Temporal Poverty Prediction In Developing Countries

Poverty reduction is the first of seventeen Sustainable Development Goals that the United Nations hopes to achieve by 2030 [1]. A prerequisite to reducing poverty levels though is understanding how wealth levels are currently distributed throughout a region and how they might change in the future. Using both current information and forecasts, governments and NGOs can then effectively allocate their limited resources. Unfortunately, in many developing countries in Africa, poverty data is particularly scarce: countries conduct surveys infrequently and collect data from only a limited portion of their population [2]. The goal of this project is to reduce this data gap using publicly available data sources. In particular, we train convolutional neural networks (CNNs) to estimate poverty levels using satellite imagery, as aerial footage of a region can identify major landmarks of wealth (e.g. crop fields and buildings). We train our deep learning model to not only estimate poverty levels in a specific year but to also predict changes across years. Each type of prediction serves its own purpose. Single year predictions help fill in data for regions where no surveys were conducted. Temporal predictions can be used to forecast how the distribution of poverty will change over time. Both ultimately would be instrumental in helping policy makers improve their country’s macroeconomical decisions.