Estimating Household-Level Economic Characteristics from High-Resolution Satellite Imagery
Satej Soman, Susana Constenla-Villoslada, Emily Aiken, Joshua E. Blumenstock
Understanding economic development, land rights and management, and structural transformation requires accurate and granular measurements of poverty and growth. Fine-grained estimates of living standards are also critical to effectively target policies and evaluate development interventions (Smythe and Blumenstock, 2022; Elbers et al., 2007). However, most low-income countries lack the resources and administrative capacity to regularly collect household-level socioeconomic information (cf. Jerven, 2013). In the past decade, there has been considerable innovation in methods for constructing estimates of living standards from non-traditional sources of data. Satellite-based poverty estimates have been produced at the village level (Jean et al., 2016; Yeh et al., 2020; Engstrom et al., 2017), the neighborhood level (Smythe and Blumenstock, 2022) and the satellite tile level (e.g., tiles that are 1-2 square kilometres in area (Chi et al., 2022; Rolf et al., 2021).) In general, these studies find that machine learning and satellite data can explain a substantial amount of the in survey-based ground-truth measures of wealth and poverty (e.g., 70% explained variation in an asset index at the village level in Sub-Saharan Africa (Yeh et al., 2020) and 60% explained variation in poverty rates at the village level in Sri Lanka (Engstrom et al., 2017)). While these methods have been successful at estimating the wealth and poverty of relatively large geographic units that contain over large numbers of households, a great deal of variation in living standards is lost when households are aggregated to larger geographic units. As an example, Figure 1a highlights a 1-square kilometer region in Bangladesh, and indicates with dots the locations of 41 households that were surveyed. We observe that, even within this one small geographic area, there is considerable variation in living standards (as measured by the Progress Out of Poverty, or PPI, score); indeed the distribution of PPI scores within this 1km2 tile has roughly the same variance as the distribution for the entire region covered by the survey (Figure 1b). In this study, we explore the extent to which machine learning and high-resolution satellite imagery can be used to accurately estimate the living standards ofindividual households. Using ground truth data from a large survey in Bangladesh, we compare methods that rely on “black-box” deep learning algorithms, to more interpretable algorithms that extract specific characteristics from satellite imagery as a precursor to supervised learning.
Event: World Bank Land Conference 2024 - Washington
Only personal, non-commercial use of this document is allowed.