Explainable AI for unveiling deep learning pollen classification model based on fusion of scattered light patterns and fluorescence spectroscopy.

Pollen monitoring have become data-intensive in recent years as real-time detectors are deployed to classify airborne pollen grains. Machine learning models with a focus on deep learning, have an essential role in the pollen classification task. Within this study we developed an explainable framework to unveil a deep learning model for pollen classification. Model works on data coming from single particle detector (Rapid-E) that records for each particle optical fingerprint with scattered light and laser induced fluorescence. Morphological properties of a particle are sensed with the light scattering process, while chemical properties are encoded with fluorescence spectrum and fluorescence lifetime induced by high-resolution laser. By utilizing these three data modalities, scattering, spectrum, and lifetime, deep learning-based models with millions of parameters are learned to distinguish different pollen classes, but a proper understanding of such a black-box model decisions demands additional methods to employ. Our study provides the first results of applied explainable artificial intelligence (xAI) methodology on the pollen classification model. Extracted knowledge on the important features that attribute to the predicting particular pollen classes is further examined from the perspective of domain knowledge and compared to available reference data on pollen sizes, shape, and laboratory spectrofluorometer measurements.

Towards European automatic bioaerosol monitoring: Comparison of 9 automatic pollen observational instruments with classic Hirst-type traps.

To benefit allergy patients and the medical practitioners, pollen information should be available in both a reliable and timely manner; the latter is only recently possible due to automatic monitoring. To evaluate the performance of all currently available automatic instruments, an international intercomparison campaign was jointly organised by the EUMETNET AutoPollen Programme and the ADOPT COST Action in Munich, Germany (March-July 2021). The automatic systems (hardware plus identification algorithms) were compared with manual Hirst-type traps. Measurements were aggregated into 3-hourly or daily values to allow comparison across all devices. We report results for total pollen as well as for Betula, Fraxinus, Poaceae, and Quercus, for all instruments that provided these data. The results for daily averages compared better with Hirst observations than the 3-hourly values. For total pollen, there was a considerable spread among systems, with some reaching R2 > 0.6 (3 h) and R2 > 0.75 (daily) compared with Hirst-type traps, whilst other systems were not suitable to sample total pollen efficiently (R2 < 0.3). For individual pollen types, results similar to the Hirst were frequently shown by a small group of systems. For Betula, almost all systems performed well (R2 > 0.75 for 9 systems for 3-hourly data). Results for Fraxinus and Quercus were not as good for most systems, while for Poaceae (with some exceptions), the performance was weakest. For all pollen types and for most measurement systems, false positive classifications were observed outside of the main pollen season. Different algorithms applied to the same device also showed different results, highlighting the importance of this aspect of the measurement system. Overall, given the 30 % error on daily concentrations that is currently accepted for Hirst-type traps, several automatic systems are currently capable of being used operationally to provide real-time observations at high temporal resolutions. They provide distinct advantages compared to the manual Hirst-type measurements.

Integration of reference data from different Rapid-E devices supports automatic pollen detection in more locations.

Pollen is the most common cause of seasonal allergies, affecting over 33 % of the European population, even when considering only grasses. Informing the population and clinicians in real-time about the actual presence of pollen in the atmosphere is essential to reduce its harmful health and economic impact. Thus, there is a growing network of automatic particle analysers, and the reproducibility and transferability of implemented models are recommended since a reference dataset for local pollen of interest needs to be collected for each device to classify pollen, which is complex and time-consuming. Therefore, it would be beneficial to incorporate the reference dataset collected from other devices in different locations. However, it must be considered that laser-induced data are prone to device-specific noise due to laser and detector sensibility. This study collected data from two Rapid-E bioaerosol identifiers in Serbia and Italy and implemented a multi-modal convolutional neural network for pollen classification. We showed that models lost their performance when trained on data from one and tested on another device, not only in terms of the recognition ability but also in comparison with the manual measurements from Hirst-type traps. To enable pollen classification with just one model in both study locations, we first included the missing pollen classes in the dataset from the other study location, but it showed poor results, implying that data of one pollen class from different devices are more different than data of different pollen classes from one device. Combining all available reference data in a single model enabled the classification of a higher number of pollen classes in both study locations. Finally, we implemented a domain adaptation method, which improved the recognition ability and the correlations of transferred models only for several pollen classes.

Real-time automatic detection of starch particles in ambient air.

Considerable amounts of starch granules can be present in the atmosphere from both natural and anthropogenic sources. The aim of this study is to investigate the variability and potential origin of starch granules in ambient air recorded at six cities situated in a region with dominantly agricultural land use. This is achieved by using a combination of laser spectroscopy bioaerosol measurements with 1 min temporal resolution, traditional volumetric Hirst type bioaerosol sampling and atmospheric modelling. The analysis of wind roses identified potential sources of airborne starch (i.e., cereal grain storage facilities) in the vicinity of all aerobiological stations analysed in this study. The analysis of the CALPUFF dispersion model confirmed that emission of dust from the location of storage towers situated about 2.5 km north of the aerobiological station in Novi Sad is a plausible source of high airborne concentrations of starch granules. This study is important for environmental health since it contributes body of knowledge about sources, emission, and dispersion of airborne starch, known to be involved in phenomena such as thunderstorm-triggered asthma. The presented approach integrates monitoring and modelling, and provides a roadmap for examining a variety of bioaerosols previously considered to be outside the scope of traditional aerobiological measurements.

Why should we care about high temporal resolution monitoring of bioaerosols in ambient air?

This is the first time that atmospheric concentrations of individual pollen types have been recorded by an automatic sampler with 1-hour and sub-hourly resolution (i.e. 1-minute and 1-second data). The data were collected by traditional Hirst type methods and state-of the art Rapid-E real-time bioaerosol detector. Airborne pollen data from 7 taxa, i.e. Acer negundo, Ambrosia, Broussonetia papyrifera, Cupressales (Taxaceae and Cupressaceae families), Platanus, Salix and Ulmus, were collected during the 2019 pollen season in Novi Sad, Serbia. Pollen data with daily, hourly and sub-hourly temporal resolution were analysed in terms of their temporal variability. The impact of turbulence kinetic energy (TKE) on pollen cloud homogeneity was investigated. Variations in Seasonal Pollen Integrals produced by Hirst and Rapid-E show that scaling factors are required to make data comparable. Daily average and hourly measurements recorded by the Rapid-E and Hirst were highly correlated and so examining Rapid-E measurements with sub-hourly resolution is assumed meaningful from the perspective of identification accuracy. Sub-hourly data provided an insight into the heterogenous nature of pollen in the air, with distinct peaks lasting ~5-10 min, and mostly single pollen grains recorded per second. Short term variations in 1-minute pollen concentrations could not be wholly explained by TKE. The new generation of automatic devices has the potential to increase our understanding of the distribution of bioaerosols in the air, provide insights into biological processes such as pollen release and dispersal mechanisms, and have the potential for us to conduct investigations into dose-response relationships and personal exposure to aeroallergens.