Nondestructive detection of apple watercore disease based on electric features

Abstract: In order to find more cost-saving and efficient technology for non-destructive detection of watercore apple, recognizing the disease by following electric feature changes of the fruit was tested in this study. With the suspected watercore fruit and sound fruit of Malus pumila cv. Qinguan as material, we collected 143 feature data of 11 electric parameters at 13 frequency points from 100 Hz to 3.98 MHz fruit by fruit. Then each fruit was crosscut to tell and record whether watercore occurred in it. All the features data were analyzed by 3 steps. The first 2 steps were to screen differential features between sound and watercore apple, and then determine principal components (PCs) whose cumulative variance contribution rate reached over 90%. In the third step, different classification models were used to discriminate the sound and watercore fruit in combination with PCs obtained. The results showed that the incidence of watercore caused the increase in feature values of dielectric loss coefficient, complex impedance angle, series equivalent capacitance, parallel capacitance, relative dielectric constant, and loss factor at low frequency region (100-10000 Hz), a total of 36 differential feature values. These findings supported theoretically the possibility to discriminate sound and watercore apple based on differences in their electric features. Using principal component analysis, 15 and 7 PCs were extracted for original group of 143 features, and the group of 36 differential features, respectively. Accuracy rates of Fisher discrimination and multilayer perceptron (MLP) artificial neural network for the groups of 143 features and 36 differential features all elevated with the increase of PCs number, and reached a stable high level when PC number reached 13 and 10, respectively. Accuracy rates of Fisher discrimination and MLP for the group of 143 features using the former 13 PCs reached 93.8% and 95.4%, while for the group of 36 features using its former 7 PCs reached 91.7% and 93.8%, respectively. It indicated that the discrimination ability between sound and watercore apples was ascribed to mainly the 36 differential electric features. Discrimination by radial basis function (RBF) modeled by using 15 PCs reached an accuracy rate of 75.1% for the group of 143 features. Quality profiles of 2 kinds of apples differed in density, firmness, and soluble solids, which presented significantly higher level in watercore fruit (P<0.05), but titrate acids content was significantly lower. Physio-chemical characteristics changes resulted in the alternation of electric feature of watercore fruit and showed multiple-to-multiple correlation of cause-effect. Values of loss factor at low frequencies (100-25100 Hz) combined with MLP or RBF classifier all achieved accuracy rates of 100% on the recognization of either watercore or sound apple, which can be selected as the simple and effective method for apple watercore detection. The result can provide theoretical and technical support for the development of on-line equipment which can non-destructively detect the disease of apple watercore in the future.