Operational pollen classification using digital holography and fluorescence

Vol.: (0123456789) Aerobiologia https://doi.org/10.1007/s10453-025-09882-w BRIEF REPORT Operational pollen classification using digital holography andfluorescence BenoîtCrouzy· Marie‑PierreMeurville· BernardClot· SophieErb· MariaLbadaoui‑Darvas· FionaTummon· GianLieberherr Received: 18 July 2025 / Accepted: 9 September 2025 © The Author(s) 2025 of operation of the Swiss automatic pollen monitoring network. Keywords Pollen monitoring· Machine learning· Real-time· Digital holography· Fluorescence· Automatic identification· Airflow cytometry Pollen monitoring networks around the world have begun a radical transition from manual monitoring (Hirst, 1952) to the use of automatic systems (see, for example, Crouzy et al. (2016), Sauvageat et al. (2020), Šauliene et al. (2019), Oteros et al. (2015) and Buters etal. (2024) for a review). Over the past 7 years across Europe, the number of automatic pollen Abstract This note introduces the newly developed MeteoSwiss operational pollen classification model based on digital holography and induced fluorescence measurements. A targeted selection of curated training datasets together with a revised model architecture result in considerable improvements compared to previous operational model. The new classification model, which has been trained specifically for Switzerland, is provided openly for use in a standard format for machine learning interoperability. In addition to the description of the new classification model, we motivate the need for this development by presenting the most significant issue met during the first 5 years Benoît Crouzy, Marie-Pierre Meurville and Gian Lieberherr contributed equally to this work. B.Crouzy(*)· M.-P.Meurville(*)· B.Clot· S.Erb· M.Lbadaoui-Darvas· F.Tummon· G.Lieberherr Surface Measurements, MeteoSwiss, Chemin de l’Aérologie, 1530Payerne, Switzerland e-mail: [email protected]h M.-P. Meurville e-mail: marie-pier[email protected] B. Clot e-mail: [email protected] S. Erb e-mail: [email protected] M. Lbadaoui-Darvas e-mail: maria.lbadaoui-dar[email protected]h F. Tummon e-mail: [email protected]h G. Lieberherr e-mail: gian.lieberher[email protected]h S.Erb Environmental Remote Sensing Laboratory (LTE), École Polytechnique Fédérale de Lausanne, Station 1, 1015Lausanne, Switzerland M.Lbadaoui-Darvas Laboratory ofAtmospheric Processes andtheir Impacts (LAPI), École Polytechnique Fédérale de Lausanne, Station 1, 1015Lausanne, Switzerland M.Lbadaoui-Darvas Institute ofChemical Engineering Sciences Hellas, Foundation forResearch andTechnology (FORTH/ICEHT), 26504Hellas,Patras, Greece Aerobiologia Vol:. (1234567890) monitoring systems has increased from less than ten devices used for research to over 60 devices, many of which are used in operational monitoring networks. One example of such automatic systems is the SwisensPoleno (Swisens AG), an airflow cytometer that measures aerosols using digital holography and fluorescence. Details on the holographic technique used can be found in Sauvageat etal. (2020) and references therein. A proof-of-concept phase, testing three SwisensPolenos in parallel, was carried out in Switzerland for the full 2020 pollen season. A year later, a large instrument intercomparison campaign was held in Germany (Maya-Manzano etal., 2023) comparing nine different types of automatic pollen monitors. The results from the proof-of-concept and the intercomparison showed that the SwisensPoleno performed well. Consequently, over the 2021–2022 period, the Federal Office of Meteorology and Climatology MeteoSwiss (‘MeteoSwiss’ hereafter) installed 15 such instruments across Switzerland following a public tender. Since January 2023 this network, the SwissPollen network, has been operational using a pollen classification algorithm based on holographic images only (Sauvageat etal., 2020). Real-time observations are publicly available, and are also directly integrated into the MeteoSwiss numerical forecasting system, resulting in significantly improved forecasts (Adamov and Pauling, 2023). In order to foster collaboration and increase the reproducibility of aerobiological studies, it is essential to document the monitoring methods and machine learning models. This note shall provide such reference for the Swiss pollen monitoring network. During the first years of operational experience, several issues have been identified. These include false detection of various taxa outside their respective main pollen seasons as well as device-to-device comparability. Specifically, as shown in Fig.1, water droplets between 10 and 100 μm in size are present under saturated atmospheric conditions (e.g. fog or heavy precipitation events) and can be confused with the smooth round grass pollen grains or sometimes other pollen types. This is particularly problematic, as grass pollen is the main allergenic taxon in Switzerland (Wüthrich etal., 2009). The supervisor approach introduced