New particle formation event detection with convolutional neural networks

Abstract

New Particle Formation (NPF) events are pivotal in climate and air quality research, impacting aerosol dynamics. This study introduces the ConvNeXt model for automated NPF event identification, comparing it to EfficientNet and Swin Transformer. ConvNeXt outperforms with 95.3% accuracy, surpassing EfficientNet (92.8%) and Swin Transformer (94.9%). Furthermore, we performed tests using different ConvNeXt variants (ConvNeXt-T/S/B/L/XL) and different pre-training weights, revealing that different configurations of ConvNeXt models exhibited improved NPF event recognition capabilities. Notably, the ConvNeXt-XL model achieves the highest accuracy of 96.4% in generalizability experiments. This underscores ConvNeXt's recognition prowess and practical utility in real-world NPF event detection, advancing our understanding of aerosol dynamics for climate and air quality research.

Data and model 

Contains the model code and all the data used in the paper. The data in train.zip is used for model training, the data in test.zip is used for model testing, and the rest of the data is used to validate the model generalization. model.py and train.py contain the model structure code and training code, respectively.