Enhancing Microplastic Raman Spectral Quality and Classification via a Dual Encoder Residual Network

When addressing noise interference and baseline drift during the rapid Raman spectral analysis of microplastics, traditional denoising and baseline correction algorithms frequently exhibit limitations such as parameter sensitivity, reliance on manual intervention, and inadequate capacity for complex signal processing. In this work, a Dual Encoder ResUNet combined with attention mechanisms is proposed. Parallel encoders with small and large convolution kernels are adopted to achieve the extraction of multiscale spectral features. Furthermore, the internal decision-making mechanism of the model is investigated using Grad-CAM visualization technology, and the rationality of the dual-branch architecture is supported. Evaluated on a data set comprising 25,000 simulated training spectra alongside an independent testing set, the proposed model achieved an over 24% improvement in key metrics, including SNR, PSNR, and SSIM, compared to traditional Savitzky-Golay (SG) filtering and Wavelet Threshold Denoising (WTD). When applied to 19,200 Raman spectra of varying quality from eight types of microplastics collected under nonideal experimental conditions characterized by insufficient laser power and short acquisition times, the method enhanced the SNR by 3-15 times relative to traditional approaches. In downstream classification tasks, a classification accuracy of 82.80% is achieved by the Raman spectra processed by this model under the most challenging testing conditions, which is a substantial improvement over the 66.83% achieved by traditional methods and significantly higher than the 41.54% obtained with unprocessed data. These results demonstrate the advantages of the proposed neural network-based Raman spectral processing method in improving spectral quality and enhancing postprocessing performance.