The presence of bioparticles in the air, including fungal spores, is a major concern for human and plant health and requires robust and precise monitoring systems. While a European norm based on the manual volumetric Hirst method exists, there's a growing interest in technologies allowing automated real-time monitoring. Most of them rely on machine learning for the identification of bioaerosols. However, the diverse nature of airborne particles in terms of size, properties and composition presents challenges, among which the availability of well-curated datasets for training algorithms. While collecting reference material for pollen is relatively straightforward, current automatic monitoring methods for fungal spores rely on limited training data, hindering broader applicability. This study, which was conducted in the frame of the SYLVA project (GA no. 101086109) and the COST Action ADOPT (CA18226), aims to address this gap by outlining best practices for collecting reference material from controlled cultivation and creating datasets specifically tailored for training algorithms to classify airborne fungal spores. Critical aspects such as access to reference fungal species, in vitro cultivation, sporulation yield, clean spore isolation, dry aerosolization, and dataset cleaning have been explored for a series of 17 fungal species from the Belgian fungi collection BCCM/IHEM, including 5 Alternaria species with contrasted morphological profiles. Simple classification models were developed as proof-of-principle to assess recognition capabilities from the holography and/or fluorescence data measured by the SwisensPoleno Jupiter (Swisens AG) and laser-induced scattering and fluorescence data measured by the Plair Rapid-E+ (Plair SA). The models were trained using 80% of reference data, while 10% was used for validation to avoid overfitting during training and the remaining 10% was left aside for testing the identification performance. For Plair Rapid-E+, classification accuracy for 7 genera was shown to vary from 0.43 to 0.75 depending on the taxon (F1 score 0.577), recognizing best Botrytis cinerea and Cladosporium (class created as a mix of 3 species). For SwisensPoleno Jupiter, the initial performance obtained for classification of 8 genera by using only holographic images (F1 score 0.77) has been significantly improved by complementing them with fluorescence measurements (F1 score 0.83). Classification accuracy varied between 0.55 and 0.95 with the best performance for Curvularia lunata and Alternaria (class created as a mix of 5 species). Differentiation of species was also shown to be possible for Cladosporium, with more difficulty for some Alternaria species, while the F1 score remained good (0.72). Overall, this protocol is paving the way for more efficient, standard and accurate automatic identification of airborne fungal spores.
