DEEP-LEARNING ASSISTED SHORT-WAVE INFRARED HYPERSPECTRAL IMAGING FOR MICROPLASTIC CLASSIFICATION

Abstract

Microplastics are small plastics with a size between a few microns and about 5mm. Due to their small size, microplastics can be ingested by living organisms including humans, which has become a global concern and a heated area of research. Various methods have been proposed to detect and characterize microplastics. In this study, we demonstrate a hyperspectral transmission imaging system operating in the short-wave range of 1100 nm–1650nm. The developed system incorporates Fourier-transform spectroscopy (FTS) to acquire a hyperspectral data cube at high spectral resolution and signal-to-noise ratio. Using the developed system, we characterize six types of microplastic powders of approximately 10 µm in size: Polyethylene Terephthalate (PET), Low Density Polyethylene (LDPE), High Density Polyethylene (HDPE), Polypropylene (PP), Polystyrene (PS) and Polyvinyl Chloride (PVC). Although the scattering of light at sample boundaries renders conventional chemometric analysis to seriously underperform, using deep learning, we show the six types of microplastics can be classified with an accuracy of 99.1%. Further, we show that the sampling number of FTS can be reduced from 1000 to 100 without much degradation in the classification accuracy. The developed instrument provides an economical solution for high-throughput classification of small microplastics between 10 and 100 µm.