Spatiotemporal prediction of aeropollen concentration using tree-based machine learning

Pollen is a major aeroallergen and an important environmental health concern, with its concentrations strongly modulated by climate, air pollution, and vegetation species composition and abundance. This study developed and compared spatiotemporal and site-specific temporal machine learning models to predict daily concentrations of 11 pollen taxa in South Korea from 2014 to 2023. Four tree-based ensemble algorithms (Random Forest, XGBoost, LightGBM, and CatBoost) were trained using cumulative meteorological and air pollution variables, along with lagged pollen concentrations. Site-specific temporal models outperformed the spatiotemporal model, particularly for highly allergenic and spatiotemporally variable taxa such as Ambrosia, Artemisia, and grass pollen. Based on SHapley Additive exPlanations (SHAP) values that quantify individual feature contributions, the one-week-lagged weekly mean pollen concentration, cumulative temperature, and solar radiation were consistently influential, whereas air pollution metrics showed limited and taxon-dependent contributions. These findings demonstrate that climatic and pollution factors at the local scales critically shape pollen dynamics and can be leveraged to develop region-specific forecasting and early-warning systems. By linking environmental predictors with allergy-relevant exposure variability, this work contributes to data-driven public health preparedness in the context of climate change, which is expected to increase variability and extremes in pollen seasons. Our results support the establishment of predictive tools for clinical allergy management and for mitigating of pollen-related morbidity in vulnerable populations.