Nondestructive determination of edible quality and watercore degree of apples by portable Vis/NIR transmittance system combined with CARS-CNN

Visible Near-infrared (Vis/NIR) transmittance spectroscopy has been employed to evaluate various quality indicators of fruits, owing to the advantages of rapid non-destructive detection, real-time monitoring, and applicability to opaque samples. The portable Vis/NIR transmittance system was autonomously developed and coupled with chemometrics methods for the detection of quality indicators including soluble solids content (SSC), watercore degree, firmness and pH of apples. Preprocessing method of Savitzky-Golay (SG) smoothing combined with standard normal variate SNV) transformation was employed to mitigate spectra noise. The feature variables were selected using variable selection methods and coupled with partial least square (PLS) to establish linear regression prediction models. Furthermore, a deep learning model was developed by combining competitive adaptive reweighted sampling (CARS) with convolutional neural network (CNN). The R_p of SSC, firmness, pH and the watercore degree of CARS-CNN model were 0.951, 0.824, 0.828 and 0.943, respectively, and the prediction performance was improved by 3%, 11.2%, 11.2% and 9.3% in comparison with the full-spectra model, respectively. The results indicate that the predictive accuracy of models utilizing CARS-CNN outperformed the other models. This research provides a portable, practical and efficient method for the field of non-destructive detecting of fruit quality to obtain fruit Vis/NIR transmittance spectra and establish quality prediction models. These results demonstrated the feasibility of utilizing the Vis/NIR transmittance spectroscopy combined with CARS-CNN method in the detection of quality indicators for apples.