Direct identification of microplastics in marine sediments by hyperspectral imaging and machine learning

ABSTRACT This study presents an innovative approach for the direct identification of microplastics (MPs) in marine sediments. Hyperspectral imaging (HSI) in the short-wave infrared range (SWIR: 1000-2500 nm) with machine learning techniques was applied to marine sediments from the Mar Piccolo basin (Taranto, Italy). Samples were collected from eight different sites across both bays of Mar Piccolo using a grab sampler and then sieved. Nine granulometric classes (from -4 mm to +180 µm) were analyzed by HSI using two instrumental setups based on particle size. Reference polymer particles were acquired and used to train classification models. Preprocessing algorithms were applied to enhance spectral differences between material classes. Principal Component Analysis (PCA) was employed to explore spectral variability and reduce data dimensionality. Two supervised classification models were developed: Hierarchical Partial Least Squares-Discriminant Analysis (Hi-PLS-DA) and Error Correcting Output Codes-Support Vector Machine (ECOC-SVM). Both models successfully identified MPs of different polymers within sediment samples. Classification results were validated using Attenuated Total Reflectance Fourier Transform Infrared (ATR-FTIR) spectroscopy. MP concentrations ranged from 58 to 17930 MPs/kg, with higher levels in finer sediment fractions and hotspot identified in MP_01 site, collected in proximity to the river mouth. Polypropylene (PP) was the most abundant polymer, followed by polyethylene (PE), polystyrene (PS), polyethylene terephthalate (PET) and polyvinyl chloride (PVC). PET was detected exclusively in the finest size classes, suggesting advanced fragmentation. The HSI-based strategy demonstrates high potential as a rapid tool for MP detection in complex environmental matrices, significantly reducing sample preparation and analysis time.