Palaeoecologists study the vegetation of the past - often thousands of years ago. What role do such experts have in a project aimed at providing relief to allergy sufferers around the world with high-tech pollen monitoring?
The answer: no one knows more about the shape and size of pollen than palaeoecologists specialised in palynology, who have distinguished millions of pollen grains under the microscope in their professional careers. Recognising different species is their bread and butter. This is precisely why the company Swisens from Emmen asked the Institute of Plant Sciences (IPS) of the University of Bern to collaborate with them. "Thanks to our expertise, we can not only identify pollen but also know what can be distinguished," says Christoph Schwörer, who works in the Palaeoecology section at the IPS and is a member of the Oeschger Centre for Climate Change Research (OCCR). For example, it is not possible to differentiate between wild grass species since pollen from different species looks the same. Therefore, it is also not known which grass species sufferers are allergic to.
Innovation funding for better pollen monitoring
A project supported by InnoSuisse, the Swiss Innovation Promotion Agency, has emerged from this request for collaboration. It is named "MARVEL - Emerging data services based on real-time pollen monitoring" and brings together various partners from research, development, and industry. The aim is real-time pollen monitoring that identifies pollen relevant to allergy sufferers and enables more precise pollen reports, among other things. Swisens, the company behind the project, specialises in the automatic identification of particles and already offers devices for pollen monitoring. However, its use is limited to Switzerland as the software was only developed for plant species found here.
This is not sufficient for the global ambitions of the company and the new project. "Already in the Mediterranean region, there are completely different plant species and pollen types than here," explains Christoph Schwörer. And this is where the expertise of palaeoecologists comes into play: they use pollen in lake sediments, for example, to reconstruct the vegetation history. In the annual deposits at the bottom of lakes, they find information about natural changes and human influences by determining the quantity and type of pollen grains. The members of this relatively small research community are internationally connected and share their knowledge. For example, there are databases that provide information about pollen types from all over the world. Christoph Schwörer's research group will now use this information to compile lists of the most important pollen types and possible distinguishing features for all regions of the world as part of the InnoSuisse project.
Algorithms for recognising pollen types
Based on these criteria, the monitoring devices are then trained to analyse the pollen load that they measure. This is done with the help of machine learning. Interestingly, the artificial intelligence will not learn from images as one might think, but will use standardised written descriptions. For pollen grains of the rattle pots (Rhinantus), for example, a plant genus within the summer root family, the description key reads as follows: "Psilate, spheroidal, poles mostly rounded, colpi long, often slightly sunken. Polar fields small to medium-sized. Colpi usually everywhere ± equally wide, also equatorially narrowed."
The fact that the algorithms for recognising the pollen types are developed based on such descriptions has to do, not least, with the hardware of the current monitoring device: The machine captures the passing pollen from different angles with laser beams and produces a three-dimensional image from this data. However, this representation is less precise than the images that palaeoecologists view under their light microscopes at 400x magnification.
Advancing research with the help of machine learning
Christoph Schwörer sees his research group in the "MARVEL" project as more than just a service provider. He is convinced that the software for recognising pollen types will also advance palaeoecology. After all, the researchers' own attempts to automate their tedious identification work at the microscope with the help of machine learning have not exactly been successful to date. Identifying hundreds of different types of pollen is currently still too difficult for artificial intelligence. Schwörer emphasises that this new development therefore has great potential for research.
But why are palaeoecologists so familiar with modern-day pollen? After all, their research involves analysing grains that are thousands of years old. "The shape and appearance are evolutionary very stable," says Christoph Schwörer. "Nothing has changed in the past ten to twenty thousand years. You'd have to go back millions of years or even further in the history of vegetation."