Automatic quantification and classification of microplastics in scanning electron micrographs via deep learning.

Area of Science:

• Environmental Science
• Materials Science
• Computer Science

Background:

• Microplastic pollution monitoring requires accurate quantification and classification.
• Manual analysis of microplastics from imaging data is labor-intensive and time-consuming.
• Scanning electron microscopy provides detailed microplastic imaging but requires efficient analysis methods.

Purpose of the Study:

• To develop and evaluate deep learning models for automated microplastic quantification and classification.
• To compare the performance of deep learning models against conventional computer vision techniques.
• To demonstrate the efficiency and accuracy of AI-driven analysis for microplastic identification.

Main Methods:

• A manually annotated dataset of 237 micrographs of microplastics was created.
• U-Net and MultiResUNet models were implemented for semantic segmentation of microplastics.
• U-Net with object-aware pixel embedding was used for instance segmentation of tangled fibers.
• A fine-tuned VGG16 neural network was employed for microplastic shape classification.

Main Results:

• Both U-Net and MultiResUNet achieved high Jaccard indices (>0.75) for microplastic quantification, outperforming traditional methods.
• The VGG16 model accurately classified microplastic shapes with 98.33% accuracy.
• Automated analysis using trained models took seconds per micrograph, drastically reducing processing time compared to manual methods.

Conclusions:

• Deep learning offers a highly accurate, efficient, and cost-effective solution for microplastic quantification and classification.
• The developed AI tools can significantly accelerate the monitoring of microplastic pollution.
• Future work can leverage growing datasets to improve microplastic identification in environmental samples.