Abstract
This paper presents an ensemble machine learning approach that combines Generalized Additive Model (GAM) with eXtreme Gradient Boosting (XGBoost) to estimate and characterize the spatiotemporal distributions of elemental PM2.5 in Taiwan. Daily field measurements of 12 PM2.5 elemental components were collected from 28 air quality monitoring stations between June 2021 and May 2022. Time-variant meteorological factors and timeinvariant land-use patterns were incorporated as predictors. Results showed that the ensemble model effectively captured spatial variations in elemental PM2.5 levels, as demonstrated by the identification of numerous timeinvariant features using Shapley additive explanations analysis. A comparative analysis was conducted with a model using only XGBoost, which outperformed the ensemble model with higher cross-validated R2 and lower prediction errors. While the XGBoost-only model is recommended for exposure prediction, the ensemble model offers superior interpretability for investigating air pollution sources and aids in formulating air quality strategies from a spatial perspective.