Accurate detection of low concentrations of microplastics in soils via short-wave infrared hyperspectral imaging

Abstract

This study evaluated the effectiveness of coupling machine learning algorithms with short-wave infrared hyperspectral imaging in detecting two types of microplastics - polyamide and polyethylene - with the maximum particle sizes of 50 and 300 ​μm, respectively, across three concentration ranges (0.01–0.10, 0.10–1.0, and 1.0–12 ​%) in soils. Using indium gallium arsenide (InGaAs; 800–1600 ​nm) and mercury cadmium telluride (MCT; 1000–2500 ​nm) sensors, we applied logistic regression and support vector machines by employing both linear and nonlinear kernels to analyze spectral features extracted via principal component analysis and partial least squares. The results demonstrated that the overall accuracy for detecting 0.01–12% microplastics was 93.8 ​± ​1.47% using the MCT sensor, which was higher than 68.8 ​± ​3.76 ​% using the InGaAs sensor. Both sensors showed high accuracy (>94 ​%) when detecting high levels at 1.0–12%) of microplastics in soil. But these accuracies greatly declined as the spiked microplastics concentrations decreased from 1.0–12 to 0.10–1.0% and further to 0.01–0.10%. Moreover, this decline was more pronounced for the InGaAs sensor compared to the MCT sensor and for sub-wavelength spans compared to the full wavelength span under each sensor. The MCT sensor consistently outperformed the InGaAs sensor across all three concentration ranges, potentially due to its extended coverage of 1600–2500 ​nm and high sensitivity of the detector. Our study highlights the feasibility of the MCT hyperspectral imaging system for rapid and effective detection of microplastics in soils non-invasively at concentrations as low as 0.01%.

Keywords

• Polyamide
• Polyethylene
• Machine learning
• Partial least squares
• Principal conponent analysis
• Linear and nonliearn kernels