Quantitative analysis of polycyclic aromatic hydrocarbons in soil by infrared spectroscopy combined with hybrid variable selection strategy and partial least squares

Infrared spectroscopy (IR) combined with multivariate calibration technology can be used as a potential method to quantitative analysis of polycyclic aromatic hydrocarbons (PAHs) in soil, which provides a rapid data support for soil risk assessment. However, IR spectrum contains lots of useless information, its predictive performance is poor. Variable selection is an effective strategy to eliminate irrelevant wavelengths and enhance predictive performance. In this study, IR combined with partial least squares (PLS) was proposed to quantify anthracene and fluoranthene in soil. In order to improve the predictive performance of the PLS calibration model, the synergy interval PLS (siPLS) method was first used for rough selection to select feature bands; on this basis, fine selection was performed to extract the feature variables. In fine selection, three different feature variables selection methods, such as successive projection algorithm (SPA), genetic algorithm (GA), and particle swarm optimization (PSO), were compared for their performance in extracting effective variables. The results show that the siPLS-GA calibration model receive a lowest root mean square error (RMSE) and a largest determination coefficient (R-2). Results of external validation demonstrate an excellent predictive performance of siPLS-GA calibration model, with the R-2 = 0.9830, RMSE = 0.5897 mg/g and R-2 = 0.9849, RMSE = 0.4739 mg/g for anthracene and fluoranthene, respectively. In summary, siPLS combined with GA can accurately extract the effective information of the target substance and improve the predictive performance of the PLS calibration model based on IR spectroscopy. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

Infrared spectroscopy combined with partial least squares (PLS) can be used for quantitative analysis of PAHs in soil, but variable selection methods are needed to extract effective information and improve predictive performance. The siPLS-GA calibration model was used in this study to extract feature variables, showing a low RMSE and high R-2, with excellent predictive performance demonstrated in external validation.