Assessing 3-D Variability of Ultrafine Particle Using a Geo-AI Modelling Approach: A Case Study in Zhunan-Miaoli, Taiwan

Abstract

Previous air pollution modeling studies have predominantly emphasized horizontal distributions, overlooking the critical vertical variability of pollutant concentrations in urban environments. Therefore, the three-dimensional (3-D) behavior of air pollutants, and of ultrafine particulate matter (PM_0.1) in particular, is insufficiently characterized. This study examined the 3-D distribution of PM_0.1 in the Zhunan and Toufen regions in Miaoli, Taiwan. Using a hexacopter drone, PM_0.1 concentrations were measured at 12 locations, at altitudes of 40, 60, and 100 m. A geospatial-artificial intelligence (Geo-AI) model was developed to estimate 3-D PM_0.1 concentrations, incorporating databases such as land use, meteorology, and 3-D building data as predictor variables. SHapley Additive exPlanations (SHAP) analysis for variable selection showed that key predictors were building height, temperature, carbon dioxide, nitric oxide, forest coverage and 3-D spatial distance from buildings. Five machine learning algorithms were used for modeling. Extreme Gradient Boosting Regressor (XGBR) achieved the best performance with a training R² of 0.98. The model’s robustness was further examined through 10-fold cross-validation and stratified validation, which yielded R² values exceeding 0.85, indicating a strong ability to capture the spatial variation of PM_0.1 across different environmental conditions. These findings underscored the crucial role of vertical pollutant dispersion in urban environments and the need to incorporate detailed 3-D measurements into urban planning and public health policies.