10th International Aerosol Conference September 2 - September 7, 2018 America's Center Convention Complex St. Louis, Missouri, USA
Abstract View
Identification of New Particle Formation Events with Deep Learning
Jorma Joutsensaari, MATTHEW OZON, Tuomo Nieminen, Santtu Mikkonen, Timo Lähivaara, Stefano Decesari, M. Cristina Facchini, Ari Laaksonen, Kari Lehtinen, University of Eastern Finland
Abstract Number: 1130 Working Group: Remote/Regional Atmospheric Aerosol
Abstract New particle formation (NPF) in the atmosphere is observed frequently in different environments in the boundary layer. It has been estimated that 30–50% of global tropospheric cloud condensation nuclei concentrations might be formed by NPF. Currently, NPF events are typically classified into different event classes manually from the measurement data by researchers. This is time consuming and the identification of event type might be inconsistent. To get more reliable and consistent results, the NPF event analysis should be automatized.
We have developed a novel deep learning-based method to identify automatically NPF events (Joutsensaari et al., 2018). The method is based on image recognition of daily-measured particle size distribution data. In image analysis, we utilized a commonly available pre-trained Convolutional Neural Network (CNN), named AlexNet, which was transfer learned to recognize NPF event classes (six different types). In transfer learning, a subset of particle size distribution images were used in the training stage of the CNN and the rest of images for testing the success of the training. The event classification was done directly from existing data (figures) without any pre-processing of the data. To our knowledge, this is the first time when an automatic method has been successfully used to classified NPF event into different classes.
We utilized the developed method for a 15-year long dataset measured at San Pietro Capofiume in Italy. The results showed that clear event (i.e., even types of Classes 1 and 2) and non-event days could be identified with an accuracy of ca. 80 %, when the CNN-based classification was compared with visual-based classification by researchers. In the event analysis, the choice between different event classes could be ambiguous and thus overlapping between different classes occurs easily. The results showed that the overlapping occurs typically between the adjacent event classes, e.g., a manually classified Class 1 was categorized into Classes 1 and 2 by CNN, etc. In general, the manual and CNN classifications were very consist for most of the days. Clear misclassifications seemed to occur more commonly in manual analysis than in the CNN categorization, mainly due to a wrong or an incorrectly listed classification by researchers.
In general, the automatic CNN-based classification seems to have a better reliability and repeatability in NPF event classification than human-made visual-based classification. Typically, an analysis of large long-term datasets requires manual labor of several researchers, which is very time consuming and the quality of the analysis may vary. Therefore, the transfer learned pre-trained CNNs are powerful tools to analyze any long-term datasets more accurately and consistently in a sufficiently short time, which helps us to understand in detail aerosol-climate interactions and the long-term effects of climate change on NPF in the atmosphere.
Joutsensaari, J., Ozon, M., Nieminen, T., Mikkonen, S., Lähivaara, T., Decesari, S., Facchini, M. C., Laaksonen, A., and Lehtinen, K. E. J.: Identification of new particle formation events with deep learning, Atmos. Chem. Phys. Discuss., https://doi.org/10.5194/acp-2017-1189, in review, 2018.