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

Giulia Dalla Fontana; Jan Appel; Jarom\'ir B\v{e}hal; Lisa Miccio; Marika Valentino; Miroslav Je\v{z}ek; Pietro Ferraro; Raffaella Mossotti; Vittorio Bianco

arxiv: 2601.15769 · v3 · pith:B74HQEQCnew · submitted 2026-01-22 · ⚛️ physics.optics · physics.data-an

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

Jan Appel , Marika Valentino , Lisa Miccio , Vittorio Bianco , Raffaella Mossotti , Giulia Dalla Fontana , Miroslav Je\v{z}ek , Pietro Ferraro

show 1 more author

Jarom\'ir B\v{e}hal

This is my paper

Pith reviewed 2026-05-16 12:22 UTC · model grok-4.3

classification ⚛️ physics.optics physics.data-an

keywords microplastic fiberspolarization-resolved holographyJones matrixdeep neural networkexplainable AIoptical anisotropyfiber classificationenvironmental monitoring

0 comments

The pith

Polarization eigen-parameters from holographic microscopy let a neural network classify microplastic fibers at 96.7 percent accuracy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that reconstructing the complex Jones matrix from polarization-resolved digital holographic microscopy yields eigen-parameters that capture optical anisotropy in fibers. Statistical summaries of these parameters feed a fully connected deep neural network that separates six material classes, including several common microplastics and natural fibers such as cotton and wool. The network reaches 96.7 percent validation accuracy, higher than standard machine-learning classifiers, while Shapley analysis reveals that eigenvalue-ratio quantities carry most of the discriminative power. A simplified version of the same network that uses only those dominant ratios still achieves 93.3 percent accuracy, confirming their central role.

Core claim

From multiplexed holograms the complex Jones matrix of each fiber is reconstructed to extract polarization eigen-parameters describing optical anisotropy. Statistical descriptors of nine polarization characteristics form a 72-dimensional feature vector for 296 fibers spanning six material classes. The fully connected deep neural network achieves 96.7 percent accuracy on validation data, surpassing common classifiers, with Shapley additive explanations identifying eigenvalue-ratio quantities as dominant predictors. A reduced-feature model using only these characteristics retains 93.3 percent accuracy.

What carries the argument

Reconstruction of the complex Jones matrix from polarization-resolved holograms to obtain eigenvalue-based descriptors of optical anisotropy that serve as material-specific fingerprints.

If this is right

Automated identification of microplastic fibers becomes practical for routine environmental monitoring without manual inspection.
Eigenvalue-ratio quantities alone support near-full accuracy, allowing simpler and faster models in future systems.
The physical distinction between synthetic and natural fibers rests on measurable differences in optical anisotropy rather than on chemical composition alone.
Explainable AI methods can be applied to polarization data to make the classification decisions transparent to regulators and analysts.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same polarization fingerprints could be tested on mixed debris from real rivers or oceans to see whether accuracy holds outside controlled laboratory fibers.
Portable holographic microscopes equipped with polarization channels might enable field deployment for rapid on-site screening of water samples.
The approach suggests that polarization-resolved imaging could be extended to other anisotropic micro-particles such as certain pigments or biological fragments.

Load-bearing premise

The 296 laboratory-prepared fibers and the polarization properties extracted from them represent the diversity, degradation states, sizes, orientations, and surface contamination found in real environmental samples.

What would settle it

Collect fibers directly from environmental water or sediment samples, run them through the trained model, and check whether accuracy remains above 90 percent.

Figures

Figures reproduced from arXiv: 2601.15769 by Giulia Dalla Fontana, Jan Appel, Jarom\'ir B\v{e}hal, Lisa Miccio, Marika Valentino, Miroslav Je\v{z}ek, Pietro Ferraro, Raffaella Mossotti, Vittorio Bianco.

