Hybrid deep learning framework for environmental microplastic classification: Integrating CNN-based spectral feature extraction and transformer models

Accurate classification of environmental microplastics (MPs) from FTIR spectra remains challenging due to spectral variability, environmental weathering, and the limitations of existing models in capturing both local and global spectral features. This study proposes a hybrid deep learning framework that integrates convolutional neural networks (CNNs) and Transformer models to classify MPs based on Fourier-transform infrared (FTIR) spectroscopy. The CNN component efficiently captures local spectral patterns, while the Transformer module models long-range dependencies and global contextual relationships across the full spectral range. This sequential hybrid architecture combines the complementary strengths of CNN and Transformer, enhancing feature extraction and improving classification robustness. We constructed a diverse spectral dataset containing 17 polymer types derived from soil, air, sediment, and water matrices. The proposed model achieved a high classification accuracy of 95.77 % on the validation set, outperforming traditional machine learning methods. The model's robustness was further confirmed through 50 independent trials, demonstrating stable and reproducible performance. These results indicate that the CNN-Transformer architecture not only improves predictive performance but also ensures generalizability across diverse environmental conditions. This integrated approach offers a reliable and scalable solution for the rapid identification of MPs across diverse environmental contexts, with potential implications for pollution monitoring and regulatory assessment.