Pollen discrimination and classification by Fourier transform infrared (FT-IR) microspectroscopy and machine learning.

ABSTRACT

The discrimination and classification of allergy-relevant pollen was studied for the first time by mid-infrared Fourier transform infrared (FT-IR) microspectroscopy together with unsupervised and supervised multivariate statistical methods. Pollen samples of 11 different taxa were collected, whose outdoor air concentration during the flowering time is typically measured by aerobiological monitoring networks. Unsupervised hierarchical cluster analysis provided valuable information about the reproducibility of FT-IR spectra of the same taxon acquired either from one pollen grain in a 25 x 25 microm2 area or from a group of grains inside a 100 x 100 microm2 area. As regards the supervised learning method, best results were achieved using a K nearest neighbors classifier and the leave-one-out cross-validation procedure on the dataset composed of single pollen grain spectra (overall accuracy 84%). FT-IR microspectroscopy is therefore a reliable method for discrimination and classification of allergenic pollen. The limits of its practical application to the monitoring performed in the aerobiological stations were also discussed.