**Figure 2.** Figure 2: The implemented neural network with input layer (green) with dimensions [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Confusion matrix of the model trained on labeled data. (a) Evaluation on the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: SHAP-FI values of nine polarization characteristics and their eight corresponding [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Class-wise SHAP-FI values of nine polarization characteristics. The most [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗

read the original abstract

Reliable identification of microplastic fibers is crucial for environmental monitoring but remains analytically challenging. We report an explainable deep-learning framework for classifying microplastic and natural microfibers using polarization-resolved digital holographic microscopy. From multiplexed holograms, the complex Jones matrix of each fiber was reconstructed to extract polarization eigen-parameters describing optical anisotropy. Statistical descriptors of nine polarization characteristics formed a 72-dimensional feature vector for a total of 296 fibers spanning six material classes, including polyamide 6, polyethylene terephthalate, polyamide 6.6, polypropylene, cotton and wool. The designed fully connected deep neural network achieved an accuracy of 96.7 % on the validation data, surpassing that of common machine-learning classifiers. Explainable artificial intelligence analysis with Shapley additive explanations identified eigenvalue-ratio quantities as dominant predictors, revealing the physical basis for classification. An additional reduced-feature model with the preserved architecture exploiting only these most significant eigenvalue-based characteristics retained high accuracy (93.3 %), thereby confirming their dominant role while still outperforming common machine-learning classifiers. These results establish polarization-based features as distinctive optical fingerprints and demonstrate the first explainable deep-learning approach for automated microplastic fiber identification.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a working polarization holographic + explainable NN pipeline that hits 96.7% on 296 clean lab fibers, with SHAP linking performance to eigenvalue ratios, but the numbers only cover controlled samples.

read the letter

The main thing to know is that the authors built a coherent pipeline: they reconstruct the complex Jones matrix from multiplexed holograms, pull 72 statistical descriptors from nine polarization parameters, train a fully connected network to 96.7% validation accuracy on 296 lab fibers across six classes, and use SHAP to show that eigenvalue ratios dominate the decisions. A reduced model keeping only those top features still reaches 93.3% and beats standard classifiers. That combination of polarization eigen-parameters with post-hoc explainability is new for this task and gives a physical story for why the classifier works on their data. The technical steps look solid for controlled conditions, and the reduced-feature result is a nice check that the performance is not just memorizing noise. The dataset is small but the numbers are reported clearly enough to see the gap over baselines. The soft spot is exactly what the stress-test note flags: everything is lab-prepared fibers. No field-collected or weathered samples appear, so the validation accuracy measures how well the 72-dimensional descriptors separate clean materials under fixed conditions, not how they hold up with surface contamination, degradation-induced birefringence shifts, or variable diameters and overlaps that real environmental monitoring would face. The abstract also omits the train-validation split details and any cross-validation or error bars, which makes the 96.7% figure harder to judge for robustness. This is for readers working on optical sensors or microplastic detection methods who need a starting point with some interpretability. It is not a major theoretical advance, but the method is distinct from prior work and the evidence is internally consistent, so it deserves peer review in an applied optics or environmental monitoring journal. A referee can ask for external testing, but the core contribution is worth the time.

Referee Report

3 major / 2 minor

Summary. The manuscript presents an explainable deep-learning framework for classifying microplastic and natural microfibers using polarization-resolved digital holographic microscopy. Jones-matrix reconstruction yields 72-dimensional feature vectors of statistical descriptors from 296 laboratory-prepared fibers across six classes (polyamide 6, PET, polyamide 6.6, polypropylene, cotton, wool). A fully connected DNN reaches 96.7% validation accuracy, outperforming standard ML classifiers; SHAP analysis identifies eigenvalue-ratio quantities as dominant predictors. A reduced model retaining only these features achieves 93.3% accuracy while preserving the architecture.

Significance. If the performance generalizes, the work would advance automated, physically interpretable microplastic identification by linking polarization eigen-parameters directly to classification via SHAP and by demonstrating that a compact set of eigenvalue descriptors suffices. The combination of holographic Jones calculus with an explainable DNN is a constructive step for environmental optics. The reduced-feature result is a clear strength, as it both confirms the physical basis and reduces model complexity.

major comments (3)

[Abstract] Abstract: the headline accuracies (96.7 % DNN, 93.3 % reduced model) are stated without any information on data partitioning, cross-validation folds, class balance, or error bars. Because the central performance claim rests on these numbers, the absence of these details prevents assessment of robustness against post-hoc selection.
[Abstract] Abstract / Results: the entire evaluation uses 296 laboratory-prepared fibers. No external test set drawn from environmental samples (with weathering, surface fouling, variable diameters, or overlapping fibers) is reported. This is load-bearing for the claim of applicability to field monitoring, as the validation accuracy only measures performance on the same controlled distribution used for training and feature selection.
[Results] Results (reduced-feature model): while the 93.3 % accuracy on the eigenvalue-based subset is encouraging, the manuscript does not specify which exact eigenvalue ratios were retained or provide a direct comparison of confusion matrices between the full 72-dimensional and reduced models. This information is needed to evaluate whether the physical interpretability claim is fully supported.

