Statistical modelling of non-stationary processes of atmospheric pollution from natural sources: example of birch pollen

A statistical model for predicting daily mean pollen concentrations during the flowering season is constructed and its parameterization and application to birch pollen in Riga (Latvia) are discussed. The model involves several steps of transformations of both meteorological data and pollen observations, aiming at a normally distributed homogeneous stationary dataset with linearized dependencies between the transformed meteorological predictors and pollen concentrations. The data transformation includes normalization of daily mean birch pollen concentrations, a switch of the independent axis from time to heat sum, a projection of governing parameters to pollen concentrations, and a reduction of non-stationarity via removal of the mean pollen season curve. These transformations resulted in a substantial improvement of statistical features of the data and, consequently, a higher efficiency of statistical procedures and better scores of the model. The transformed datasets are used for the model construction via multi-linear regression. For the application in Riga, the model coefficients were calculated using 9 years of birch pollen observations. The model was evaluated using years withheld from the training dataset. The evaluation showed robust model performance with the overall Model Accuracy exceeding 80% and Odds Ratio = 30.