Research on Citrus grandis Granulation Determination Based on Hyperspectral Imaging through Deep Learning

• Published in: Food analytical methods Vol. 14; no. 2; pp. 280 - 289
• Main Authors: Jie, Dengfei, Wu, Shuang, Wang, Ping, Li, Yan, Ye, Dapeng, Wei, Xuan
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.02.2021; Springer Nature B.V
• Subjects: Accuracy; Analytical Chemistry; Artificial neural networks; Back propagation networks; Chemistry; Chemistry and Materials Science; Chemistry/Food Science; Citrus fruits; Convolution; Deep learning; Food Science; Fruits; Functional groups; Granulation; Hyperspectral imaging; Microbiology; Model accuracy; Neural networks; Nondestructive testing; Support vector machines; Granulation; Convolution neural network; Spectrum analyze; Internal quality; Batch normalization
• DOI: 10.1007/s12161-020-01873-6
• ISSN: 1936-9751

Citrus fruit granulation is a major physical disorder during late maturity and post-harvest storage, and it greatly undermines the quality of fruit. Currently, there is still a lack of rapid and nondestructive methods to detect citrus granulation. This study proposes a nondestructive granulation detecting method based on deep learning. Different models were established with the input of preprocessed transmission spectra obtained by hyperspectral imaging. Conventional convolution neural network (CNN) got the best accuracy at 88.02% for training, compared with the least-square support vector machine (LS-SVM) and back-propagation neural network (BP-NN). After adding the batch-normalization layer to the CNN, the experimental results show that the detection model obtained a 100% accuracy in train set and 97.9% in validation set, respectively. And then, through analyzing the well-trained model layer by layer, bands of 660.2–721.1 nm, 708.5–750 nm and 806.5–847 nm were the spectra greatly related to granulation. The model rebuilt with these feature bands obtained 90.1% and 85.4% accuracy in train set and validation set, respectively. This way, effective wavelength selection can find bands highly correlated with granulation.Combined with some research on functional group, it is possible that inference to internal matter changes in granulation process, which may provide some hints to explore the reason of granulation. It is also meaningful to develop granulation-detecting equipment for citrus fruits.