In this study, near-infrared (NIR) spectroscopy was exploited for non-destructive determination of theanine content of oolong tea. The NIR spectral data (400-2500 nm) were correlated with the theanine level of 161 tea samples using partial least squares regression (PLSR) with different wavelengths selection methods, including the regression coefficient-based selection, uninformative variable elimination, variable importance in projection, selectivity ratio and flower pollination algorithm (FPA). The potential of using the FPA to select the discriminative wavelengths for PLSR was examined for the first time. The analysis showed that the PLSR with FPA method achieved better predictive results than the PLSR with full spectrum (PLSR-full). The developed simplified model using on FPA based on 12 latent variables and 89 selected wavelengths produced R-squared (R2) value and root mean squared error (RMSE) of 0.9542, 0.8794 and 0.2045, 0.3219 for calibration and prediction, respectively. For PLSR-full, the R2 values of 0.9068, 0.8412 and RMSEs of 0.2916, 0.3693, were achieved for calibration and prediction. Also, the optimized model using FPA outperformed other wavelengths selection methods considered in this study. The obtained results indicated the feasibility of FPA to improve the predictability of the PLSR and reduce the model complexity. The nonlinear regression models of support vector machine regression and Gaussian process regression (GPR) were further utilized to evaluate the superiority of using the FPA in the wavelength selection. The results demonstrated that utilizing the wavelength selection method of FPA and nonlinear regression model of GPR could improve the predictive performance.
Consumption of agricultural products with pesticide residue is risky and can negatively affect health. This study proposed a nondestructive method of detecting pesticide residues in chili pepper based on the combination of visible and near-infrared (VIS/NIR) spectroscopy (400-2498 nm) and deep learning modeling. The obtained spectra of chili peppers with two types of pesticide residues (acetamiprid and imidacloprid) were analyzed using a one-dimensional convolutional neural network (1D-CNN). Compared with the commonly used partial least squares regression model, the 1D-CNN approach yielded higher prediction accuracy, with a root mean square error of calibration of 0.23 and 0.28 mg/kg and a root mean square error of prediction of 0.55 and 0.49 mg/kg for the acetamiprid and imidacloprid data sets, respectively. Overall, the results indicate that the combination of the 1D-CNN model and VIS/NIR spectroscopy is a promising nondestructive method of identifying pesticide residues in chili pepper.