Deep learning based classification of biological aerosols

Biological aerosols mainly include viruses, bacteria, fungal and pollen, which can significantly affect the human health and environments. Accurate classification of biological aerosols contributes to further understand the implications of these aerosols in different domains. In this work, we collected the real-time fluorescence intensity, size and scattering images data of bioaerosols over a six-month period in Hong Kong by using Rapid-E particle identifier. To clustering the different types of bioaerosols, two deep leaning methods: autoencoder neural network (AE) and bidirectional long short-term memory neural network (Bilstms) were designed to extract the main features of bioaerosol fluorescence intensity and scattering images. The results showed that both AE and Bilstms could reconstruct the input bioaerosol data quite well, which illustrated that the main features they exacted were accurate. Then two clustering methods: K-means, and genie clustering were used to assign the extracted main features of bioaerosol into different clusters respectively. According to the aerosol number distribution in different clusters, the K-means clustering always presented a more uniform aerosol number distribution than genie clustering, especially for bioaerosol features extracted by Bilstms, genie believed that no matter how the number of clusters and the type of bioaerosol data changed, most aerosols were only distributed in one or two clusters. In order to assess the accuracy of clustering and obtain the species of bioaerosol in different clusters, different clusters were identified by analyzing their diurnal variation, average scattering images pattern and the relationship to the meteorological variables temperature, relative humidity, wind speed and wind direction. Based on the identification results, the accuracy of different combinations of two deep learning methods and two clustering methods in bioaerosol classification was evaluated. We believed that this work could provide the potential aid in aerosol classification methods development to achieve the easy and accurate bioaerosol identification.