by Crouzy etal. (2022) showed 0 200 400 600 Jun 21 Jun 22 Jun 23 Jun 24 Jun 25 Time Event Count Poaceae Correct identification Model prediction Wrong identification Correct identification Holography Fluorescence spectra A B C = = ≠ ≠ Fig. 1 Example of a timeseries from the SwissPollen monitoring site in Davos, Switzerland, highlighting the problem of misclassification of water droplets. A Photo of the Davos site from 23 June 2024 at 06:30 UTC. B From left to right: example of individual event identified by the algorithm as grass pollen (Poaceae) [true positive], water droplet misidentified as grass pollen [false positive], and correctly identified water droplet [true negative]. Each event is described by two holograms (200 × 200 pixels, greyscale) and measurements from 13 fluorescence channels. C Timeseries of aggregated hourly counts of aerosols classified as grass pollen between the 20–25 June 2024. The peak highlighted in grey is mostly composed of water droplets misclassified as grass pollen Aerobiologia Vol.: (0123456789) some level of success in handling out-of-season falsepositive classifications. However, in-season false-positive detection remains an issue that has so far been addressed by operational quality control procedures (manual and automatic correction). Another issue is related to species within the Betulaceae family that are difficult to distinguish from each other when using only digital holography (Sauvageat etal., 2020). To address these problems, new classification algorithms that also take fluorescence measurements as input have been developed. Limited controlled experiments have proven the potential for fluorescence to resolve confusions between certain allergenic pollen taxa (Erb etal., 2024). Following the work of Erb etal. (2024), we developed a new operational classification algorithm based on the following four principles: (1) It should flexibly handle situations where no fluorescence observations are available, since not all SwisensPoleno devices were (in the first years) equipped with fluorescence modules and not all events display a fluorescence signal above the noise threshold. This requirement can readily be accommodated by a two-branch neural network architecture able to provide classifications with and without fluorescence input. In order to avoid over-training linked to the fluorescence part of the signal we used for training a significant number of datasets without fluorescence and, on the datasets including fluorescence, we discarded 20% of the fluorescence signals during training. The fluorescence branch of the network includes few neurons in order to avoid over-complex fluorescence features (see GitHub model repository for details (MeteoSwiss biometeorology team, 2025)). (2) The algorithm must identify the main allergenic pollen taxa in Switzerland (Poaceae (grasses), Betula sp., Corylus sp., Alnus sp., Fraxinus sp., Quercus sp., and Fagus sylvatica). In addition to these taxa, several additional training datasets were included covering other pollen taxa present in significant quantities, as well as water droplets (a full list describing the training datasets can be found on GitHub (MeteoSwiss biometeorology team, 2025)). Non-biological particles are filtered out using the procedure described by Sauvageat etal. (2020) which relies on a deterministic morphological filter applied prior to the neural network. Water droplets training datasets were created from operational data during fog or rain events by manually filtering out unwanted particles based on morphology. (3) The architecture should be optimised for handling greyscale images as produced by the SwisensPoleno digital holography module. This led to the application of a simple, customised convolutional neural network rather than a pre-trained neural network optimised for RGB images as applied, for example, by Erb etal. (2024). The detailed network architecture is inspired by the VGG architecture similar to the one used in Sauvageat etal. (2020) and is presented in detail on GitHub (MeteoSwiss biometeorology team, 2025). Note that the model was trained from scratch and no pre-trained neural network was used. Finally, (4) the training datasets need to reflect the full diversity of characteristics of each pollen taxon. This can be assured by creating training datasets from different plants, locations and under different meteorological conditions (Erb etal., 2025). These datasets need to be manually cleaned to remove various artefacts such as aggregates and debris. Here we briefly present results obtained using the new MeteoSwiss operational algorithm developed in 2025. The algorithm evaluation is performed for five sites during the 2024 pollen season (Luzern: 15 January–10 June, Neuchâtel: 11 January–1 June, Payerne: 3 May–1 September, Buchs: 16 January–14 September, and Basel: 12 January–4 May). Three metrics are calculated using parallel manual measurements from Hirst-type traps. The Kendall’s Tau correla‑ tion