Predicting atmospheric particle formation days by Bayesian classification of the time series features
Research output: Contribution to journal › Article › Scientific › peer-review
|Number of pages||10|
|Journal||Tellus, Series B: Chemical and Physical Meteorology|
|Publication status||Published - 2018|
|Publication type||A1 Journal article-refereed|
Atmospheric new-particle formation (NPF) is an important source of climatically relevant atmospheric aerosol particles. NPF can be directly observed by monitoring the time-evolution of ambient aerosol particle size distributions. From the measured distribution data, it is relatively straightforward to determine whether NPF took place or not on a given day. Due to the noisiness of the real-world ambient data, currently the most reliable way to classify measurement days into NPF event/non-event days is a manual visualization method. However, manual labor, with long multi-year time series, is extremely time-consuming and human subjectivity poses challenges for comparing the results of different data sets. These complications call for an automated classification process. This article presents a Bayesian neural network (BNN) classifier to classify event/non-event days of NPF using a manually generated database at the SMEAR II station in Hyytiälä, Finland. For the classification, a set of informative features are extracted exploiting the properties of multi-modal log normal distribution fitted to the aerosol particle concentration database and the properties of the time series representation of the data at different scales. The proposed method has a classification accuracy of 84.2 % for determining event/non-event days. In particular, the BNN method successfully predicts all event days when the growth and formation rate can be determined with a good confidence level (often labeled as class Ia days). Most misclassified days (with an accuracy of 75 %) are the event days of class II, where the determination of growth and formation rate are much more uncertain. Nevertheless, the results reported in this article using the new machine learning-based approach points towards the potential of these methods and suggest further exploration in this direction.