Microplastics in the Rough: Using Data Augmentation to Identify Plastics Contaminated by Water and Plant Matter

Abstract

Microplastics are present in nearly all environments. The detection of microplastics in the field is an important step toward understanding and regulating the proliferation of plastic waste, particularly in natural environments. Real-time surveys require robust instruments, rapid acquisition, and minimal processing. Near infrared (NIR) spectroscopy is an ideal technique to detect polymer composition regardless of spectral interference by water and/or organic matter. Here we report a fiber-based NIR instrument designed for simple and efficient spectral acquisition of consumer plastic particles across a range of sizes. Data augmentation with measured interferent spectra has been used to generate machine-learning based classification models that can identify polymer compositions in plastic particles that are wet and/or mixed in with organic plant material. These models achieve 98.5% accuracy on synthetic data and 86.4% accuracy when transferred to spectra of plastic particles of nine common polymers with particle sizes as small as 500 μm. Our model paves the way for the development of equipment to perform real-time surveys of microplastic compositions in the field.