기계학습 기반 전국 1 km 해상도 유해대기오염물질의 위해지도 구축과 건강영향 평가

Abstract

Hazardous air pollutants (HAPs) pose substantial long-term health risks, but nation-scale exposure assessment remains challenging because routine monitoring networks provide limited spatial coverage. A machine learning framework was developed to estimate annual HAP concentrations across South Korea from 2015 to 2022 at 1-km resolution and to evaluate their relevance to health risk assessment. Toxicity-weighted concentrations for polycyclic aromatic hydrocarbons (PAHs), volatile organic compounds (VOCs), and toxic metals, expressed as TEQBaP, TWCBD, and TWCAs, respectively, were estimated from 1-km gridded meteorological data, emissions, satellite products, ground-based air quality data, and national HAP monitoring network measurements. Annual held-out validation correlations were 0.83 for TEQBaP, 0.81 for TWCAs, and 0.67 for TWCBD. SHAP analysis and non-negative matrix factorization indicated that the major predictors differed among pollutants and identified four major sources. Cause-specific fixed-effects panel regression showed that TEQBaP was positively associated with lung cancer mortality, TWCBD was associated with ischemic heart disease, cerebrovascular disease, and cancer mortality, whereas TWCAs showed weaker associations with mortality outcomes. Local bivariate Moran’s I showed strong hotspot overlap of TEQBaP with respiratory mortality in the central inland region, limited hotspot overlap for TWCBD, and clearer hotspot overlap of TWCAs in the northeastern region. These findings suggest that high-resolution HAP estimation can support national-scale health risk assessment and spatial risk prioritization.