Dynamic, high-resolution poverty measurement in data-scarce environments

Accurate and comprehensive measurement of household livelihoods is critical for monitoring progress towards poverty alleviation and targeting social assistance programs for those who most need it. However, the high cost of traditional data collection has historically made comprehensive measurement a difficult task in many locations. This paper evaluates alternative satellite-based deep learning approaches to local-level livelihoods measurement, using detailed and multi-year household census extracts from four African countries as training and evaluation data. We show that new machine-learning architectures based on transformers solve multiple open measurement problems, including high performance on measuring changes in livelihoods over time and accurate measurement of local-level variation in household asset wealth within cities, offering general improvement over previous benchmarks especially when training datasets are large. Experiments that artificially restrict data availability show that satellite-based models can make accurate predictions with limited training data. The proposed approach demonstrates the promise of combining satellite imagery and new deep learning architectures for hyperlocal and dynamic measurement of poverty in data-scarce environments.