Fast Mueller matrix digital holography with anisotropy-informed machine learning for automated non-destructive microplastic characterization

Non-invasive detection and classification of microplastics (MPs) remain a significant challenge in environmental monitoring. In this work, we introduce a polarization-sensitive digital holography (PSDH) technique, termed Mueller matrix Imaging with Machine learning for Identification and Classification of MicroPlastics (MIMIC-MP) , for automated and label-free MP analysis. MIMIC-MP combines 3 × 3 Mueller matrix digital holography (M 2 DH) with anisotropy-informed machine learning for MP detection and classification. The proposed imaging system integrates Mach–Zehnder interferometry (MZI) and Mueller matrix polarimetry (M 2 P), enabling simultaneous measurement of five anisotropy-driven polarization features (ADPF), i.e., diattenuation (D), retardance (R), linear depolarization (Δ), degree of linear polarization (DoLP), angle of linear polarization (AoLP), along with the sample’s crucial quantitative phase (QP). We show that, while individual ADPFs do not provide sufficient discrimination for reliable MP classification, their combined representation yields a robust polarization fingerprint across different MP polymers. An anisotropy-informed k-nearest neighbours (AI-kNN) classifier, tailored using the five polarization parameters and coupled with multi-tier outlier detection, is developed to exploit this multi-feature representation. The proposed framework demonstrates the ability to both distinguish MP from non-MP samples and accurately classify MPs into their respective types. Experimental validation achieves a cross-validation accuracy of 96.4% and strong robustness to false acceptance of non-MP samples. The proposed MIMIC-MP system provides a scalable and effective platform for MP detection and classification, with strong potential for integration into automated and field-deployable microplastic monitoring systems.