Explainable deep-learning detection of microplastic fibers via polarization-resolved holographic microscopy

Area of Science:

• Optics
• Materials Science
• Artificial Intelligence

Background:

• Accurate identification of microplastic fibers is vital for environmental monitoring.
• Current analytical methods for microplastic fibers face significant challenges.

Purpose of the Study:

• To develop an explainable deep-learning framework for classifying microplastic and natural microfibers.
• To utilize polarization-based features from digital holographic microscopy for fiber identification.

Main Methods:

• Reconstruction of complex Jones matrix from multiplexed holograms to extract polarization eigen-parameters.
• Creation of a 72-dimensional feature vector from nine polarization characteristics for 296 fibers across six material classes.
• Development of a fully-connected deep neural network and explainable artificial intelligence (Shapley additive explanations) for classification and feature importance analysis.

Main Results:

• The deep neural network achieved 96.7% accuracy on validation data, exceeding common machine-learning classifiers.
• Explainable AI identified eigenvalue-ratio quantities as dominant predictors, revealing the physical basis for classification.
• A reduced-feature model using only eigenvalue-based characteristics maintained high accuracy (93.3%), outperforming other classifiers.

Conclusions:

• Polarization-based features serve as distinctive optical fingerprints for microfibers.
• The developed explainable deep-learning approach enables automated and reliable microplastic fiber identification.
• This framework offers a promising tool for environmental monitoring and microplastic research.