At Swisens, we focus on advancing technology for measuring aerosol particles as part of our ongoing efforts. Our SwisensPoleno Jupiter showcases our commitment to automated bioaerosol observation. It utilizes holographic imaging to capture and distinguish different pollen types among individual aerosol particles.

Previous efforts of collaborative research have shown the potential improvement in classification quality when combining fluorescence with holography images. In 2025, Swisens released a multimodal classifier to overcome challenges from visually indistinguishable pollen for the first time.

Model architecture

We have developed a multimodal classifier that combines the information from the holography images and fluorescence spectra. The model architecture is designed to handle multimodal data and to learn the complex relationships between the two data sources. To allow the classifier to work on Jupiter and Mars, it was trained with a 50% chance of missing fluorescence data. This allows the model to make predictions even when the fluorescence data is not available.

The graph above shows the model architecture of the multimodal classifier. The model combines holography images (green trail) and fluorescence spectra (orange trail) to improve the classification quality of pollen particles.

Increased performance for EUMETNET Autopollen 2021

The multimodal classifier was tested on data from the EUMETNET Autopollen Intercomparison Campaign 2021 (Maya-Manzano et al., 2023 ). The results show a significant improvement in classification quality compared to the previous holography-only classifier (referenced as MSw in the publication) .

The most important taxa for allergenic pollen were evaluated in close collaboration with our customers were (taxa in the EUMETNET Autopollen Intercomparison Campaign are bold):

• Alnus
• Ambrosia
• Artemisia
• Betula
• Carpinus
• Corylus
• Cupressus
• Fagus sylvatica
• Fraxinus excelsior
• Ostrya sp.
• Picea
• Pinus
• Plantago lanceolata
• Poaceae
• Populus
• Quercus robur
• Ulmus
• Urtica

Water detection

Detection of water droplets in the air is crucial for the accurate classification of airborne pollen and has been a challenge in the past. Especially in the case of holography images, water droplets can be misclassified as pollen grains (mainly Poaceae), leading to inaccurate results.

The multimodal classifier has been trained to detect water droplets in the air and there forth to exclude them from the pollen classification, providing more accurate results. Especially due to the additional information from the fluorescence spectroscopy, the classifier can distinguish between water droplets (non-fluorescent) and pollen grains (fluorescent).

The plot above shows the false positive rates scaled to the MSw-Model. The false positive rate for water droplets is significantly reduced with the multimodal classifier. In combination with the water suppressor, the classifier can now almost perfectly distinguish between water droplets and pollen particles.

Future Work

The performance of the multimodal classifier is being tested on data from across Europe. The results are promising and show a significant improvement in classification quality compared to the previous holography-only classifier. However, it seems that the fluorescence data only slightly improves the classification quality when comparing with Holography-only*. We are working on improving the model architecture to even better utilize the fluorescence data and to further improve the classification quality and robustness to out-of-season detections.

* The new classifier can classify both without and with fluorescence data due to the chosen training strategy.