minor comments (2)

[Abstract] Abstract: the phrasing 'microplastic fibers' versus 'natural microfibers' should be made consistent throughout to avoid reader confusion about the six-class taxonomy.
[Methods] The manuscript would benefit from an explicit statement of the number of fibers per class and any class-imbalance mitigation strategy, even if only in supplementary material.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed comments, which have helped improve the clarity and robustness of our manuscript. We address each major comment below and have made revisions where feasible to strengthen the presentation of our results.

read point-by-point responses

Referee: [Abstract] Abstract: the headline accuracies (96.7 % DNN, 93.3 % reduced model) are stated without any information on data partitioning, cross-validation folds, class balance, or error bars.

Authors: We agree that these methodological details are necessary to evaluate the reliability of the reported accuracies. In the revised manuscript, we have expanded the abstract and added a dedicated paragraph in the Methods section specifying that the 296-fiber dataset was partitioned using stratified 70/30 train/validation splits to preserve class balance (approximately 49 fibers per class). We performed 5-fold cross-validation, with the reported accuracies representing mean values and standard deviations as error bars (96.7 ± 1.1% for the full model and 93.3 ± 1.4% for the reduced model). revision: yes
Referee: [Abstract] Abstract / Results: the entire evaluation uses 296 laboratory-prepared fibers. No external test set drawn from environmental samples (with weathering, surface fouling, variable diameters, or overlapping fibers) is reported.

Authors: We acknowledge this as a limitation of the current study. The work was designed to establish the physical basis and explainability of polarization features under controlled laboratory conditions, which enabled precise Jones-matrix reconstruction and SHAP analysis. We have revised the Discussion to explicitly state that the reported performance applies to lab-prepared samples and to outline plans for future validation on environmental samples; however, we do not possess such data in the present dataset. revision: no
Referee: [Results] Results (reduced-feature model): while the 93.3 % accuracy on the eigenvalue-based subset is encouraging, the manuscript does not specify which exact eigenvalue ratios were retained or provide a direct comparison of confusion matrices between the full 72-dimensional and reduced models.

Authors: We thank the referee for this observation. In the revised Results section, we now explicitly list the top three eigenvalue-ratio features retained by the SHAP analysis (λ_max/λ_min for linear retardance, circular birefringence, and diattenuation). We have added a supplementary figure showing side-by-side confusion matrices for the full 72-dimensional and reduced models, confirming that the reduced model maintains comparable per-class performance with only minor increases in confusion between polyamide variants. revision: yes

standing simulated objections not resolved

Absence of an external test set from real environmental microplastic samples, as the study is restricted to controlled laboratory-prepared fibers and acquiring new environmental data would require additional experiments beyond the scope of this work.

Circularity Check

0 steps flagged

No circularity: standard feature extraction and empirical training on held-out lab data

full rationale

The paper reconstructs Jones matrices from holographic data via established polarization optics, computes 72-dimensional statistical descriptors of eigen-parameters, and trains a fully connected DNN on a split of the 296-fiber experimental dataset to report validation accuracy. SHAP explanations are applied after training to rank features, and the reduced model retrains on the top-ranked subset; neither step defines its output by construction from the labels or reduces to a self-citation. All performance numbers are direct empirical measurements on the controlled lab distribution, with no parameter fitted to the target accuracy itself and no uniqueness theorem or ansatz imported from prior author work.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on established optical physics and standard machine-learning practice rather than new physical postulates; the only free parameters are the neural-network weights learned from the 296-fiber dataset.

free parameters (1)

Neural-network weights and hyperparameters
The fully connected network parameters are optimized on the training portion of the 296-fiber dataset to achieve the reported classification accuracy.

axioms (2)

standard math The complex Jones matrix fully captures the polarization transformation induced by a thin fiber
Invoked when reconstructing the matrix from multiplexed holograms.
domain assumption Statistical descriptors of the polarization eigen-parameters form a discriminative feature space for material classes
Used to construct the 72-dimensional input vector.

pith-pipeline@v0.9.0 · 5542 in / 1379 out tokens · 40201 ms · 2026-05-16T12:22:24.779302+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

From multiplexed holograms, the complex Jones matrix of each fiber was reconstructed to extract polarization eigen-parameters... eigenvalue-ratio quantities as dominant predictors

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.