Comparison of quantitative analysis models for atmospheric bioaerosols based on mixed fluorescence spectra

Abstract Real-time monitoring of atmospheric bioaerosol concentrations primarily involves analyzing the total fluorescence signal, which does not provide the specific content of each component. Since mixed fluorescence spectra contain information about the composition and content of bioaerosols, by employing multivariate analysis of the spectra and using multivariate regression algorithms, it is possible to extract the components from the mixed spectra and predict the proportion of each component. In this study, upon the extracted components in the mixed fluorescence spectra by vertex component analysis, we compared several quantitative analysis models for their accuracy in predicting bioaerosol proportions. The models can obtain the proportion of each component effectively when three components exist in the mixed spectra, with 2D convolutional neural networks (CNNs) having the best effect. To ascertain the algorithms’ stability, discussions on scenarios involving multiple components (four-component and five-component mixtures) as well as low-frequency noise conditions have also been included. The degree of impact on Bacillus Subtilis is the biggest when the number of mixed components increased. For the CNNs, the interference caused by the increase in the number of components is relatively higher than that of low frequency noise. When the number of the component is increased, the prediction effect of the algorithms is affected with considerable low frequency noise. Principal component regression and back propagation neural networks can maintain a relatively stable and reliable prediction effect. It is possible to match the most suitable algorithms for different components, thereby obtaining more accurate measurement results and providing a certain reference value for the detection of bioaerosol concentrations.