Development of a Weighted Average Ensemble Model for Predicting Officially Assessed Land Prices Using Grid Map Data and SHAP

This study proposes a weighted average ensemble model to predict the Officially Assessed Land Price in Sejong City, South Korea, using 500m $\times 500$ m grid-based spatial data. The model analyzes the impact of key variables through SHAP for improved interpretability. Independent variables were grouped into three categories: building characteristics, population demographics, and accessibility to social infrastructure. To construct the predictive model, three boosting algorithms—Gradient Boosting Machine (GBM), eXtreme Gradient Boosting (XGBoost), and LightGBM—were evaluated. The optimal ensemble model, weighted as GBM:XGBoost:LightGBM =0:9:1, achieved a high predictive performance with R2 of 0.8964. SHAP analysis revealed five major factors influencing land price: floor area ratio, accessibility to fire stations, total male population, number of reinforced concrete buildings, and accessibility to general hospitals. The model’s outputs were visualized through a grid-based land price map, revealing elevated land values within Sejong’s Multifunctional Administrative City and around key transportation hubs. By employing fine-scale spatial data and enhancing model interpretability through SHAP, this research presents a robust empirical framework that addresses the limitations of traditional administrative-bound analyses. The findings provide valuable insights for data-driven policy formulation and the evaluation of the real estate market.