and predicting the poverty can be helpful for government to take proactive measurements to reduce poverty. In this research, I develop a machine learning model that leverages transfer learning, deep learning, and random forest algorithm to predict the poverty level of three African countries based on satellite images. I extracted features from satellite images through VGG-11 network and then feed them into random forest model. Through tuning the hyperparameters of the entire machine learning pipeline using GPyOpt package, I improved the model performance from the average R2 score of 20% to 46%. Furthermore, I implemented a perturbation based algorithm occlusion in Captum package to explore the feature importance of CNN model and used locally linear embedding to visualization the distribution of extracted features from different regions. My work proves that deep learning and satellite images can be used to accurately predict the poverty in regional level. The feature importance studies increase the interpretability of our model and enables the decision maker to look into the most important features for the prevention of poverty.