Problem

The accurate identification of microplastic fibers (MFFs) is a critical yet challenging aspect of environmental monitoring. Microplastics, particularly MFFs from textiles, are pervasive pollutants in various ecosystems, posing potential environmental and health risks due to their persistence and ability to carry chemical additives. Existing methods for microplastic identification, such as visual inspection, spectroscopy, and thermal analysis, often suffer from limitations like operator dependence, destructiveness, high cost, complex sample preparation, or susceptibility to interference, highlighting the need for a more reliable, non-destructive, and automated classification approach.

Method

The researchers developed an explainable deep-learning framework utilizing polarization-resolved digital holographic microscopy for MFF classification. The method begins by using a Mach-Zehnder interferometer-based digital holographic microscope to capture multiplexed holograms. From these, the complex Jones matrix of each fiber is reconstructed, allowing for the extraction of nine polarization eigen-parameters describing optical anisotropy. These include the magnitude and phase of the Eigen-Numbers (ENs) ratio, their real and imaginary parts, the inner product of Eigen-Vectors (EVs), and the ellipticity and orientation angles of the Eigen-Polarizations (EPs).

For each of these nine polarization characteristics, eight statistical descriptors (mean, median, mode, mean absolute deviation, median absolute deviation, standard deviation, skewness, and kurtosis) were computed across the fiber pixels, creating a 72-dimensional feature vector for each MFF. A dataset of 296 fibers across six material classes (polyamide 6, polyethylene terephthalate, polyamide 6.6, polypropylene, cotton, and wool) was used. A fully connected deep neural network, with an input layer of 72 elements, four hidden layers (256, 128, 64, 32 neurons) with ReLU activation, and a Softmax output layer for six classes, was trained using the Adam optimizer and categorical cross-entropy loss. Regularization techniques like batch normalization and dropout were employed to prevent overfitting. Explainable AI was integrated using Shapley additive explanations (SHAP) to identify the most influential features for classification.

Results

The developed deep neural network achieved an impressive classification accuracy of 96.7% on the validation dataset and 98.6% on the complete dataset, significantly outperforming common machine learning classifiers such as Gaussian naive Bayes, random forest, support vector machine, K-nearest neighbors, gradient boosting, and logistic regression (which ranged from 82% to 92% accuracy). SHAP analysis revealed that eigenvalue-ratio quantities, particularly the absolute value of the ENs ratio and its statistical parameters (median absolute deviation and median), were the most dominant predictors for overall classification. Class-wise analysis showed that EN-based features were highly significant for synthetic fibers, while EP ellipticity (χEP2) was most important for cotton due to its twisted ribbon-like geometry, and the imaginary part of the ENs ratio for wool. A reduced-feature model, utilizing only the most significant EN-based characteristics, maintained high accuracy at 93.3%, further confirming their dominant role and still outperforming conventional methods.

Implications

This research establishes polarization-based features as distinctive optical fingerprints for microplastic fibers, demonstrating a highly accurate and interpretable deep-learning approach for automated MFF identification. The method’s non-destructive and label-free nature, combined with its superior classification performance, positions polarization-sensitive digital holography as a powerful tool for environmental monitoring and analysis of microplastic pollution. By identifying the specific optical properties that drive classification, this explainable AI framework not only improves detection capabilities but also offers deeper insights into the physical basis for distinguishing different fiber types, which can aid in developing more targeted mitigation strategies for microplastic contamination.