Teaching Machines to Tell Pollen from Pollution in Arctic Air

High in the Arctic air, some of the most consequential particles are invisible to the naked eye. Coarse-mode aerosols — particles larger than about one micrometer — include mineral dust, sea spray, pollen fragments, fungal spores, bacteria, and even microplastics. Though they share a size range, their atmospheric roles diverge dramatically. Mineral dust can trigger ice formation in mixed-phase clouds at temperatures below about minus 15 degrees Celsius, while certain airborne bacteria can nucleate ice at temperatures as warm as minus 2 degrees. Sea spray particles, meanwhile, provide hygroscopic surfaces that host important chemical reactions. Because a handful of biological particles can disproportionately alter cloud properties, scientists urgently need reliable ways to identify what kind of particle they are looking at, in real time, in some of the most remote places on Earth.

A new study published in Atmospheric Chemistry and Physics by Aiden Jönsson of Stockholm University and colleagues tackles this identification problem head-on. The team built a comprehensive reference library of the fluorescence and optical morphology signatures of coarse aerosols from major sources — pollen, marine bacteria, dust, cellulose, and microplastics — measured in controlled laboratory experiments with a Multiparameter Bioaerosol Spectrometer, or MBS. They then trained machine learning models on these data and tested the resulting classification algorithm against a full year of field observations from the Zeppelin Observatory in Svalbard, using independent chemical tracer measurements as a reality check. The result is an open-source framework that improves bioaerosol detection while candidly exposing the limits of fluorescence-based identification.

The MBS is a single-particle ultraviolet light-induced fluorescence (UV-LIF) spectrometer developed to detect bioaerosols in ambient air. Each particle passes first through a low-power laser for sizing via Mie scattering theory, then through a high-power pulsed laser whose scattered light is recorded by two 512-pixel CMOS arrays positioned on either side of the beam. These arrays capture chords across the particle’s diffraction pattern, yielding statistical descriptors — mean signal, variance, skewness, kurtosis, peak count, peak width, and several asymmetry measures — that serve as proxies for particle shape and surface roughness. Finally, a xenon flash lamp emitting 280-nanometer light excites fluorescence, which is dispersed across eight detection channels spanning roughly 305 to 655 nanometers. The instrument can measure particles from 0.5 to about 20 micrometers at rates up to roughly 160 particles per second.

Fluorescence works as a biological marker because fluorophores such as the amino acid tryptophan and the coenzyme riboflavin are abundant in living material. But the technique has a well-known Achilles’ heel: many non-biological particles fluoresce too. Polycyclic aromatic hydrocarbons from biomass and fossil fuel combustion are highly fluorescent and can coat soot particles, causing them to glow. Plastic polymers fluoresce due to aromatic groups in their base materials or additives, or fluorophores produced during aging. The new laboratory characterization quantified these confounding similarities in detail — and found some striking examples.

Among the biological samples, bacteria cultured from Baltic seawater produced a distinctive fluorescence peak in the channel centered near 364 nanometers, consistent with tryptophan emission, and this feature proved to be a robust marker unique among the samples tested. Pollen fragments, by contrast, showed broader spectra peaking near 414 nanometers when dry, shifting toward 461 nanometers when wet-nebulized — evidence that moisture and aerosolization method can substantially alter pollen fluorescence, with implications for detecting pollen fragments produced by humid atmospheric processes such as thunderstorms. Most provocatively, freshly ground polyethylene and pure cellulose both produced fluorescence spectra closely resembling those of dry pollen fragments, meaning synthetic particles and plant debris can masquerade as bioaerosols in fluorescence-only measurements. Combustion particles from ship exhaust, observed during Arctic Ocean cruises, showed broad, intense fluorescence, and about 10 percent met the standard fluorescence-based criterion for biological particles.

