Comparison of polynomial and machine learning regression models to predict LDPE, PET, and ABS concentrations in beach sediment based on spectral reflectance

Abstract Microplastic (MP) contamination on land has been estimated to be 32 times higher than in the oceans, however, despite most MPs potentially found in soils there is a distinct lack of research on soil MPs compared to marine MPs. Beaches are bridges between land and ocean and present equally understudied sites of microplastic pollution. Visible-near-infrared (vis-NIR) has been applied successfully for the measurement of reflectance and prediction of low-density polyethylene (LDPE), polyethylene terephthalate (PET), and polyvinyl chloride (PVC) concentrations in soil. The rapidity and precision associated with this method make vis-NIR particularly promising. The present study explores two novel data processing approaches in this field. First, using a spectroradiometer, the spectral reflectance data was measured from treated beach sediment spiked with virgin microplastic pellets (LDPE, PET, and acrylonitrile butadiene styrene (ABS)). The concentration of spiked microplastics in the treated sediment was increased sequentially. Using the recorded spectral data, polynomial regression and machine learning-based regression methods were applied and predictive models developed for each microplastic. Both methods generated models of good accuracy with R2 values greater than 0.7, root mean squared error (RMSE) values less than 3 and standard deviation (SD) < 0.06. Additionally, the optimum wavelengths of each microplastic for their detection in the beach sediment were found to be similar by both the methods in the used vis-NIR spectrum (325 nm – 1075 nm), indicating a higher vis-NIR spectrum range is not required for detection. This study is the first assessing predictive abilities of models created by polynomial regression and machine learning algorithms for soil microplastic contamination and another step towards standardizing the quantification of microplastics in soil samples using vis-NIR spectroscopy.