In order to solve the problem of the reduction in the yield of camellia oil caused by inaccurate judgment of the picking date of Camellia oleifera, hyperspectral imaging technology combined with chemometrics was used to qualitatively classify the maturity of C. oleifera. The changes in the physicochemical and spectral characteristics of C. oleifera fruit at different maturity stages were examined. Curvature correction of hyperspectral images was carried out. Four classification algorithms were used to establish a discriminant model for C. oleifera fruit maturity based on the full-band spectral data, and it was found that the support vector machine (SVM) model had the highest classification accuracy (97%). Five feature variable selection methods were used to reduce the dimension of the full-band spectral data, and it was found that the accuracy of the model based on the feature wavelengths selected by competitive adaptive reweighted sampling (CARS) was the highest (82%). The accuracy of the SVM model based on a combination of the color and texture features extracted from the hyperspectral images was higher than that of the model based on single spectral features (whose dimension was reduced by CARS), with classification accuracy for the training and test sets of 95% and 93%, respectively. In conclusion, hyperspectral imaging technology can be used to classify C. oleifera with different maturities, which will provide a scientific basis for the judgment of the best picking date of C. oleifera, and is of great significance in ensuring the maximum tea seed yield and the optimum oil quality.

This study explored the feasibility of applying visible and near-infrared (400‒1000 nm), and mid- and short-wave near infrared (900‒1700 nm) spectroscopy for detecting Camellia oleifera fruit with different degrees of natural mildew. The near infrared spectra of the equatorial shady and sunny sides, and the junction surfaces of samples with different degrees of mildew were collected in two wavelength bands. The average spectra were analyzed by principal component analysis (PCA), revealing that samples with different mildew degrees could be clustered into different groups, and the first and second principal components (PC1 and PC2) were effective in distinguishing the samples in each category. The full-spectrum partial least squares-discriminant analysis (PLS-DA) model based on the original spectra performed better than of its counterpart based on the preprocessed spectra. In the selection of characteristic wavelengths, successive projections algorithm (SPA) was found to be superior to PC loadings in establishing the simplified models for both spectral ranges. The correct classification rate and kappa coefficient were 84.4% and 0.7667 for the prediction set, respectively. Based on the observation of confusion matrices for the prediction set, the specificities of the two optimal simplified models for the prediction of each degree of mildew were equivalent to each other and above 0.84. However, mid- and short-wave near infrared spectra in the wavelength range of 900–1700 nm provided slightly higher sensitivity (0.72) in classifying samples with moderate degree of mildew. Our study showed that near-infrared spectroscopy could be used to detect the degree of natural mildew of C. oleifera fruit, and visible and near-infrared spectroscopy showed comparable results to mid- and short-wave near infrared spectroscopy. Considering its lower cost, visible and near-infrared spectroscopy has a better application prospect for real-time detection.

This study aimed to develop a method based on hyperspectral imaging (400–1 000 nm) combined with chemometrics to identify the diverse geographical origins of Lycium barbarum fruit. Hyperspectral images of L. barbarum fruit from Ningxia, Gansu, Inner Mongolia, Qinghai and Xinjiang were acquired, and spectral data was extracted from the region of interest (ROI) by threshold segmentation method. Multiple preprocessing methods were employed to eliminate the interference information from the spectra, and the results showed that the discriminant model based on normalized reflectance spectrum (NR) exhibited better performance. Furthermore, the successive projection algorithm (SPA), competitive adaptive reweighted sampling (CARS), particle swarm optimization (PSO), iteratively retaining informative variables (IRIV), and CARS + IRIV were used to select characteristic wavelengths. The results showed that the simplified model based on the wavelengths selected by CARS + IRIV had the best performance. In the models ranging from binary to quintuple classifications, the selected characteristic wavelengths accounted for only 15.6% to 27.7% of the full spectra. The prediction accuracy was 97.7%, 90.9%, 89.2%, and 87.1%, respectively. In addition, a confusion matrix was employed to visualize the optimal simplified classification model in order to intuitively distinguish the classification categories. Satisfactory sensitivity, specificity and Kappa coefficients were obtained in classifying L. barbarum. The results illustrated that hyperspectral imaging technology combined with chemometric methods could effectively identify the geographical origin of L. barbarum and provide crucial technical support for the development of the L. barbarum industry.