Mapping poverty in low- and middle- income countries using asset-based measures and geospatial big data

Keywords: Poverty, Machine learning, Big data
Abstract Type: Virtual Paper Abstract

Abstract

It is at the research frontier to utilize geospatial big data (such as nighttime lights, land cover, climate variables, mobile networks, etc.,) and machine learning algorithms (such as random forest, support vector machines, deep learning, and transfer learning, etc.,) to predict high-resolution poverty maps in low- and middle-income countries, which supports policies of poverty eradication and inclusive development. However, current scientific progress has focused primarily on improving predictive accuracy through experimenting with different algorithms or feature sets, while the poverty measurements to be predicted are usually obsolete to greater policy impact. This work brings together a policy-relevant asset-based structural poverty measurement with machine learning algorithms to predict not only the prevalence of poverty but also the depth of poverty at high resolution. Moreover, we design and analyze different methods to generalize high-quality predictions to countries without training data. Specifically, we compare model performance when using training data from only contiguous countries, only regions sharing the same agroecological zones, and all available places. Space matters in our machine learning approach. Through innovative exploration of the outcome measurement design and spatially-explicit generalization method, we fill in the research gaps and strengthen current machine learning methods to predict high-resolution poverty maps.