Improving the classification performance of microplastics by noise reduction and baseline correction of Raman spectra with a neural network-based algorithm

Microplastics have emerged as global environmental pollutants, with Raman spectroscopy serving as an effective method for detecting and identifying them. However, conventional Raman preprocessing methods are often constrained by parameter sensitivity, a heavy reliance on manual intervention, and limited efficacy in handling complex signals. To overcome these limitations, we report the application of a ResUNet model integrated with Squeeze-and-Excitation (SE) blocks for the denoising and baseline correction of Raman spectra of microplastics, acquired under nonideal conditions characterized by low laser power and short acquisition times. Compared with the traditional combination of Wavelet Threshold Denoising and AirPLS baseline correction, a more than 15-fold improvement in the signal-to-noise ratio was achieved. In downstream classification tasks, even under stringent conditions (29.63 mW laser power and 750 ms integration time), the identification accuracy for microplastics was significantly enhanced from 35.13% in the raw data to 96.90%, notably outperforming the 55.70% accuracy attained by traditional methods. This work demonstrates the effectiveness of the SE-ResUNet neural network in enhancing spectral quality and optimizing post-processing outcomes.