coefficient, which is less susceptible to outliers than the Pearson coefficient, quantifies the correlation between manual and automatic measurements. The scaling factor represents the ratio between manual and automatic measurements; the lower the scaling factor, the better the sampling using the automatic system. Finally, the off‑season noise ratio serves as an indicator of the noise induced by false-positive detections of the classifier. It is defined as the ratio of the averaged daily pollen concentrations outside the defined pollen season to the averaged daily pollen concentrations during the season. The out-of-season period is defined when, over any sliding window of seven consecutive days, at least four days have average daily pollen concentrations below 20 particles/ m³, based on the manual measurements. A low offseason noise ratio means fewer incorrect warnings to the public outside of the pollen season and indicates a better signal-to-noise ratio in general. All three metrics are presented as a function of the confidence Aerobiologia Vol:. (1234567890) threshold applied to the classifier output (Crouzy et al., 2022), which provides useful information to select the optimum confidence threshold. We do not use the supervisor approach (Crouzy etal., 2022) so Aerobiologia Vol.: (0123456789) as to provide an impartial assessment of the algorithms out of the main pollen seasons. The two upper panel of Fig.2 show how the new 2025 algorithm improves grass pollen classification. The average values for all three metrics are better than for the 2022 model. Depending on the selected threshold, the spread remains low, indicating acceptable coherence between the five sites. The increase of spread for higher confidence threshold in panel (B) of Fig.2 highlights which confidence threshold should be selected as a trade-off optimising correlation while keeping sufficient sampling. Note that in an operational setup not too high confidence thresholds need to be used in order to ensure reproducibility of results. The lower panels of Fig.2 show the same three metrics for the 2025 model for the four most relevant (Wüthrich etal., 2009) additional pollen taxa, Alnus sp., Betula sp., Corylus sp., and Fraxinus sp. in order to present model performance for aero-allergen monitoring and to assess device-to-device variability. The high Kendall’s Tau correlations and low scaling factors indicate that the 2025 algorithm performs well, particularly given the limitations of manual measurements (Oteros etal., 2017). These results also show that the algorithm is capable of discriminating between taxa that are typically misclassified (Šauliene etal., 2019; Erb etal., 2024). Interestingly, for lower confidence thresholds all metrics are good, while limiting the spread across the five sites. A more detailed evaluation and discussion of various classification algorithms is beyond the scope of this short note and will be presented in a companion paper. We provide however here a self-contained reference to allow new developments based on this model and to ensure reproducibility of studies based on data from the MeteoSwiss operational pollen monitoring network. The neural network architecture, as well as the confidence thresholds, corresponding scaling factors optimised for the Swiss network and training sets, are available on GitHub (MeteoSwiss biometeorology team, 2025). To facilitate reuse, the open ONNX format is used (Bai etal., 2019). This classification algorithm may serve as a benchmark for Switzerland to evaluate future algorithms. Following the introduction of the new model in the Swiss operational network, a reanalysis of previous raw data will provide consistent timeseries for climatological applications. However, it is important to note that the model should not be used operationally in environments without prior validation, as other pollen taxa or aerosol particles as well as other environmental conditions could affect the performance in unexpected ways. Acknowledgements The authors acknowledge support of the following research funding: EU Horizon project SYLVA (Grant nbr 101086109) and 23NRM03 BioAirMet Project (BioAirMet - bioAirmet.ptb.de)). The present manuscript is a contribution to the EUMETNET AutoPollen Program. Sophie Erb was funded by the Swiss National Science Foundation (IZCOZ0_198117). We are grateful to Julia Burkard for providing Fagus pollen data from the SwisensPoleno Jupiter from the University of Vienna and to Swisens AG for sharing their model training pipeline. Author contributions BC, GL, and MM wrote the main manuscript text, BC trained, MM evaluated the model, GL supervised model development and evaluation, and MM prepared figures1 and 2. All authors contributed to drafting the manuscript and model development. Data availability The neural network architecture, as well as the confidence thresholds, corresponding scaling factors optimised for the Swiss network and training sets, is available on GitHub (https:// github. com/ Meteo Swiss/ swiss pollenmodels). To facilitate reuse, the open ONNX format is used. Declarations Conflict of interest The authors declare no conflict of interest. