Convolutional neural network for predicting poverty using satellite imagery

Abstract

Eradication of poverty in all its forms worldwide by 2030 is the rst of 17 Sustainable Development Goals outlined by the United Nations. In Africa, where the bulk of the world's poorest people live, national household surveys are used to collect data on poverty. It is di cult to obtain accurate and timely information due to the signi cant resource costs associated with conducting such surveys. Public access to abundant data sources such as daytime satellite images and nighttime lights in conjunction with advances in computer vision provides a feasible solution to the data scarcity problem. The purpose of this research is to follow previous research works in utilising machine learning techniques to process daytime satellite images and nighttime lights in order to predict the distribution of poverty at the village level in three African countries using more advanced technologies and up-to-date data. The procedure involves designing and training a Convolutional Neural Network, which is then used to extract poverty indicators from daytime satellite images. The extracted features are mapped to poverty statistics and used in a regression model to estimate poverty. In the initial nighttime light prediction task, the model attained an accuracy of 75:29% on the training data and 77:97% on the validation data over 81,000 iterations. In the target task of poverty estimation, the model explained 18% to 43% of the variation in average household expenditures in each country. The results demonstrate that the transfer learning technique is generally applicable to forecasting poverty in other countries, although its e ectiveness is highly dependent on the hyperparameters used to tune the algorithm. Additionally, the technique does not easily transfer to other poverty indicators such as child mortality and levels of education.