Gunwon Lee

Land surface temperature (LST) is a representative indicator for predicting urban heat island. The single-buffer approach, which has been used to analyze the relationship between LST and explanatory variables, cannot capture the effects of both the distance of independent variables and the area of land cover included within the buffer zone. Therefore, we aimed to investigate the relationship between LST and natural and artificial land cover in urban regions, using a multi-ring buffer model to refine independent variables in 50-m increments, by land cover type. Four tree-based machine learning models—Random Forest, LightGBM, CatBoost, and XGBoost—were compared. XGBoost exhibited the highest performance, with R² range of 0.88–0.98 across 50–500-m intervals. LST prediction analysis was effective up to 350 m. Further, the significant spatial range of variables shifted with 50-m intervals, enabling a more granular analysis of the significant spatial range of independent variables. The restoration simulation showed that greening efforts on the Western Arterial Road and Jemulpo Tunnel reduced LST within an approximate range of 150 m. These findings provide crucial scientific evidence for determining the placement of urban parks and formulating urban planning and management strategies aimed at optimizing LST reduction in urban environments.