Spatiotemporal modeling of historical airborne pollen concentrations in Switzerland

Pollen is one of the most common causes of seasonal allergies. Pollen exposure is related to
allergic and respiratory diseases. Allergic disease is an important public health problem that
has risen dramatically in recent years. Climate change, which has led to rising temperature, can
affect the different allergenic pollen types, timing, and length of the pollen season. Therefore,
predicting pollen concentration is valuable to assess the risk level, and also to determine the
effects of pollen exposure on cardiovascular, respiratory, and cognitive health. However, the
prediction of airborne pollen concentrations is challenging due to the complex relationships
between biotic and environmental variables.
This dissertation investigated the spatiotemporal distributions predicted for five allergenic
pollen types, including hazel, alder, ash, birch, and grass pollen, across Switzerland. A set of
spatial and temporal predictors, including wind speed, wind direction, temperature,
precipitation, relative humidity, Normalized Difference Vegetation Index (NDVI), elevation,
land use, and tree type were used to develop spatiotemporal models to predict pollen
concentration. In addition, statistical and machine learning algorithms including LASSO,
Ridge, Elastic net, Random forest, XGBoost, and ANNs along with ensemble model evaluated
to national-wide prediction of airborne allergenic pollen concentration. Finally, a comparison
between the developed spatiotemporal machine learning model and the dispersion model
COSMO-ART was made to model pollen concentration of grass, alder, and birch pollen in
Switzerland.
The results showed the importance of the meteorological parameters in all the developed
models. Furthermore, results indicated that the land use information such as broadleaf forest,
mixed forest, agricultural land, and urban areas had an impact on pollen concentration,
typically in large buffer sizes. Despite the limited number of monitoring stations, the models
were able to explain and predict variations in the concentration of pollen at high spatial
resolution. Results indicated that the concentrations of grass and birch pollen types are notably
higher in northern Switzerland, with higher concentrations observed mainly within southern
valleys, while alder pollen concentrations reach high levels within southern valleys.
Evaluation of the performance of the various statistical, machine learning, and ensemble
methods revealed that the Random forest model is a better choice compared to alternative
machine learning methods for modelling pollen concentration. However, an ensemble model
with a combination of machine learning algorithms is a better technique for estimating
accurately pollen concentration. The impact of the different spatial and temporal variables
varies based on selected algorithms and pollen types.
The statistical assessment of the spatiotemporal machine learning model and the dispersion
model indicated a weak correlation of alder pollen and a strong correlation of grass pollen.
Furthermore, findings suggested a higher agreement of statistical models with the measured
concentrations at stations, compared to the dispersion model for all pollen types.
This was the first spatiotemporal statistical model for pollen, predicting daily exposure maps
with a fine spatial resolution of 1x1 km across Switzerland. The results of this project will be
used to estimate the pollen concentration for the panel participants in the main project named
Effects of Pollen on Cardiorespiratory Health and ALlergic symptoms (EPOCHAL) as well as
the SNC (Swiss National Cohort) and various other health studies.