Early detection of freezing damage in oranges by online Vis/NIR transmission coupled with diameter correction method and deep 1D-CNN

Abstract

Realizing online detection of early freezing damage of citrus fruits is meaningful and profitable in the existing postharvest sorting system of fruits and vegetables. Transmission spectra of 114 oranges in the range of 644–900 nm were obtained using a self-designed online spectral measurement system in this study. Fruit size seriously affected the intensity of transmission spectra and thus reduced the detection accuracy of the model for early freezing damage. To solve this problem, a new diameter correction method (DCM) was proposed. The results showed that DCM could eliminate the effect of fruit size on transmission spectra more effectively than multiplicative scattering correction (MSC) and standard normal variable (SNV), and partial least squares discrimination analysis (PLSDA) and support vector machine (SVM) models based on DCM pretreated spectra had better performance. To eliminate the collinearity variables in the original spectra and simplify the model, competitive adaptive reweighted sampling (CARS) and successive projections algorithm (SPA) were used to extract effective wavelengths (EWs). The accuracy of DCM-CARS-SPA-PLSDA model established by 15 EWs for early freezing damage identification met the requirement of online detection. A one-dimensional Convolutional Neural Network (1D-CNN) architecture was proposed in this study to further improve the detection accuracy. The model established by combining DCM and 1D-CNN had the best performance. In the prediction set, the recall of the optimal model for the early freeze-damaged oranges and unfrozen oranges was 95.15 % and 88.54 %, and the overall accuracy was 91.96 %. Therefore, the DCM and 1D-CNN method proposed in this study can effectively eliminate the effect of fruit size on transmission spectra, and enable the model to effectively identify freezing damage.