Ahmad Daudsyah Imami, Jurng-Jae Yee
Busan is one of the southernmost metropolitan areas with the highest ozone pollution levels influenced by urban development and specific coastal meteorological conditions. The study present Cluster-Based Ensemble Regression (CBER), which consist of two-stage workflow which are benchmarking stage (Pre-CBER) and operational stage (CBER). During Pre-CBER, six machine-learning algorithms were compared, and multiple unsupervised-learning techniques were evaluated in parallel to cluster stations with similar meteorological characteristics and ozone patterns. Hyper-parameter-tuned XGBoost emerged as the most accurate regressor (RMSE = 3.69 ppb, R2 = 0.95). Nine clustering scenarios were assessed with the Silhouette score, ultimately retaining solutions based on both centroid based and density based clustering. In the CBER phase, XGBoost models were trained within each shortlisted cluster scenario and validated through leave-one-station-out tests. KNN based Meteorological Regionalization preserved fine-scale variability, sustaining R2 > 0.90 and RMSE <7 ppb in 12 of 14 clusters, while still achieving R2 averagely 0.75–0.80 in the emissions-intensive port and mountainous northeast. SHAP interpretation ranked nitrogen dioxide, temperature, solar radiation, and diurnal timing as dominant predictors. The computationally light, transparent pipeline thus converts sparse monitoring into hourly subdistrict ozone maps, providing actionable decision for Busan and other coastal cities with limited AQMS networks. © 2025
Department of ICT Integrated Ocean Smart City Engineering, Dong-A University, Busan, 49315, South Korea; Department of Environmental Engineering, Institut Teknologi Sumatera, Terusan Ryacudu, Way Hui, Jati Agung, Lampung Selatan, 35365, Indonesia