Recently, a research team led by Prof. GAO Xiaoming from the Hefei Institutes of Physical Science of the Chinese Academy of Sciences, improved residual neural network to accurately classify and identify microplastics using low-quality Raman spectra, even under non-ideal experimental conditions.
"It detects and classifies microplastics when the data is cluttered with noise," explained Prof. GAO, "and it does this without overloading computing power."
The research results were published in Talanta.
Microplastics, plastic particles under 5mm, are found everywhere. They pose serious environmental and health risks, making their rapid and accurate identification crucial for pollution control. Raman spectroscopy, with its non-destructive and high-resolution capabilities, is ideal for detection, while machine learning enables precise classification of complex spectra. However, analyzing microplastics in challenging environments or under interference remains difficult.
In this work, the team proposed an improved residual network model capable of classifying and identifying microplastics based on Raman spectra obtained under non-ideal experimental conditions, such as insufficient laser power and short spectral acquisition times.
Compared to traditional convolutional neural networks, the improved residual network with the Squeeze-and-Excitation module achieves higher accuracy in classifying low-quality microplastic Raman spectra with significant noise interference and low signal-to-noise ratios, without significantly increasing the number of parameters or computational load.
Additionally, Grad-CAM visualization, a kind of "AI X-ray vision," reflects the basis for spectral classification by machine learning.
This work demonstrates the capability of machine learning to analyze and process low-quality Raman spectra in more complex environments and under interference, according to the team.
The detected Raman spectra of microplastics enter into the neural network, which is trained to capture the features of different kinds of spectra for effective classification. SE block is used to enhance the performance of the network. Grad-CAM visualization reflects the basis for spectral classification by machine learning. (Image by CHEN Jiajin)