To move beyond these ambiguities, the researchers built a supervised machine learning algorithm with three tasks: flag fluorescent particles that are likely combustion-derived interferents, classify fluorescent biological particles into broad subgroups (pollen fragments, bacteria, and fungal spores), and distinguish non-fluorescent dust-like from sea-spray-like particles. The interferent and dust tasks used logistic regression models, while the biological classification combined uniform manifold approximation and projection (UMAP), a dimensionality-reduction technique, with a k-nearest neighbors classifier operating in the transformed space. In testing, the pollution model identified combustion particles with 92 percent precision, and the dust model — trained only on particles larger than 2.5 micrometers, where dust and sea spray scattering signatures separate most cleanly — achieved 99 percent precision for both classes. Yet the multiclass model still misclassified roughly 12 percent of combustion particles as pollen fragments, reflecting genuine overlaps in both fluorescence and irregular morphology.

The most revealing test came when the algorithm was applied to real observations from the Zeppelin Observatory during 2020, part of the Ny-Ålesund Aerosol Cloud Experiment campaign. In its raw form, trained purely on laboratory data, the algorithm failed: its biological particle estimates correlated strongly with levoglucosan and equivalent black carbon, tracers of combustion, rather than with biological markers like arabitol and mannitol. The problem, the authors argue, is a domain gap — laboratory-generated aerosols differ from their real-world counterparts, and important particle classes such as biomass burning aerosols and plant debris were missing from the training data entirely.

The solution was domain adaptation, a transfer learning technique in which a model trained in one domain is adjusted for use in another. Using chemical tracer concentrations, the team constructed soft, continuous labels describing the degree of pollution versus biological influence for particles observed at Zeppelin, then further trained the models on these data with deliberately limited weight — capped at 20 percent of the laboratory training influence — to avoid overfitting to uncertain labels. The tuned algorithm transformed the results. It reproduced the previously published annual bioaerosol cycle, with near-zero biological particle concentrations in winter and early spring, a July peak of roughly 10 particles per liter, and a secondary autumn peak that coincided with elevated warm-activating ice-nucleating particles. In August, when biological tracer concentrations were highest, the machine learning approach detected fluorescent biological particle concentrations nearly two orders of magnitude higher than the older fluorescence-only decision tree method — levels consistent with independent offline analyses from Ny-Ålesund.

The study is equally instructive about what remains unresolved. The tuned classifier attributed most biological particles to the pollen fragment class, yet intact pollen counts in Svalbard are known to be very low, suggesting that plant debris or combustion particles with broad fluorescence spectra were likely misidentified as pollen-like. The dust model, before tuning, labeled about 80 percent of ambient non-fluorescent particles as dust — implausible given that total coarse particle concentrations correlated with sea salt tracers, not mineral dust tracers. Ambient sea spray particles, modified by atmospheric aging, apparently resemble laboratory dust more than freshly generated laboratory sea spray. These mismatches underscore that laboratory aerosolization methods can bias particle properties and that naturally aged aerosols may behave quite differently from their nascent counterparts.

The broader lesson is methodological humility paired with practical progress. No UV-LIF instrument configuration or classification scheme, the authors conclude, will fully eliminate non-biological interference, and no single method can establish definitive ground truth for bioaerosol concentrations. Parallel measurements of black carbon and source-specific chemical tracers remain essential for validating any fluorescence-based quantification. Still, the framework offers a flexible, expandable foundation: it explicitly incorporates morphology alongside fluorescence, accounts for a wider range of particle classes, and can be retrained as new characterization data arrive — particularly from biomass burning experiments, real plant debris, and studies of atmospheric aging. With the code and data openly available, the approach could be applied to reanalyze existing bioaerosol datasets worldwide, bringing climate models one step closer to accurately representing the tiny biological particles that punch far above their weight in Earth’s cloud system.

Subject of Research: Machine learning classification of coarse atmospheric aerosols using single-particle fluorescence and optical morphology

Article Title: Tracing biological, anthropogenic, and inorganic sources of coarse aerosols via single-particle fluorescence and optical morphology

Article References: Tracing biological, anthropogenic, and inorganic sources of coarse aerosols via single-particle fluorescence and optical morphology. (n.d.). https://doi.org/10.5194/acp-26-13617-2026

Image Credits: AI Generated

DOI: 10.5194/acp-26-13617-2026

Keywords: bioaerosols, coarse-mode aerosols, UV-LIF spectroscopy, fluorescence, machine learning, domain adaptation, mineral dust, sea spray aerosol, pollen fragments, Arctic, Zeppelin Observatory, ice nucleation