Automatic classification of microplastics and natural organic matter mixtures using a deep learning model

Area of Science:

• Environmental Science
• Analytical Chemistry
• Data Science

Background:

• Microplastic (MP) classification in aquatic systems often requires extensive sample preprocessing, such as oxidation to remove natural organic matter (NOM).
• These preprocessing steps are time-consuming, costly, and can introduce errors due to subjective operator judgment during spectroscopic analysis.
• Accurate and efficient MP identification is crucial for understanding their environmental impact.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) model for improved classification accuracy of microplastics (MPs) in mixtures with natural organic matter (NOM).
• To assess the applicability of a convolutional neural network (CNN) with a spatial attention mechanism for classifying MPs from Raman spectra.
• To compare the DL model's performance against conventional Raman spectral library software.

Main Methods:

• A convolutional neural network (CNN) with a spatial attention mechanism was employed to analyze Raman spectra of microplastic-natural organic matter (MP-NOM) mixtures.
• The DL model's classification results were compared with those from traditional Raman spectral library software.
• Gradient-weighted class activation mapping (Grad-CAM) was utilized to investigate critical spectral bands for DL model training and interpretation.

Main Results:

• The developed DL model achieved a high classification accuracy of 99.54%, significantly outperforming conventional Raman spectral library software (31.44%).
• Grad-CAM analysis confirmed the DL model's ability to identify MPs by focusing on prominent Raman spectral peaks.
• The model demonstrated effectiveness in classifying MPs even with less prominent spectral features, showing robustness against variations in peak intensity.

Conclusions:

• Deep learning offers a powerful, automated, and objective approach for microplastic classification in aquatic environments.
• The proposed DL model eliminates the need for laborious NOM preprocessing, saving time and resources.
• This research presents a promising direction for advancing microplastic analysis and monitoring technologies.