Machine learning based workflow for (micro)plastic spectral reconstruction and classification

With the advancement of artificial intelligence, it is foreseeable that computer-assisted identification of microplastics (MPs) will become increasingly widespread. Therefore, exploring a machine learning-based workflow to facilitate the identification of MPs is both meaningful and practically significant. However, interferences present in MPs spectra often compromise identification accuracy, making the improvement of spectral quality a critical prerequisite for precise identification. This study developed a fully machine learning-based workflow that combines spectral reconstruction and identification of MPs. To enhance the quality of MPs spectra, two reconstruction models named autoencoders (AE) and V-like convolutional neural networks (VCNN) were employed. Then, four classification models including decision tree, random forest, linear support vector machines (LSVM) and 1D convolutional neural networks were developed to accurately identify MPs. In terms of reconstruction, VCNN outperformed AE with a higher R² value of 0.965, while both models outperformed conventional widely used Savitzky-Golay algorithm. For classification, LSVM exhibited the best performance with an overall accuracy of 91.35% on the original dataset and 98.00% on the VCNN-reconstructed dataset. When applied to real environmental datasets, a slight decrease in performance was observed, but a maximum top-1 accuracy of 71.43% and top-3 accuracy of >90% was still practically significant, indicating that the combined workflow has great potential for spectral reconstruction and identification of MPs.