A Novel Approach for Fast Microplastic Quantification in Sediments Using Machine Learning—Spectrometer Combinations

Abstract

The accumulation of microplastics in surface soils and sediments has raised significant concerns due to their potential environmental risks. Conventional quantitative methods for microplastics often require time-consuming pretreatment and statistical counting, rather than providing direct concentration data, complicating cross-study comparisons. To rapidly investigate microplastic pollution in environmental samples, machine learning (ML) algorithms combined with spectrometers have been employed to estimate microplastic concentrations without the need for extraction. While previous research has primarily focused on microplastic-spiked soils, this study explores the use of river and loess sediments spiked with four commonly used plastic polymers: polyethylene (PE), polypropylene (PP), polystyrene (PS), and polyvinyl chloride (PVC) at concentrations ranging from 0.1 wt% to 5 wt%. Visible near-infrared (vis–NIR, 350–2500 nm) and Fourier transform infrared (FTIR, 4000–400 cm⁻¹) spectroscopy were employed to acquire spectra, which were then preprocessed using the first derivative (FD) and Savitzky-Golay (SG) filtering (FD-SG) methods. Support Vector Regression (SVR), Partial Least Squares Regression (PLSR) and Back Propagation Neural Network (BPNN) models were trained and tested using river sediment datasets and subsequently applied to predict microplastic concentrations in loess sediment samples. The SVR models, constructed with preprocessed vis–NIR data using the FD-SG method, exhibited the best performance, with root mean square error (RMSE) for PE, PP, PS, and PVC in loess sediments of 0.32 wt%, 0.46 wt%, 0.74 wt%, and 0.59 wt%, respectively. These results demonstrate the potential of this method to mitigate the matrix effect in the quantification of microplastics across diverse sediment types.