ABSTRACT
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.
Subject(s)
Artificial Intelligence , Deep Learning , Spectrometry, Fluorescence , Data Collection , PollenABSTRACT
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.