Microplastic detection in soil by THz Time-Domain hyperspectral imaging combined with unsupervised learning analysis

Year: 2026

Authors: Martinez A., Di Sarno V., Maddaloni P., Castaldo R., Cerruti P., Cocca M., Paturzo M., Paparo D.

Autors Affiliation: Univ Napoli Federico II, Scuola Super Meridionale, Naples, Italy; CNR, Ist Nazl Ott INO, Pozzuoli, Italy; CNR, Inst Polymers Composites & Biomat, Pozzuoli, Italy; CNR, Inst Appl Sci & Intelligent Syst, ISASI, Pozzuoli, Italy.

Abstract: This work introduces a novel approach for detecting and spatially mapping microplastics in soil using Terahertz (THz) Time-Domain Hyperspectral Imaging. To this end, we applied a range of statistical and machine-learning techniques. First, we conducted a statistical analysis that enabled the identification of spectral and temporal signatures of the THz signal associated with low-density polyethylene contamination. These features informed a similarity analysis, which generated heatmaps exhibiting clear spatial patterns characterized by an inverse relationship between similarity to soil and to plastic. Although effective, this approach relies on prior knowledge of pure reference spectra, thereby limiting its applicability to uncharacterized soil environments. To address this limitation, we explored unsupervised learning methods-Principal Component Analysis, K-means clustering, Agglomerative Clustering, Spectral Clustering and Gaussian Mixture Models-that leverage internal variance to segment the data without reference spectra. All five methods consistently separated the mixtures into distinct phases associated with soil and plastic. Averaged spectra within these regions recovered the characteristic features of each material. The convergence of independent unsupervised methods to similar spatial and spectral distinctions highlights the robustness of the approach. Note that unsupervised approaches offer a key advantage when labeled samples are unavailable for training supervised algorithms. The key contribution of this work is the demonstration that, under controlled-albeit simplified-experimental conditions, a robust separation between plastic and soil consistently emerges across multiple, conceptually independent analytical methods, indicating that the observed discrimination is rooted in underlying physical contrasts rather than being specific to any single algorithm. This controlled regime allows the fundamental physical behavior of the plastic-soil system to be characterized before addressing more challenging scenarios.

Journal/Review: ENVIRONMENTAL TECHNOLOGY & INNOVATION

Volume: 42      Pages from: 104889-1  to: 104889-12

More Information: The Terahertz ASOPS system was acquired through SHINE project funding (Strengthening the Italian Nodes of E-RIHS) .
KeyWords: THz-TDS; Machine learning; THz imaging; Principal components analysis; K -Means clustering; Spectral similarity analysis; Agglomerative clustering; Spectral clustering; Soil contamination; Microplastic
DOI: 10.1016/j.eti.2026.104889