CNN 기반 주간 위성 이미지를 활용한 격자 단위 인구추정

Abstract

The traditional methods of collecting population data are population census and vital data surveys. Although most countries still use this method, changes are taking place in the collection and utilization of population data according to the ICT environment and the trend of opening public data. An example is the register-based census using administrative data introduced by Korea in 2015 and population estimation for subregions that are more detailed than administrative districts. In particular, studies using satellite images for population estimation of subregions are being introduced. In the early days, night-time satellite images were mainly used, but after the deep learning CNN algorithm was introduced, studies using day-time satellite images have been introduced. In this study, we tried to estimate the 1km gridded population using day-time satellite images and CNN transfer learning in Korea. As a result of the study, the possibility of estimation was confirmed, but satisfactory results were not obtained like previous studies. In order to obtain improved gridded population estimates, it was confirmed that it may be advantageous to apply the CNN algorithm to a grid smaller than 1 km in a relatively small area such as a city rather than a large area such as the entire country.