A research team from the Hefei Institutes of Physical Science of the Chinese Academy of Sciences developed a new deep learning method that increases the accuracy with which mixed microplastics are classified using infrared spectroscopy to 98%.

Their findings were published in Microchemical Journal.

Microplastics are plastic fragments smaller than five millimeters that come in various shapes. They are one of the four major emerging pollutants receiving global attention. Due to their small size, microplastics are more harmful than larger plastics. In practice, they often appear in mixtures, and changes in the mixing ratios alter spectral signals, making microplastics difficult to analyze. Traditional machine learning methods capture only a limited number of spectral features, reducing the accuracy of microplastic identification.

In this study, the researchers applied a highly efficient attention mechanism-convolutional block attention mechanism (CBAM) to a two-branch convolutional neural network. The two branches then concatenate the outputs of the CBAM attention module to extract more spectral features. This optimizes the model's classification performance, achieving a classification accuracy of up to 98% and outperforming traditional algorithms.

First, the CBAM module uses a channel attention module to identify key channels. Then, it uses a spatial attention module to locate important spatial regions within each channel. Finally, it generates an attention map and multiplies it element-wise with the input feature map to refine the features.

"Visualizing convolutional neural networks through Grad-CAM more clearly shows the important features selected by the model in characterizing microplastics," said TONG Jingjing, a member of the team. 

This study was funded by the National Key Research and Development Project of China, the National Natural Science Foundation of China and other projects. 

Hybrid microplastic recognition method combining attention mechanism and dual-branch convolutional neural network (Image by TONG Jingjing)