Comparing alder, birch and grass pollen concentrations derived from spatiotemporal statistical and dispersion models across Switzerland

Abstract

While several studies have compared the performance of statistical and dispersion modelling methods for air pollution, none have done so for pollen.

We developed a statistical machine learning model for daily pollen concentrations of five highly allergenic pollen types (hazel, alder, birch, ash, and grasses) across Switzerland (2000–2023). Daily average predictions for grass, alder and birch pollen were available for 2017–2023 from the COSMO-ART dispersion model, the operational forecast model at the Swiss Federal Office of Meteorology and Climatology. In this study, we have compared estimated concentrations for overlapping pollen types and years at three levels: (1) pollen measurement stations, (2) a 1 × 1 km national grid, and (3) residential addresses of the Swiss National Cohort.

At the grid and cohort address levels, statistical and dispersion models showed a strong correlation for grass (0.75) and birch (0.70) pollen and a moderate correlation for alder pollen (0.41). Cross-validated Pearson's correlations between statistically modelled and measured pollen concentrations ranged from 0.80 (alder), 0.85 (birch) to 0.86 (grass), with root-mean-squared logarithmic error (RMSLE) values of 0.33, 0.31 and 0.35, respectively. Pearson correlations between COSMO-ART predictions and measured concentrations were 0.41 (alder), 0.75 (birch) and 0.63 (grass), with the highest RMSLE of 0.88 for alder pollen, while RMSLEs of grass and birch models were 0.63 and 0.67, respectively.

Statistical models showed higher agreement with measured pollen concentrations at stations than COSMO-ART for all pollen types. Correlations between the two models were high for grass and birch pollen predictions, but notably lower for alder pollen.