Improving the identification accuracy of sugar orange suffering from granulation through diameter correction and stepwise variable selection

Granulation is one of the main diseases for citrus fruit, causing the loss of water and nutrients. To prevent citrus with granulation from flowing into the market, it is essential to identify them. In this study, sugar oranges suffering from granulation were detected online using visible/near-infrared (Vis/NIR) spectroscopy technology. Diameter correction and stepwise variable selection were optimized to improve the identification accuracy. To eliminate or weaken the effect of different sample sizes on the Vis/NIR transmission spectrum which leads to the decline of accuracy, the average extinction coefficient inside the fruit was calculated to correct the transmittance spectra of different sizes of citrus, and the effective variables of the spectrum were selected stepwise using variable importance of projection (VIP), selectivity ratio (SR) and competitive adaptive reweighted sampling (CARS). Four different pretreatment methods (standard normal variate (SNV), multiplicative scatter correction (MSC), mean center, 1st derivative) were used to process the spectra before and after correction, and two modeling methods (partial least squares discriminant analysis (PLSDA) and support vector machine (SVM)) were combined to develop the identification model. The results showed that the recognition accuracy of the models built from the corrected spectra was generally better than that of the uncorrected ones. The SNV-Mean Center-CARS-PLSDA model was optimal, with a discrimination accuracy of 94.00 % and an average discrimination error rate of 5.84 % for healthy and diseased samples. This study demonstrates that the proposed fruit diameter correction method combined with effective variable preference can effectively improve the discrimination accuracy of citrus granulation online, which is important for improving fruit quality and protecting consumers' interests.