Determination of Huanghua Pear's Harvest Time Based on Convolutional Neural Networks by Visible-Near Infrared Spectroscopy

The maturity of fruit at harvest time determines its final eating quality. Choosing the optimal harvest time of fruit is one of the key issues to improve fruit quality. Visible/near-infrared spectroscopy technology is suitable for fruit maturity and harvest time determination because of its rapid and non-destructive detection characteristics. Due to the large difference in fruit quality on the tree, traditional chemometric methods require complex spectral pretreatment, and the model is not suitable for different seasons, orchards, etc. In this paper, the discrimination model of Huanghua pear's harvest time by full convolutional neural networks (CNNs) based on visible/near infrared spectroscopy (Vis/NIR) was proposed. The CNNs was used for spectral feature extraction, and the error backpropagation algorithm combined with the random gradient descent method was used to adjust the connection weights between layers, and output the Logistic regression of harvest time determination, which implemented the end-to-end discrimination of Huanghua pear's harvest time, and the result was compared with the PLSDA method. The one-dimensional convolutional neural networks included one input layer, two convolution layers, a pooling layer and one Softmax output layer, using cross-entropy as the loss function, and the L2 regularization was used as the regular term to avoid overfitting, without preprocessing. A total of 450 samples were collected for two years. Three hundred samples in the first year constituted training set, 90 samples constituted test set one, and 60 samples in the second year constituted test set two. The results have shown that when the test set and the training set were collected from the same year, correct discrimination rate of PLSDA and CNNs models was 100%, when the test was from different years, correct discrimination rate reduced to 41.67% and 88.33%, respectively. The correlation coefficient and the mutual information of the modes indicated that the CNN model could take advantage of full-spectrum information. Therefore, the CNNs method optimizes convolution kernels through iteration to achieve more flexible preprocessing, which can reduce the difficulty of model training. The established model has good ability of explanation and generalization. The proposed method could be applied in discrimination of fruit's harvest time.