Browsing by Subject "atmospheric science"

Atmospheric new-particle formation (NPF) plays a crucial role in generating climate-influencing aerosol particles. Direct observation of NPF is achievable by tracking the evolution of aerosol particle size distributions in the environment. Such analysis allows researchers to determine the occurrence of NPF on specific days. Currently, the most dependable method for categorizing days into NPF event (class Ia, class Ib, class II) or non-event categories relies on manual visual analysis. However, this manual process is labor-intensive and subjective, particularly with long- term data series. These issues underscore the need for an automated classification system to classify these days more objectively. This paper introduces feature-engineering based machine learning classifiers to discern NPF event and non-event days at the SMEAR II station in Hyytiälä, Finland. The classification utilizes a suite of informative features derived from the multi-modal log-normal distribution fitted to the aerosol particle concentration data and time series analysis at various scales. The proposed machine learning classifiers can achieve an accuracy of more than 90% in identifying NPF event and non-event days. Moreover, they are able to reach an accuracy of around 80% in further categorizing days into detailed subcategories including class Ia, class Ib, class II, and non-event. Notably, the machine learning classifiers reliably predict all event Ia days where particle growth and formation rates are confidently measurable. Moreover, a comparative analysis is conducted between feature-engineering machine learning methods and image-based deep learning in terms of time efficiency and overall performance. The conclusion drawn is that through reasonable feature engineering, machine learning methods can match or even surpass deep learning approaches, particularly in scenarios where time efficiency is paramount. The results of this study strongly support further investigation into this area to improve our knowledge and proficiency in automating New Particle Formation (NPF) event detection.