Reducing Spectral Confusion in Microplastic Analysis: A U-Net Deep Learning Approach

Among the various analytical techniques that have been proposed with the growing significance of microplastic detection, Raman spectroscopy is a powerful technique for detecting microplastics. However, the structural similarity in Raman spectra between fatty acids and polyethylene (PE) frequently causes misclassification by HQI-based methods, particularly when analyzing environmental samples containing mixed fatty acids. Herein, a U-net-based deep learning model was employed to precisely classify PE, stearic acid (SA), oleic acid (OA), mixtures of SA and OA, sodium dodecyl sulfate (SDS), and polypropylene based on their Raman spectra. Additionally, by incorporating a binarization technique commonly utilized in material chemistry, high scalability for both qualitative and quantitative analyses is provided. Consequently, the U-net model achieved accuracy improvements over the Pearson correlation coefficient of 2.05% to 11.09% for spectra with high signal-to-noise ratio (SNR) and 21.21% to 48.97% for spectra with nonaveraged spectra. Additionally, it demonstrated at least 36.69% higher accuracy compared to metrics such as Spearman correlation coefficient, cosine similarity, and Manhattan/Euclidean distance. This deep learning-based approach significantly reduces the confusion between PE and fatty acids observed in conventional Raman spectral analyses of microplastics, thereby demonstrating its potential applicability in microplastic standardization and analysis fields.