Predicting Building Height from Footprint and Urban Planning information for Digital Twin Generation

Abstract. Accurate building height data is essential for constructing realistic and analytically useful 3D city models within digital twin systems. However, such data are often incomplete, particularly in small to medium-sized cities. This study presents a machine learning-based approach to predict building heights by integrating multi-source urban data, including building footprints, zoning regulations, and roof-type classifications. To enhance prediction accuracy, we clustered zoning types by height profiles and trained models separately for each group. We evaluated three regression algorithms—Random Forest (RFR), Support Vector Regression (SVR), and XGBoost—using stratified sampling and cross-validation. Among them, RFR achieved the highest overall accuracy, particularly in homogeneous zoning areas (R² = 0.67), while SVR showed lower generalizability. The proposed model has been implemented in an open-source “3D City Model Generation Simulator,” enabling automated LOD3-level urban model generation using only remote sensing and street-view inputs. This work contributes a scalable and cost-effective solution for urban digital twin applications in data-sparse environments.