Identification of New Particle Formation Events Using a You Only Look Once (YOLO) Deep Learning Algorithm

Résumé fourni par la source

Atmospheric new particle formation (NPF) events, which involve the formation and growth of molecular clusters, affect air quality, weather, climate, and human health. Traditional NPF event classification schemes in the literature, primarily through visual inspection of particle number size distributions, are subjective, time-consuming, and laborious. Here, we introduced a fast object detection deep learning algorithm, You Only Look Once (YOLO), for the first time to detect NPF events. We used more than one year of asynchronous particle number size distribution data from 20 diverse geographical locations globally. The YOLO algorithm was first trained and validated using a small subset of manually annotated 252 NPF events and 195 non-events (approximately 25 images from each measurement site). The trained YOLO algorithm was then evaluated to detect NPF events against the remaining 6462 observation days across all measurement sites. The performance metrics of the trained YOLO algorithm revealed high precision and accuracy in detecting NPF events. The model accuracy for each measurement site was calculated by taking the ratio of the total number of NPF events detected by the trained YOLO algorithm at a confidence score (CoS) > 0.1 to the total number of visually identified NPF events. Considering all measurement sites, the trained YOLO algorithm’s accuracy (in fraction) for detecting NPF events ranged from 0.74 to 0.97 at a CoS > 0.1, although the accuracy decreased with increasing threshold CoS. This work underscores the efficacy and robustness of the YOLO algorithm and demonstrates its applicability in accurately detecting NPF events in diverse environmental conditions worldwide.