This study, led by the World Bank Water Global Practice, aimed to use remote sensing to generate highly granular poverty maps of Luanda. The poverty maps were meant to be used by Angola's Regulatory Institute for Water and Energy Services (IRSEA) to better allocate subsidies to those sectors of the population which need them the most. A better allocation of subsidies would help offset household costs of connecting and paying for water services, thereby significantly expanding the percentage of the population with access to clean water. The objective of this project therefore closely aligns with marking progress towards the achievement of SDG 6.1: "By 2030, achieve universal and equitable access to safe and affordable drinking water for all." Additional effects were expected in progress along other SDGs, particularly, in helping end poverty, in reducing inequalities (SDG 1 and 10, respectively), and in strengthening government institutions (SDG 16.1). The study combined sub-meter resolution DigitalGlobe (DG) imagery with data from 1,200 household surveys, and deployed a suite of classification algorithms to explore the correspondence between survey-based poverty measures and remotely sensed household information. The exercise sought to use high-resolution satellite imagery and street-view imagery to maximize the predictive power of satellite-based measures using two continuous measures of poverty:
the household’s aggregate monthly income and
a multidimensional poverty index (MPI).
Using state-of-the-art machine learning algorithms, findings suggest that the model had relatively low levels of predictive accuracy, insufficient to provide IRSEA with any dependable mapping of urban poverty in Luanda. A city-wide prediction using nighttime and daytime imagery on an additional measure of household-wealth–a combined Wealth Index (WI), built from a principal-component analysis (PCA) of household infrastructure and assets–is provided as proof-of-concept of how these models could be used for targeting poor households for public support programs. Nevertheless, the goodness-of-fit estimates of this proof-of-concept exercise remain low.
Overall, the project was unable to produce a reliable, high-definition, poverty prediction map of Luanda. A series of lessons learned which could help improve the precision of this remote sensing model moving forward include:
refine household data-collection in order to improve the accuracy of georeferenced data,
improve the street imagery captured by the surveyors,
test the machine-learning algorithms model using alternative indicators of poverty, and
apply the model to different cities, in the hope that these are more responsive to high-quality satellite and street-view imagery, and can thus be more easily calibrated in the remote sensing exercise.
Disclaimer: As many of the links are external, some links may not work.