Spatial Demography Humanities and social sciences

Gridded Population Projection of South Korea Using Ensemble Machine Learning (2025–2060)

Jung, Jaeheon, Kim, Oh Seok, Matthews, Stephen A., Kim, KeeWhan
Springer Nature · 2026-11-05 · DOI 10.1007/s40980-026-00173-7

Abstract

This study proposes a machine learning-based dasymetric model (ML model) to spatially downscale existing population projections of South Korea at the 500 m grid level. The ML models include eXtreme Gradient Boosting (XGBoost) and Categorical Boosting (CatBoost), applied to the population projections from Statistics Korea, Korea University, and the International Institute for Applied Systems Analysis. The spatial downscaling was applied separately to urban and non-urban regions to reflect their distinct population structures; the outcomes were re-combined to portray the national population. All combinations were modeled and compared for 2025 to determine which yielded the highest performance; the final gridded population projection was produced for 2060, accordingly. All models were validated against mobile phone-based hourly de facto population data of Seoul for the urban regions and de jure population data for the non-urban regions. CatBoost performed best for both urban and non-urban nighttime, while Ordinary Least Squares (OLS) performed best for urban daytime. The Korea University projection downscaled by CatBoost emerged as the final combination for representing the nighttime gridded population. SHapley Additive exPlanations (SHAP) analysis was applied to identify the most salient input feature, Residential area. Our research provides the ideal combination of population projections and the ML models in producing fine-scale spatial resolution and long-term temporal coverage of demographic profiles of South Korea in a raster form.

Read at publisher →