Abstract
Nighttime light (NTL) imagery has been widely used to predict economic impacts, electricity consumption, and light pollution. In this research, we evaluated NTL imagery to predict the county-level population in the United States using deep learning methods. Rather than condensing the brightness information into a single statistic (e.g., mean annual brightness per county), this research aims to exploit the temporal nature of NTL imagery. Monthly composites of Visible Infrared Imaging Radiometer Suite (VIIRS) imagery were analyzed for each county to create quantiles of brightness. These quantiles were provided as input to two neural network architecture variants: a feedforward neural network (FFNN) and a recurrent neural network variant (long short-term memory, LSTM). A rigorous leave-one-group-out (division holdout) analysis showed that the LSTM was able to achieve a higher weighted average R2 (0.68) value across all the divisions than the baseline (−0.01) and FFNN (0.26) models, suggesting that treating NTL data as sequence data can yield more accurate predictions than traditional structured datasets. This approach of relating temporal NTL data to the population can be generalized to other parts of the world where data collection is sparse or uncertain.
