Identifying Multicomponent Microplastics in Complex Matrices Using a Fast Fourier Convolutional Neural Network with Hierarchical Feature Mapping

Identifying environmental microplastics (EMPs) via Raman spectroscopy (RS) is well-advanced, whereas differentiating multicomponent MPs in impure mixtures remains challenging. In this work, we developed a Fast- Fourier-Convolutional neural network (FFCNN) to identify MP components in mixtures based on laser-confocal RS without complex isolation and introduced a hierarchical feature mapping (HFM) method to visualize and interpret the learned multilayer spectral features. The involved mixtures included several common types of MPs (polypropylene, polyethylene, polystyrene, polyvinyl chloride, and polyethylene terephthalate) and minor additives or impurities. The results indicated that the macro F1-score of the FFCNN was 93.6%, 10.18% higher than the second-place Random Forest, in the database with 3600 spectra from commercial MPs, mixed MPs, and EMPs. FFCNN achieved probability classification of MP mixtures via multilabel recognition. The HFM method visualized and tracked multilayer spectral feature evolution, revealing that FFC learned precise component signals through nonlocal receptive field and cross-scale fusion, eliminating noise to extract and integrate MPs-related characteristics in spectra. Of 22 environmental samples yielding one Raman spectrum each, MP component identification was 100% accurate, aligning with the pyrolysis-GC-MS cross-validation results. The work provides a practical framework for the rapid detection and identification of typical MP mixtures in the environment.