Application of AI methods for classification of individual fine particles based on morphology and elemental composition

Abstract

Atmospheric fine particles have become a significant concern due to their impacts on global climate and adverse health impacts. These impacts are largely dependent on their physical and chemical properties including particle morphology, particle size and particle composition which undergoes various changes during their atmospheric lifetimes. Characterizing these properties is critical. However, conventional methods such as bulk particle analysis often fail to provide detailed insights on physio-chemical properties of individual particles. To overcome this limitation, individual particle analysis techniques such as electron microscopy are employed. However, the traditional manual approach to analyze the electron microscopic data are time- consuming, labor intensive, subjective and prone to human error. This has led to a growing interest in artificial intelligence (AI) to automate the classification of fine particles. This dissertation is mainly focused on exploring different machine learning and deep learning methods for rapid, accurate, scalable and integrated morphological and elemental classification of fine particles from various sources and ambient atmosphere obtained from transmission electron microscopy-energy dispersive X-ray spectroscopy (TEM-EDX). In this study, TEM images of fine particles were collected from diverse ambient atmosphere and sources. Different pre-trained convolutional neural network (CNN) models were fine-tuned using three different strategies (Freeze the convolutional base, Train few layers and Train the entire model) and compared to accurately classify the individual fine particles into 4 different morphological classes (Agglomerate, Dendrite, Polygonal and Spherical) (Chapter III). Weighted average ensemble model (DenseNet169 and InceptionV3) was found to be the best model when the entire model was trained achieving 99.0% overall classification accuracy with high sensitivity (0.99) and specificity (1.00). However, this high performance was achieved under controlled conditions with a closed-set classification problem where the model was only trained to classify four specific morphological classes. But in the atmosphere, particles exist in diverse shapes. When encountered with these out-of-distribution (OOD) particles, supervised models forces the classification into one of the known classes, leading to misclassification and reduced reliability in real-world applications. To address this limitation, this study explored different OOD methods (Maximum SoftMax probability, Out-of- Distribution detector for Neural networks, Mahalanobis distance, Grad Norm and Energy-based) so that the best-performing ensemble model can handle both known and unknown particle morphology (Chapter IV). Mahalanobis distance achieved the best performance (overall test accuracy, 78%) for separating known and unknown particles but with the tradeoff of improved unknown particle identification but reduced accuracy for known classes. Previous studies predominantly focused on elemental composition for classification with morphological features used primarily for visual validation and particle size are reported rather than quantitative integration, thus underutilizing rich individual particle information obtained from TEM-EDX for particle type classification. Automated studies have also focused primarily on elemental composition and were limited to site-specific sources with manual post-processing and restricted generalizability. So, in this study, the multimodal data (morphology, elemental composition and particles size) of individual fine particles were fused together using different multimodal fusion methods (Early fusion, traditional joint fusion and attention- based fusion) to classify the individual fine particles into seven different particle types (Dust, Fly ash, Metals, Organic-rich, Sea spray, Soot and Sulphur-rich) (Chapter V). The model with attention-based fusion performed the best with 89.6% overall classification accuracy with high sensitivity (0.90) and specificity (0.98) compared to simple fusion methods which lacked interaction among the modalities. This dissertation contributes to addressing the bottleneck of the conventional individual particle analysis by studying the potential application of different artificial intelligence methods to develop automated, rapid, accurate, scalable and integrated methods for individual particle analysis from diverse sources. That could help in better understanding of atmospheric particle sources, their transformation processes and impact on health and climate.