Over the years, one of the main trends in science and industry has become an automatic detection and classification of objects in real time. The cooperation between IGF and INOE 2000 (Magurele, Romania) allowed to develop a methodology for pollen classification using machine learning for in-situ measurements (Rapid-E+ sampler) with the application of neural networks. The work compares the trained models and characterizes the boundary conditions for identifying individual pollen groups. 

The developed algorithms allow to go through the entire process of pollen typing, from machine learning to visualization of predictions, taking into account supervised and unsupervised learning. In the original study, features such as size, spectrum, scattering and fluorescence lifetime were taken into account for 14 different pollen taxa. The models were trained using different variants of input feature filtration, which were analysed in terms of detection accuracy for individual cut-off thresholds of classification probability. The initial comparison of the obtained models allows to determine the ranges of feature filtration. The trained models were used to analyse the measurements collected during the operation of the sampler in the field. 

The system allows pollen grains detection in real time, taking into account the possibility of processing historical data. Work is currently underway on the use of the methodology at the University of Wrocław using the SwisensPoleno pollen meter. The acquired capabilities are the basis for the development of automatic pollen detection at the IGF station. 

The research was carried out with the financial support of the projects "Impact of allergenic pollen on the optical and microphysical properties of the urban aerosol (prePOLLEN)": NCN Preludium BIS-2 (2020/39/O/ST10/03586) and "Synergic use of Lidar and Pollen sensor for Aerosol Typing (SLaP4AT)" financed under the NAWA PRELUDIUM BIS 2 competition (PN/PRE/2022/1/00024).

Join with Google Meet: https://meet.google.com/cuq-qzgb-mzu