Deep Learning-Based Shape Classification for Hyperspectral-Imaged Microplastics.

Area of Science:

• Environmental Science
• Analytical Chemistry
• Computer Science

Background:

• Microplastic (MP) shape classification is crucial but traditionally labor-intensive and prone to bias.
• Automating MP shape analysis is needed for efficient environmental monitoring.

Purpose of the Study:

• To investigate deep learning models for automated microplastic shape classification using hyperspectral imaging.
• To compare the performance of various deep learning architectures on diverse datasets.

Main Methods:

• Tested nine deep learning architectures (NNs, CNNs, transfer learning models) on 11,042 MP hyperspectral images.
• Utilized micro-Fourier transform infrared spectroscopy across seven environmental matrices.
• Evaluated models on original, augmented, refined, and augmented refined datasets.

Main Results:

• Convolutional Neural Networks (CNNs) outperformed Neural Networks (NNs), and transfer learning models surpassed non-transfer learning models.
• MobileNet achieved the highest accuracy (0.93 validation, 1.00 test).
• Model architecture and data quality significantly impact classification accuracy; complex models showed robustness.

Conclusions:

• Deep learning provides an efficient, automated alternative to manual MP shape classification.
• Model selection and high-quality data are critical for optimal performance.
• Further research into robust model designs is needed for advanced MP analysis.