Improved neural networks for the classification of microplastics via inferior quality Raman spectra

Machine learning algorithms are proficient in the rapid extraction of features for the classification of microplastic Raman spectra. Nevertheless, the classification of Raman spectra from microplastics, particularly in the presence of complex environmental interference, remains a substantial challenge. In this study, an improved ResNet model incorporating the Squeeze-and-Excitation (SE) module is employed for the classification and identification of Raman spectra of microplastics across varying quality levels under diverse experimental conditions with insufficient laser power and short spectrum acquisition time. The improved ResNet model exhibits superior accuracy in classifying inferior quality Raman spectra characterized by significant noise and low signal-to-noise ratios, as compared to traditional CNN, without a considerable escalation in parameter size or computational burden. Even under the most adverse experimental conditions assessed, the model achieved a notable recognition accuracy of 97.83 %. Moreover, the application of Grad-CAM visualization provides insights into the criteria underlying machine learning-based spectral classification. This research underscores the capacity of machine learning algorithms in the analysis and interpretation of inferior quality Raman spectra within complex and non-ideal experimental scenarios.