Low-cost, multispectral machine learning classification of simulated airborne micro/nanoplastics

Abstract

This study presents a novel smartphone-based, machine-learning-assisted multispectral classification method for identifying airborne micro- and nanoplastics (MNPs). Instead of commercial polymeric microspheres, coffee grinder-based cryogrinding generated nonuniform MNPs from real-world plastic products with highly irregular shapes and heterogenous size distributions. The low-cost handheld device comprises a smartphone, a spectral mask array made from plastic color films, and a discrete multiplexed illumination device. A stack of images was captured across multiple wavelength ranges, and the RGB ratios were extracted without using morphological information. An XGBoost model was trained on two datasets: dry and wet MNP samples passively collected on a glass slide, simulating two types of airborne MNPs. The model successfully distinguished plastics from clay with 89-99% accuracy and classified six plastic types with 79-87% accuracy for dry and wet MNPs. This method offers a promising toolkit for airborne MNP monitoring.