AAAR 37th Annual Conference October 14 - October 18, 2019 Oregon Convention Center Portland, Oregon, USA
Abstract View
Pollen Collection Campaign: Clustering and Classification Applications Utilizing a High-Spectral Resolution UV-LIF Instrument
BENJAMIN E. SWANSON, Samir Rezgui, J. Alex Huffman, University of Denver, CO
Abstract Number: 674 Working Group: Bioaerosols
Abstract Biological aerosols are omnipresent in the atmosphere, and play a multitude of roles in both atmospheric processes and human health. Pollen, in particular, has a myriad of applications to human health, particularly in terms of allergenic response. Pollen detection, however, is frequently done either by a trained technician through visual microscopy. This, however, is time and cost-prohibitive for widespread usage to effectively map pollen allergies. As such, the ability to map real-time pollen populations in the atmosphere is relatively limited. Even recently developed real-time/commercial techniques have various hurdles, despite ultra-violet laser-induced fluorescence instrumentation being extremely power tools.
Here, we discuss our efforts over the previous pollination season (2018) to collaborate with the Denver Botanic Gardens and collect a large serious of species in attempts to discover a good mechanism for pollen clustering and classification. Spectral data was collected from 34 individual species of pollen over the 2018 growing season and classified using three different clustering/classification algorithms: Hierarchical Agglomerative Clustering, KMeans Clustering, and Random Forest Classification. Data was collected on a previously developed inexpensive single-particle fluorescence spectrometer [1] that has been shown to have efficacy in pollen detection and separation [2]. The collected pollen was broken up into various delineation types, including pollination mechanism, allergenicity, and species, and classified using the Random Forest algorithm. Species were also classified along pollination season (spring, summer, fall) in order to produce a more realistic classification scenario. In these analyses, the random forest algorithm was able to classify pollen species at a general error rate of below 10%. We also show efforts to apply this analysis to ambient collections of pollen. In efforts towards successful classification in developing an inexpensive pollen sensor, we also demonstrate initial efforts in source-reduction and the effects removing any one piece of data would have on the overall classification efficacy.
References: [1] D. R. Huffman, B. E. Swanson, and J. A. Huffman, “A wavelength-dispersive instrument for characterizing fluorescence and scattering spectra of individual aerosol particles on a substrate,” Atmos. Meas. Tech. 9(8), 3987–3998 (2016). [2] B. E. Swanson and J. A. Huffman, “Development and characterization of an inexpensive single-particle fluorescence spectrometer for bioaerosol monitoring,” Opt Express 26(3), (2018).