Issues with the detection and classification of microplastics in marine sediments with chemical imaging and machine learning

Abstract

Numerous studies have attempted to detect microplastic litter directly in environmental sediments via spectral imaging and powerful classification algorithms. Spectral imaging is attractive largely due to the benefits of adding a spatial element to spectral data, the relative measuring speed, and minimal sample processing. Despite this promise, important concerns related to the spatial and spectral selectivity must be considered along with the appropriateness of classification algorithms. Here we evaluate the performance of near infrared hyperspectral imaging (NIR-HSI) and four commonly used classification algorithms on a simple test case in which images of individual microplastics of known size on top of sand were collected. The results highlight major weak points of NIR-HSI and machine learning as applied to the detection of the microplastics, with a large proportion of false positives and negatives in most of the situations studied, and alerts the reader to important concerns about the use of this methodology.