Machine Learning-Enhanced Raman Spectroscopy for Microfiber Detection: From Model Development to Coastal Investigation.

Microfibers are recognized as the most prevalent form of microplastics, with a widespread distribution across various ecosystems. The distinctive morphology of microfibers provides additional insights into their detection. In contrast to most studies targeting microplastic identification, this study proposed a rapid detection method for microfibers utilizing Raman spectroscopy and machine learning. A Raman spectral data set comprising 15 types of plastic fibers was created, and an autoencoder (AE) model was employed to process the data, effectively reducing spectral interferences such as noise and fluorescence through spectral reconstruction. Four machine learning models, including support vector machine (SVM), random forest (RF), and two convolutional neural networks (CNN, one-dimensional CNN1D, and two-dimensional CNN2D), were developed and evaluated by using the reconstructed spectra. CNN models demonstrated superior performance, with CNN1D achieving the highest accuracy of 99.03%. When applied to microfiber samples from real-world environments, the CNN1D-based method achieved an overall classification accuracy of 85.71%, achieving perfect identification of polyethylene (PE) and polyethylene terephthalate (PET). Finally, AE and CNN1D were applied to surface water samples from the East China Sea to validate the proposed microfiber identification approach in real-world scenarios. A total of 775 environmental microfibers were efficiently classified, achieving high-throughput classification with an average abundance of 2.92 ± 2.30 items/L, and the primary material types included cotton and polyester. Through the method in this study, the time needed to complete spectral processing and matching for all microfiber samples was shortened to less than 5 min. This study established a rapid detection framework for microfibers, addressing a critical gap in environmental monitoring. Importantly, the integration of morphology in our analytical pipeline shows potential for the future development of source tracking systems, which could significantly contribute to microfiber pollution control strategies.