Nondestructive Detection of Soluble Solids Content of Apples from Dielectric Spectra with ANN and Chemometric Methods

To investigate the feasibility of using dielectric spectra in nondestructively determining the soluble solids content (SSC) of fruits, the dielectric constants and loss factors of 160 apples of three varieties (Fuji, Red Rome, and Pink Lady) were obtained at 51 discrete frequencies from 10 to 1800 MHz with an open-ended coaxial-line probe and an impedance/material analyzer. Based on the joint x-y distances sample set partitioning (SPXY) method, 106 apples were selected for the calibration set and the other 54 samples were used for the prediction set. The principal component analysis (PCA), uninformative variables elimination method (UVE-PLS), based on partial least squares, and successive projection algorithm (SPA) were applied to extract characteristic variables from original full dielectric spectra. The generalized regression neural network (GRNN), support vector machine (SVM) and extreme learning machine (ELM) modeling methods were used to establish models to predict SSC of apples, based on the original full dielectric spectra and characteristic variables, respectively. Results showed that four principal components were selected as characteristic variables by PCA, 15 dielectric constants and 14 loss factors at different frequencies were selected as characteristic variables by UVE-PLS, and one dielectric constant and ten loss factors were chosen as feature variables by SPA. ELM combined with SPA had the best SSC prediction performance, with calibrated correlation coefficient and predicted correlation coefficient of 0.898 and 0.908, respectively, and calibrated root-mean-square error and predicted root-mean-square error of 0.840 and 0.822, respectively. The study indicates that dielectric spectra combined with artificial neural network and chemometric methods might be applied in nondestructive determination of SSC of apples.