Hyperspectral techniques were used to study the sugar content of different parts of citrus, and the sugar content detection models with hyperspectral information of calyx, stem and equator part were established respectively. The results showed that the model established by calyx was better than that of stem and equator. The detection models of partial least squares regression (PLSR), principal component regression (PCR), and stepwise multivariate linear regression (SMLR) were established respectively, and the results of these three models were close. The PLSR model was found to the best among them, after Norris derivative pretreatment methods were applied, the prediction correlation coefficient (r_pre) and the root mean square error of prediction (RMSEP) were 0.950 and 0.636 °Brix. This result inclined that it was feasible to use the hyperspectral technology to detect the sugar content in different parts. The study indicated that the calyx part could be the prior choice for the sugar content detection site in the citrus quality testing, and the conclusion has great significance for the way of citrus place in the actual production. Moreover, the PLSR method was used to establish the model of hyperspectral information and average sugar content in calyx, stem and equator part. The highest prediction r_pre and RMSEP of models was in the calyx and only to be 0.913 and 0.621 °Brix, which was not excellent enough. Therefore, it was limited to predict the citrus average sugar content with the hyperspectral information of a certain part.