Comprehensive Quality Evaluation of Blueberries Based on Near-infrared Spectroscopy and Fiber Optic Droplet Analysis

Two-dimensional near-infrared (2D-NIR) correlation spectroscopy integrated with droplet analysis was used for comprehensive evaluation of the storage quality of blueberries. NIR spectra and droplet fingerprints of ‘Emerald’ blueberries with 8 storage periods were collected, and 15 physicochemical indexes such as hardness, anthocyanin content, VC content and solid/acid ratio were considered, among which a close correlation was found. Therefore, membership function analysis (MFA) combined with principal component analysis (PCA) was employed to calculate comprehensive scores for blueberries from the 15 physicochemical indexes, and based on the comprehensive scores, the storage quality of blueberries was graded. The spectral data were preprocessed by Savitzky-Golay (SG) convolutional smoothing, standard normal variate (SNV) transformation, multiple scattering correction (MSC) or adaptive iteratively reweighted penalized least squares (airPLS). After comparative analysis, it was found that SG convolutional smoothing yielded the highest predictive model accuracy of 82.67%. After averaging and dimensional reduction, the droplet data were preprocessed by SG convolutional smoothing, Gauss filtering or median filtering. SG convolutional smoothing was found to provide the highest predictive model accuracy of 86.67%. Using the comprehensive scores of blueberries as the external disturbance, two-dimensional correlation analysis was carried out on the spectral and droplet data, and the positions of the autocorrelation peaks at 879, 1 019, 1 220 and 1 636 nm and at 789, 1 653, 2 386 and 2 703 ms were selected as the feature variables, respectively. Support vector machine (SVM) and random forest (RF) models were established using the fusion of spectral and droplet feature data as the input, and the predictive accuracy of the models were 100.00% and 98.33%, respectively, which were higher than those obtained using the single features as the input; the SVM model had a better prediction effect than the RF model. Then, nine blueberry varieties such as ‘Sapphire’, ‘Legacy’ and ‘Bluecrop’ were used for validation of the SVM model established using a series of data processing methods. The results showed that the SVM model exhibited good predictive effects for all varieties. Therefore, visible/near-infrared spectroscopy combined with droplet analysis enables the prediction of the comprehensive storage quality of blueberries, providing a new method for the quality evaluation of blueberries.