Using publicly available satellite imagery and deep learning to understand economic well-being in Africa

Nationally representative consumption or asset wealth surveys are conducted infrequently, with at least 4 years passing between surveys in the majority of African countries. This limited data availability constrains efforts to characterize who and where the poor are. Local-level measurements of human well-being are essential for informing public service delivery and policy choices by governments, targeting and evaluating livelihood programs by governmental and non-governmental organizations, and developing new products and services by the private sector. However, existing data are scarce, and traditional collection methods are expensive to scale. Other potentially relevant data for the measurement of well-being are being collected increasingly frequently, such as cloud-free imagery from multiple satellite-based sensors. Satellite imagery can be used to accurately measure local-level well-being over both space and time in Africa. Earlier work demonstrated that coarse nighttime lights imagery can measure country-level economic performance over time, and high-resolution imagery from private-sector providers can be used to measure spatial variation in local economic outcomes in a handful of developing and middle-income countries. This study focuses on using multiple sources of spatially coarser public imagery to infer both spatial and temporal differences in local-level economic well-being across sub-Saharan Africa. A deep learning model trained on multispectral satellite imagery is able to explain ~70% of the spatial variation in ground-measured village-level asset wealth across Africa, and up to 50% of temporal variation when aggregating to the district level. The model performance is limited in large part by noise in the training data. The study demonstrates how satellite-based estimates could potentially be used to help target social programs and further understand the determinants of well-being across the developing world. Asset wealth is a key economic indicator that is thought to be a less-noisy measure of households' longer-run economic well-being. A wealth index is computed from the first principal component of survey responses to questions about ownership of specific assets. The index is highly correlated with log consumption expenditure in a small subset of countries where consumption data are available. The study uses a convolutional neural network (CNN) to predict the village- and year-specific measure of wealth, using temporally and spatially matched multispectral daytime imagery from 30m/pixel Landsat and <1 km/pixel nighttime lights imagery as inputs. The study assembles data on asset wealth for >500k households living in 19,669 villages across 23 countries in Africa, drawn from nationally representative Demographic and Health Surveys (DHS) conducted between the years 2009 and 2016. The data are used to train the deep learning model, which is then applied to predict wealth estimates for a large number of villages across Africa. The study demonstrates the utility of satellite-based estimates for research and policy, and demonstrates their scalability by creating a wealth map for Africa's most populous country.