A Laboratory Evaluation of the New Automated Pollen Sensor Beenose: Pollen Discrimination Using Machine Learning Techniques

Abstract

The monitoring of airborne pollen has received much attention over the last decade, as the prevalence of pollen-induced allergies is constantly increasing. Today, the most common technique to identify airborne pollen species and to monitor their concentrations is based on manual analysis. Here, we present a new, low-cost, real-time optical pollen sensor, called Beenose, that automatically counts and identifies pollen grains by performing measurements at multiple scattering angles. We describe the data pre-processing steps and discuss the various statistical and machine learning methods that have been implemented to distinguish different pollen species. The analysis is based on a set of 12 pollen species, several of which were selected for their allergic potency. Our results show that Beenose can provide a consistent clustering of the pollen species based on their size properties, and that pollen particles can be separated from non-pollen ones. More importantly, 9 out of 12 pollen species were correctly identified with a prediction score exceeding 78%. Classification errors occur for species with similar optical behaviour, suggesting that other parameters should be considered to provide even more robust pollen identification.