A Framework for the Interpretable Modeling of Household Wealth in Rural Communities From Satellite Data

Data-driven policy development and investment are necessary for aligning policies across administrative levels, targeting interventions, and meeting the 2030 Sustainable Development Goals. However, local-level economic well-being data at timely intervals, critical to informing policy development and ensuring equity of outcomes, are unavailable in many parts of the world. Yet, filling these data gaps with black-box predictive models like neural networks introduces risk and inequity to the decision-making process. In this work, we construct an alternative interpretable model to these black-box models to predict household wealth, a key socioeconomic well-being indicator, at 5-km scale from widely available satellite data. Our interpretable model promotes transparency, the identification of potential drivers of bias and harmful outcomes, and inclusive design for human-ML decision-making. We model household wealth as a low-order function of productive land use that can be interpreted and integrated with domain knowledge by decision-makers. We aggregate remotely sensed land cover change data from 2006-2019 to construct an interpretable linear regression model for household wealth and wealth change in Uganda at a 5-km scale with r2 = 72%. Our results demonstrate that there is not a clear performance-interpretability tradeoff in modeling household wealth from satellite imagery at high spatial and temporal resolution. Finally, we recommend a tiered framework to model socioeconomic outcomes from remote sensing data.