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. Fig. 2 Dependence of the performance metrics on the model confidence threshold for five classes with the proposed 2025 model and one class with the former 2022 model. The three metrics are computed for A grass (Poaceae) using the 2022 model and the 2025 model, for B grass (Poaceae), C Alnus sp., D Betula sp., E Corylus sp. and F Fraxinus sp., over different five sites. The red, green, and orange lines illustrate the averaged Kendall’s Tau, the scaling factor, and the off-season noise ratio, respectively. The shading delimits the minimum and maximum values for each metric across the five sites ◂ Aerobiologia Vol:. (1234567890) References Adamov, S., & Pauling, A. (2023). A real-time calibration method for the numerical pollen forecast model COSMOART. Aerobiologia, 39(3), 327–344. https:// doi. org/ 10. 1007/ s1045302309796-5. Bai, J., Lu, F., Zhang, K., etal. (2019). ONNX: Open Neural Network Exchange. GitHub. Buters, J., Clot, B., Galán, C., Gehrig, R., Gilge, S., Hentges, F., O’Connor, D., Sikoparija, B., Skjoth, C., Tummon, F., Adams-Groom, B., Antunes, C. M., Bruffaerts, N., Çelenk, S., Crouzy, B., Guillaud, G., Hajkova, L., Seliger, A. K., Oliver, G., … Stjepanovic, B. (2024). Automatic detection of airborne pollen: An overview. Aerobiologia, 40(1), 13–37. https:// doi. org/ 10. 1007/ s1045302209750-x. Crouzy, B., Lieberherr, G., Tummon, F., & Clot, B. (2022). False positives: Handling them operationally for automatic pollen monitoring. Aerobiologia, 38(3), 429–432. https:// doi. org/ 10. 1007/ s1045302209757-4. Crouzy, B. M. S., Konzelmann, T., Calpini, B., & Clot, B. (2016). All-optical automatic pollen indentification: Towards an operational system. Atmospheric Environment, 140, 202–212. https:// doi. org/ 10. 1016/j. atmos env. 2016. 05. 062. Erb, S., Berne, A., Clot, B., Lbadaoui-Darvas, M., Lieberherr, G.-D., Meurville, M.-P., Tummon, F., & Crouzy, B. (2025). Pollen holographic images and light-induced fluorescence measurements at the species level. Scientific Data, 12(1), 786. https:// doi. org/ 10. 1038/ s4159702505139-w. Erb, S., Graf, E., Zeder, Y., Lionetti, S., Berne, A., Clot, B., Lieberherr, G., Tummon, F., Wullschleger, P., & Crouzy, B. (2024). Real-time pollen identification using holographic imaging and fluorescence measurements. Atmospheric Measurement Techniques, 17(2), 441–451. https:// doi. org/ 10. 5194/ amt174412024. Hirst, J. M. (1952). An automatic volumetric spore trap. Annals of Applied Biology, 39(2), 257–265. https:// doi. org/ 10. 1111/j. 17447348. 1952. tb009 04.x. Maya-Manzano, J. M., Tummon, F., Abt, R., Allan, N., Bunderson, L., Clot, B., Crouzy, B., Daunys, G., Erb, S., Gonzalez-Alonso, M., Graf, E., Grewling, Ł, Haus, J., Kadantsev, E., Kawashima, S., Martinez-Bracero, M., Matavulj, P., Mills, S., Niederberger, E., … Buters, J. (2023). Towards European automatic bioaerosol monitoring: Comparison of 9 automatic pollen observational instruments with classic Hirst-type traps. Science of the Total Environment, 866, 161220. https:// doi. org/ 10. 1016/j. scito tenv. 2022. 161220. MeteoSwiss biometeorology team: Swisspollen-models. GitHub (2025). https:// github. com/ Meteo Swiss/ swiss pollenmodels. Oteros, J., Buters, J., Laven, G., Stefani, R., Reinhard, W., Carsten, S., & Frieder, H. (2017). Errors in determining the flow rate of Hirst-type pollen traps. Aerobiologia, 33, 201–210. https:// doi. org/ 10. 1007/ s104530169467-x. Oteros, J., Pusch, G., Weichenmeier, I., Heimann, U., Möller, R., Röseler, S., Traidl-Hoffmann, C., Schmidt-Weber, C., & Buters, J. T. M. (2015). Automatic and online pollen monitoring. International Archives of Allergy and Immunology, 167(3), 158–166. https:// doi. org/ 10. 1159/ 00043 6968. Šauliene, I., Šukiene, L., Daunys, G., Valiulis, G., Vaitkevičius, L., Matavulj, P., Brdar, S., Panic, M., Sikoparija, B., Clot, B., Crouzy, B., & Sofiev, M. (2019). Automatic pollen recognition with the rapid-e particle counter: The firstlevel procedure, experience and next steps. Atmospheric Measurement Techniques, 12(6), 3435–3452. https:// doi. org/ 10. 5194/ amt1234352019. Sauvageat, E., Zeder, Y., Auderset, K., Calpini, B., Clot, B., Crouzy, B., Konzelmann, T., Lieberherr, G., Tummon, F., & Vasilatou, K. (2020). Real-time pollen monitoring using digital holography. Atmospheric Measurement Techniques, 13(3), 1539–1550. https:// doi. org/ 10. 5194/ amt1315392020. Wüthrich, B., Schindler, C., Leuenberger, P., & AckermannLiebrich, U. (2009). Prevalence of atopy and pollinosis in the adult population of Switzerland (SAPALDIA Study). International Archives of Allergy and Immunology, 106(2), 149–156. https:// doi. org/ 10. 1159/ 00023 6836https:// arxiv. org/ abs/ https:// karger. com/ iaa/ artic lepdf/ 106/2/ 149/ 29440 12/ 00023 6836. pdf.