Estimating Daily Taxon-Specific Tree Pollen at a 1-km Resolution in Atlanta, GA from 2020 to 2024

Abstract While tree pollen is a major trigger of respiratory allergies, the lack of high-resolution, taxon-specific exposure data has limited our ability to identify local taxa. Traditional pollen monitoring networks rely heavily on human efforts, thereby being limited to one site per city. In this study, we leveraged pollen measurements from a monitoring network equipped with automated sensors at eight sampling sites across Atlanta, Georgia, and developed a modeling framework integrating atmospheric dispersion, phenology, and machine learning (Random Forest, XGBoost, and ensemble) to predict daily concentrations (grains/m3) of 13 tree taxa at a 1-km resolution from 2020 to 2024. Model performance was evaluated using random 10-fold cross-validation (CV), leave-one-site-out (LOSO) CV, and external validation against measurements from a colocated Rotorod sampler. XGBoost showed higher R2 than other models, and the median 10-fold CV R2 for pollen season was 0.64 (ranged 0.21 to 0.93) across study taxa and years. The R2s were lower for the nonpollen season (0.28) and LOSO–CV (0.31), but the Emory site (0.61) and Pinus taxa (0.76) retained higher R2s. The median Spearman R between our predictions and the Rotorod sampler was 0.40 (<0 to 0.87). These validation analyses highlighted the importance of site locations for future automated pollen monitoring and modeling.