This research paper explores the extent to which information extracted from satellite images can help predict the distribution of extreme urban poverty in Ethiopia, Tanzania, and Mozambique.
Household surveys have become the main source of data and information on economic wellbeing in low income countries and policymakers have made extensive use of poverty mapping techniques (Elbers et al., 2003) that combine such surveys with concurrent census data to obtain more detailed knowledge on the spatial distribution of welfare. However, household survey data often has gaps.
Other poverty mapping techniques are evolving rapidly. Using machine learning algorithms to detect and classify objects in images, these high-resolution poverty maps could be useful for policy makers and urban planners, especially with the scaling up of social protection programmes post Covid-19 lockdowns.
This research paper explores the extent to which information extracted from satellite images can help predict the distribution of extreme urban poverty in Ethiopia, Tanzania, and Mozambique. Geo-referenced estimates of daily per capita consumption from nationally representative household surveys are matched with observable neighbourhood level characteristics, such as the average size, perimeter, and density of buildings, as well as distances to different types of road and other points of interest. By comparing cities across three countries with different poverty profiles and material characteristics, a broader range of possible outcomes can be described.
The main data sources were household surveys, building footprints, and data on roads and distance, used with two of the most widely accepted methods in machine learning predictions today- Random Forest (RF) and Extreme Gradient Boosting (XGB).
The main purpose of this paper is to predict welfare at high levels of detail in three large urban centres of East Africa using available spatial data derived from satellite images.
Available at: https://doi.org/10.55158/DEEPWP10