Identification of the proximate geographical origin of wolfberries by two-dimensional correlation spectroscopy combined with deep learning

Abstract

In this study, two-dimensional correlation spectroscopy (2D-COS) of near-infrared hyperspectral images combined with convolutional neural networks (CNN) was developed to identify the origin of wolfberries for the first time. 2D-COS was adopted to identify characteristic wavelengths and resolve the change orders of corresponding chemical bonds. Competitive adaptive reweighed sampling (CARS), iteratively retaining information variables (IRIV) and interval variable iterative space shrinking analysis (iVISSA) methods were used to select characteristic wavelengths. Linear discriminant analysis (LDA), partial least squares discriminant analysis (PLS-DA), support vector machine (SVM) and CNN classification models of the original spectra and characteristic wavelengths were established. Wolfberry texture information was extracted by the grey-level co-occurrence matrix (GLCM) method, and fused with optimal characteristic wavelengths to optimize the identification results of the models. The results showed that the sequence of changes in the correlation spectra caused by fluctuation in geographical origins in sequence was 1556 nm, 1437 nm, 1058 nm, 1368 nm. The stretching vibration of the NH bonds and CN bonds (1556 nm) in the amide II bands preceded the bending vibration of the NH bonds and CN bonds (1437 nm) in the amide III bands. Stretching vibration of the COH bonds (1058 nm) preceded double-frequency absorption bands of the CH bonds (1368 nm). For the original spectral dataset, the 2D-COS-CNN model performed the best, with the calibration set and prediction set accuracies of 100% and 95.29%, respectively. For the characteristic wavelength dataset, the 2D-COS-iVISSA-CNN model exhibited the best accuracy, with the calibration set and prediction set accuracies of 100% and 96.67%, respectively. Using the optimized fusion dataset, the CNN discrimination model showed the best results, with the calibration and prediction set accuracies of 100% and 97.71%, respectively. 2D-COS combined with deep learning algorithm can effectively distinguish the origin of wolfberries and provide crucial technical support for the development of wolfberry industry.