AI Pollen Counter Passes Its Toughest Test in Thunderstorm Asthma Country

On a rooftop at the University of Melbourne, two very different machines spent the spring of 2022 breathing the same air. One was a Burkard spore trap, a device whose basic design has barely changed since 1952: it sucks air through a narrow inlet at ten litres per minute, deposits pollen grains onto a sticky microscope slide, and waits for a trained technician to peer through a microscope and count what landed there. The other was a SwisensPoleno Mars, an automated instrument that fires lasers at individual airborne particles, captures two digital holograms of each one, and hands the images to a machine-learning classifier that identifies the pollen type in near real time. The question was simple but consequential: could the robot match the microscope? The answer, published in Environmental Monitoring and Assessment, is a resounding yes for the pollen that matters most in Victoria, and a qualified yes for everything else.

The stakes are unusually high in Melbourne. In November 2016, the city experienced the world’s largest epidemic thunderstorm asthma event, a night in which a severe storm shattered grass pollen grains into respirable fragments and sent thousands of asthmatics to emergency departments, overwhelming ambulance services and contributing to multiple deaths. Since 2017, the Victorian government has operated a daily epidemic thunderstorm asthma risk forecasting system during the grass pollen season from October to December. That system combines severe weather forecasts with district-level grass pollen forecasts, and the pollen forecasts depend entirely on a statewide network of manual Hirst-type traps staffed by trained counters. Those human measurements carry a well-recognised uncertainty of roughly plus or minus thirty percent in daily pollen levels, arising from variable sampling efficiency and the inherent fallibility of human identification and counting. Any upgrade to that surveillance backbone must therefore be proven, not assumed.

The trial ran for 92 days, from October 1 to December 31, 2022, spanning the southern hemisphere spring and early summer. Over that period Melbourne recorded 265.3 millimetres of rain across 37 wet days, a maximum temperature of 37 degrees Celsius in late December, and a minimum of 5.2 degrees in late October. Both instruments sat 28.4 metres above ground level on the roof of the Earth Sciences building, roughly two kilometres north of the central business district, a siting that complies with Australian and New Zealand standards for ambient air sampling. The Burkard trap produced data for 88 of the 92 days; the automated instrument ran continuously for the entire trial. Grass and cypress pollen were the focus because together they account for nearly seventy percent of the total pollen load in Melbourne’s springtime air.

The technical comparison followed the analytical framework of a large 2023 European evaluation of nine automated pollen instruments, allowing direct benchmarking against international results. The researchers calculated the coefficient of determination, or R-squared, to measure the strength of the linear relationship between the two systems, the slope of the best-fit line to detect proportional bias, and the mean absolute error to quantify typical disagreement. For total daily pollen, the automated instrument achieved an R-squared of 0.92 with a slope of 1.1, meaning it tracked the manual counts closely while reading slightly high on average. For grass pollen, the single most important taxon for thunderstorm asthma forecasting, the agreement was equally strong: R-squared of 0.92, a slope of 1.1, and a mean absolute error of 11.6 grains per cubic metre. Cypress pollen came in at R-squared of 0.86. Those figures match or exceed the performance reported for the same instrument in Munich, a result the authors attribute partly to the dominance of grass and cypress in Melbourne’s air, both of which the classifier can recognise.

The category analysis is where the results become genuinely striking for public health. Victoria does not report raw grass pollen numbers to the public; instead, daily levels are assigned to three operational thresholds: low, from zero to 19 grains per cubic metre; moderate, from 20 to 49; and high, at 50 or more. These categories feed directly into the thunderstorm asthma risk matrix and public advisories. Across the 88 paired days, the two instruments placed 72 days in the same category. Of the 16 mismatches, 12 fell between the low and moderate boundaries and only four between moderate and high, a pattern consistent with the known tendency of co-located pollen instruments to diverge most at low concentrations. Critically, the automated system detected high-pollen days with a probability of 0.96, achieved a Gerrity skill score of 0.82, and produced no false positives and no false negatives for high-pollen days. That last detail matters enormously: false positives trigger unnecessary alerts, eroding public trust and costing agencies money, while false negatives leave vulnerable people unprotected during genuinely dangerous conditions.

