During the production and processing of food or raw materials, contamination may be caused by various external causes and illegal human factors. The detection of such substances by infrared spectroscopy is of great importance.

Arzu Yazici et al. ^[24] established a method for rapid nondestructive detection of pesticide residues in strawberries based on near-infrared spectroscopy. By performing the second-order derivative, principal component analysis (PCA) on the spectral data, the prediction models for aminopyralid and azoxystrobin were developed using PLSR: the prediction correlation coefficient (R_P) for aminopyralid was 0.93, the root mean square error of prediction (RMSEP) was 3.22 mg kg⁻¹, and the relative prediction error (RPD) was 2.28. The prediction correlation coefficient (R_P) for azoxystrobin was 0.91. Lu et al. ^[25] used visible-NIR spectroscopy combined with chemometric methods to quantify chlorpyrifos and carbendazim residues in cabbage. The quantitative models were developed using the partial least squares regression (PLSR) and least squares support vector machine (LS-SVM) methods. The feature variables were selected using the continuous projection algorithm (SPA), random frog algorithm, and PLSR methods. The LS-SVM model performed better than the PLSR model. The prediction correlation coefficient (R_P) of chlorpyrifos samples was 1 and the root mean square error of prediction (RMSEP) was 0.03 mg kg⁻¹. The average spiked recoveries ranged from 98.95–102.26% and the relative prediction errors ranged from 0.88–9.97% when the chlorpyrifos concentration was greater than 5 mg kg⁻¹. The prediction correlation coefficient (R_P) of chlorpyrifos samples was close to 1, the prediction mean-spiked recoveries ranged from 99.10% to 100.66%, and the relative prediction errors ranged from 0.39% to 5.01% when the concentration of carbendazim was greater than 1 mg kg⁻¹, and the two samples had good consistency and reproducibility.

Liu et al. ^[26] investigated a method based on the radial basis function (RBF) neural network combined with NIR spectral data to predict the talc content in wheat flour. In this method, sample data were processed by multiple scattering correction, and the correlation coefficient method was used to reduce the spectral redundancy to determine the maximum relevant information wavelength. The prediction correlation coefficient (R_P) of the optimal MSC-CCM-RBF model was 0.9999, the root mean square error of prediction (RMSEP) was 0.0765, and the relative prediction error (RPD) was 65.0909. Che et al. ^[27] used Vis-NIR reflectance spectroscopy based on physical property analysis to predict the concentration of azodicarbonyl in flour. An amount of 101 samples with a concentration gradient of 3 mg kg⁻¹ were prepared using a stepwise dilution method in the concentration range of 0–300 mg kg⁻¹. The abnormal samples were identified and rejected by combining the Marxian distance method and leave-one-out cross-validation, and the radial basis function model could better predict the concentration of azodicarbonyl in flour by selecting the characteristic band through correlation and using the preprocessing method of the first-order derivatives and SNV. The method yielded prediction correlation coefficients and the root mean square error of prediction (RMSEP) of 0.99996 and 0.5467, which were within 0.01 for each sample, and LOD and LOQ of 3.2 and 9.7 mg kg⁻¹, respectively. Li et al. ^[28] used Fourier transform transmission infrared (FTIR) spectrometry to determine Pb–Cr green in green tea. Partial least squares discrimination (PLS-DA) was used for the qualitative analysis of Pb–Cr green, and the classification was 100% correct. The interval partial least squares (iPLS) regression combined with the successive projection algorithm (SPA) was proposed to select the characteristic wavenumbers for quantitative analysis of Pb–Cr green, and the least squares support vector machine algorithm (LS-SVM) was used to obtain the optimal model with R_P² of 0.864 and a root mean square error of prediction (RMSEP) of 0.291. The results showed that infrared spectrometry was feasible for the detection of Pb–Cr green in green tea. Kurrey et al. ^[29] used diffuse reflectance Fourier transform infrared spectroscopy (DRS-FTIR) for the rapid quantitative determination of antibiotics, ciprofloxacin (CIP), and norfloxacin (NOR) in poultry egg samples. The linear range of DRS-FTIR for the detection of CIP and NOR in poultry egg yolk was 0.05–0.50 ng mL⁻¹, the limits of detection (LODs) were 0.032 ng mL⁻¹, and the recoveries were 83.1%-102.3%. The method is simple, sensitive and suitable for high-throughput analysis of food samples. Gu et al. ^[30] studied the chemical morphological changes of squid immersed in different concentrations of formaldehyde using three infrared spectra (FTIR, SD-IR, and 2DCOS-IR). The predicted values of the constructed model were close to the actual formaldehyde concentration values in squid, with a prediction correlation coefficient of 0.9774, a root mean square error of prediction (RMSEP) of 6.08 and a limit of quantification of 15 mg kg⁻¹. The method took 5 min to determine formaldehyde, while the HPLC method took 1.5 h to determine formaldehyde.

The analysis of the above studies shows that infrared spectroscopy is a good technique for the nondestructive detection of pesticide residues in fruits and vegetables, illegal additives in flour and green tea, and antibiotics in eggs. However, differences in the manufacturers of fruits, vegetables, and eggs from different origins and illegal additives used in the experiments may lead to the differentiation of the final models. Therefore, the establishment of a set of standardized and perfect data models for detecting illegal food additives, pesticides, and antibiotics could well improve the universality of the models and broaden the application of this technology.