Development of daily 1 km resolution estimation models for outdoor BTEX using random forest with land-use data and...

Artificial intelligence (AI) can facilitate the prediction and monitoring of air pollution. Benzene, toluene, ethylbenzene, and xylene (BTEX) are harmful aromatic compounds that pose significant health risks. Conventional short-term and sparsely collected stationary monitoring data cannot fully capture the spatial and temporal variations of BTEX. Land use regression combined with machine learning has been shown to enhance the predictive capacity for outdoor BTEX, effectively addressing monitoring gaps. However, because of the scarcity of long-term BTEX measurements, only a few studies have employed this approach to estimate outdoor BTEX. In this study, we developed daily 1 km resolution BTEX models using machine learning, utilizing hourly BTEX measurements from 2011 to2020 at ten monitoring stations across Taiwan. Criteria air pollutants, land use, and meteorological variables were incorporated into the random forest models. Model performance was evaluated using 10-fold cross-validation (CV), and temporal and spatial validation. The CV coefficient of determination (R²) values for all BTEX components exceeded 0.8, with o-xylene and m,p-xylene achieving values of 0.85. Temporal validation R² values exceeded 0.8, whereas spatial validation ranged between 0.50 and 0.64. The key predictors include nitrogen dioxide, roads, crops, temples, commercial areas, and industrial areas. The model demonstrated high accuracy in identifying BTEX hotspots in major cities in Western Taiwan. Traffic was identified as the primary source of outdoor BTEX compounds, highlighting the need to implement traffic control measures to mitigate exposure. The model estimations can be applied to epidemiological research to assess both the long- and short-term health risks of BTEX exposure.