DPESS: Daytime Satellite Imagery-based Prediction of Demographic Attributes Using Embedding Spatial Statistics

Hyunji Cha, Sungwon Han, Donghyun Ahn, Sungwon Park, Meeyoung Cha

http://doi.org/10.5626/JOK.2020.47.8.742

Studies are being actively conducted to predict or analyze demographics used as socioeconomic factors using satellite images. We present a new approach, called DPESS, for estimating demographic attributes from daytime satellite imagery based on a deep neural network model. The four steps of the DPESS summarize any number of input images into a fixed-length embedded vector without a considerable loss of information, which is possible because of its unique structure and technique like transfer learning and embedded spatial statistics. Our extensive validation demonstrates that the DPESS model can predict various advanced demographics such as population density (R² =0.94), population count by age group (0.80), population count by education degree (0.79), and total purchase amount per household (0.80). We discuss future applications of this method in terms of applying our algorithm to other countries.