Point-to-Region Co-learning for Poverty Mapping at High Resolution Using Satellite Imagery

Poverty eradication is the first of 17 Sustainable Development Goals set by the United Nations to reach by 2030 [Jerven, 2014] . Measuring poverty is an important step to alleviating poverty, as it helps to inform research studies, target aid efforts, guide policy decisions, and generally monitor the progress of such an initiative. In the developing world, where the need for poverty data is the most pressing, poverty data is particularly scarce due to the resource cost associated with conducting surveys. This data gap is one of the crucial challenges to overcome in order to alleviate poverty. This thesis aims to cover a variety of methods towards closing this poverty data gap using remote sensing and satellite imagery data. We introduce a high-resolution poverty mapping method using only publicly available satellite data. The approach utilizes transfer learning to leverage knowledge from data-rich sources and combat the data gap. We also attempt to reduce the data gap by incorporating additional data, detailed in preliminary work using multiple resolutions of satellite imagery to improve predictive performance. We develop a semi-supervised method which narrows the data gap by utilizing abundant unlabeled satellite imagery. We show that can use this method to also take advantage of spatial correlations of poverty measures to improve our model predictions. Experiments are conducted on a variety of real-world datasets, as well as poverty measure prediction problems for 5 African countries -Malawi, Tanzania, Uganda, Nigeria, and Rwanda -demonstrating that our methods based on only publicly available data can approach the predictive performance of surveys conducted in the field and potentially transform efforts to track and alleviate poverty. Finally, we detail the implementation of a deployment pipeline system designed to support automated production of global scale poverty maps that is flexible enough to incorporate any dataset and model. This represents the first step towards providing up-to-date poverty maps to guide the decision-making process of nonprofit organizations and policymakers. a great collaborator on much of this work. I would like to thank Matthew Davis for his help on acquiring, processing, and understanding the data. I would like to thank Jake Kim for his work and collaboration on extending the transfer learning method for multiple resolutions. Lastly, I would like to thank Christopher Yeh and Jason Liu for their hard work on the development and debugging of the deployment pipeline.