Mapping Sectoral Employment at Fine Spatial Resolution

Abstract

We develop a deep learning approach to estimate employment by sector at fine spatial resolution using features of the built environment. A two-stage model first predicts total employment per tile (R^2 = 0.92) and then allocates employment across sectors, achieving strong accuracy. We show that commonly used gridded GDP methods substantially mischaracterize local sectoral composition, whereas our approach closely recovers observed patterns and generalizes to new countries without retraining. These results indicate that open geospatial data can be used to infer the spatial structure of economic activity, providing a new tool for economic measurement.

JEL Classification: R12, R39, Q54