Global GDP Prediction With Night-Lights and Transfer Learning

Nighttime lights (night-lights) data are useful in predicting gross domestic product (GDP), a key economic indicator used by policymakers and economists. A persistent problem in such prediction is that night-lights under-represent economic activity in rural areas. Attempting to disaggregate night-lights using urban and rural regions is problematic, as the urban-rural dichotomy is increasingly tenuous due to changing economic structures. In response, this article presents a regionalization approach, which is data-driven. Utilizing transfer learning, we trained a model that takes fine spatial resolution daytime satellite sensor imagery and learns an optimal regionalization to disaggregate visible infrared imaging radiometer suite (VIIRS) night-lights for GDP prediction. To make national scale inference feasible, we formulate a novel Monte Carlo importance sampling scheme, and then performed a single-year cross-sectional study across 178 countries using 178 000 images. This achieved an R^{2} between predicted and actual \log _{10} \text{GDP} of 0.86 on the validation set and 0.89 on the whole study area. To benchmark, we perform a subnational study over 396 U.S. counties using 98 500 images in which our method outperformed comparable methods. Interpreting the regionalization, we found that the utility of the urban-rural dichotomy is not supported by the model and argue that seeing the night-lights of some land types as representative of the overall economy is a better way to understand the model.