Beyond daily totals, the trial tested whether the automated counter could resolve the sub-daily structure that manual counting simply cannot see. Twelve days spanning the November midpoint of the season were selected for three-hourly analysis, yielding 96 paired observations. Both instruments showed the expected diurnal rhythm, with 65 to 75 percent of each day’s pollen released during daylight hours and concentrations falling overnight. At three-hourly resolution the linear relationship weakened modestly, to R-squared of 0.86 for total pollen and 0.81 for grass, with correspondingly larger mean absolute errors, but the slopes remained close to those of the daily comparison. To validate that this rhythm was real rather than an artefact, the team compared both instruments against the Victorian Grass Pollen Emissions Module, a physics-based dispersion model driven by downscaled meteorological data. Both instruments correlated strongly with the model’s predictions, confirming that the automated system faithfully captured the day-to-night cycle that underpins understanding of how pollen exposure evolves ahead of storm events.

Cypress pollen exposed the limits of importing a classifier trained on European air. The MeteoSwiss 2022 model distinguishes Cupressus from Taxus, two conifer genera whose grains look nearly identical and are usually merged into a single category in aerobiological studies. Because yew trees are uncommon in Melbourne, the researchers reassigned all automated Taxus counts to Cupressus, applying the appropriate scaling factor. Even after this correction, the automated cypress integral came out more than double the manual one, driven largely by discrepancies late in the season. The statistics were also heavily influenced by an extreme event on October 12, when roughly forty percent of the season’s entire cypress load arrived in a single day. Removing that outlier pushed the regression slope from 1.0 to 2.2 while leaving R-squared largely intact. The authors are candid that this signals a need for locally adapted classifier training, though they stress the problem was confined to cypress and did not contaminate the grass or total pollen comparisons.

The study is honest about its constraints. The reference instrument itself is not an absolute ground truth; manual Hirst-type counting carries its own sampling and preparation uncertainties, so the trial measured operational comparability with the existing standard rather than perfect accuracy. The classifier was built for Swiss aerobiology, not Australian, which is precisely why some taxa needed local interpretation. The sub-daily analysis covered a focused subset of days, and the trial did not link pollen measurements directly to health outcomes such as ambulance call-outs or emergency presentations. Generalising to regions where native Myrtaceae pollen dominates, a family the classifier cannot identify at all, will require further evaluation across additional Australian sites and seasons. Notably, the automated instrument still recorded a total pollen integral 25 percent higher than the manual trap despite lacking classifiers for several common local pollen types, suggesting the two systems weight the airborne particle load differently.

What the trial ultimately demonstrates is a turning point for aerobiology. The SwisensPoleno Mars sampled air at 40 litres per minute, four times the flow of the Burkard trap, identified particles between 2 and 300 micrometres using light scattering and holographic imaging at a resolution of 0.595 micrometres per pixel, and streamed results continuously to an offsite server without a single technician touching a slide. For grass pollen, the taxon that decides whether Melbourne’s emergency services brace for a thunderstorm asthma event, the machine matched the human standard almost perfectly and never once confused a dangerous day with a safe one. The researchers conclude that automated counters are ready to supplement or replace manual counting in Victoria’s monitoring network, and that their real added value lies in the high temporal resolution they unlock: the ability to watch pollen concentrations rise hour by hour, exactly the kind of rapid change that precedes the storms that turned a November evening in 2016 into a public health emergency.

Subject of Research: Evaluation of an automated pollen monitor against manual Hirst-type trapping for thunderstorm asthma forecasting in Melbourne, Australia

Article Title: Evaluating automated pollen monitoring for thunderstorm asthma and public health forecasting: a comparative study in Melbourne, Australia

Article References: Evaluating automated pollen monitoring for thunderstorm asthma and public health forecasting: a comparative study in Melbourne, Australia. (n.d.). https://doi.org/10.1007/s10661-026-16017-2

Image Credits: AI Generated

DOI: 10.1007/s10661-026-16017-2

Keywords: pollen monitoring, thunderstorm asthma, automated pollen counter, SwisensPoleno Mars, Hirst trap, grass pollen, cypress pollen, machine learning, digital holography, Melbourne, public health forecasting, aerobiology