Pollen clustering strategies using a newly developed single-particle fluorescence spectrometer

Pollen is routinely monitored and forecasted with respect to public health and allergies, but monitoring networks generally utilize a manual process of collection, analysis, and modeling that leads to poor sampling density and high measurement cost. Here, we discuss application of a single-particle fluorescence sensor recently developed for the purpose of real-time detection and recognition of pollen and spores. The sensor operates by collecting fluorescence emission spectra from many individual pollen grains sampled onto a microscope slide for each of four excitation wavelengths (280, 350, 405, and 450 nm) associated with pollen fluorophores. The sensor also records major and minor diameters of each particle. Approximately 25–30 particles for each of eight commercially purchased pollen species were interrogated. Data were analyzed using four classification methods: hierarchical agglomerative and k-means clustering (unsupervised) and random forest and gradient boosting algorithms (supervised). The purpose of the manuscript is to show development of a computational strategy to analyze spectral input data of this kind in order to support further efforts to automate sensor data collection and analysis. Both unsupervised methods showed insufficient accuracy for separating pollen species (76% k-means, 9% HAC) whereas supervised methods performed similarly well (94–95%). The random forest algorithm was then utilized to further optimize operational parameters, based on its higher computational speed. Analyzing the relative importance of each optical source for sensor performance highlighted ways that may be useful to lower sensor cost with minimal reduction to analysis quality. The results provide a framework for the application of this and similar sensors to ambient pollen detection and classification.

Copyright © 2020 American Association for Aerosol Research