The AI pollen counter inside Melbourne’s asthma forecasting chain

The chain from a pollen grain to a health forecast starts with a physical problem: what is floating past a rooftop? Victoria combines weather forecasts with grass-pollen information. Its health department lists automated monitors at seven locations, including Melbourne, Geelong and Wodonga.[3]

A Melbourne rooftop study published on 13 October compared a SwisensPoleno Mars with a manual Burkard trap in spring 2022. They assigned grass pollen to matching categories on 72 of 88 paired days. Several authors disclosed relationships with manufacturer Swisens or project supporter AirHealth.[1]

Two views of a passing grain

Digital holography makes pollen recognition possible while particles are in flight. Light interacting with a grain produces a recorded pattern from which its image can be reconstructed. Two views preserve more shape information than one: a particle that looks ambiguous from one direction may be more recognisable from another.[2]

The foundational Poleno research described a two-stage approach. Conventional image analysis separated smooth, intact pollen shapes from rougher dust and debris. A neural network then assigned a pollen type from the reconstructed images. Training used known pollen samples introduced into the instrument, so an image came with a botanical label rather than a guess about what happened to be in the air.[2]

Recognising a grain is only half the measurement. The machine must also relate detections to the amount of air sampled and account for particles lost along the way. The original instrument research tested counting against reference particles in a calibration chamber. That is what turns image classification into an environmental measuring instrument.[2]

What the Melbourne comparison actually found

Using the MeteoSwiss 2022 classifier and taxon-specific scaling, AI identified 22 of 23 manual high-pollen days (at least 50 grains/m³); three extra high readings corresponded to moderate manual readings. No day jumped between low and high. These classify pollen, not the occurrence of an asthma emergency.[1]

At three-hour resolution, grass-pollen measurements had a mean absolute difference of 27.8 grains per cubic metre, against 11.6 for daily readings. That shorter-interval comparison used 96 observations across 12 selected days. Cypress required correction for confusion with yew, illustrating the limits of a classifier trained on European pollen.[1]

A daily average and a sequence of shorter measurements answer different questions. Imagine, for illustration, a day with a brief, heavy pollen episode followed by cleaner air. The average describes that day’s overall burden; the sequence shows when the episode passed. A warning system concerned with changing weather has reason to care about both.

From measurement to readiness

Victoria’s current system combines pollen modelling with forecasts of an uncommon combination of thunderstorms and strong winds. The monitors help verify and refine the pollen model. The forecasts cover three days during the October-to-December season and serve both the public and health and emergency services preparing for possible surges in demand.[3]

This gives automation a concrete public role: repeatedly checking the environmental input while forecasters interpret the weather. A pollen classification does not by itself say that a dangerous storm will occur. The health department also stresses that a high forecast is not certainty, and a low forecast does not rule an event out.[3]

The study tested instrument agreement, without measuring reductions in asthma admissions. It supplies evidence about the observations on which forecasting depends.[1]

Victorians can follow the combined forecast through VicEmergency and pollen information through Melbourne Pollen. The practical question remains: what is in the air, what weather is approaching, and when might they meet?[3]

Report an error or suggest a correction