In order to improve the predictive precision for on-line determination of soluble solids content (SSC) of Gannan navel orange, the dynamic detecting system was applied to optimize online detection model by visible and near-infrared reflectance spectroscopy. The spectra were obtained at the constant velocity of 5 navel oranges per second. After employing various preprocessing methods, the sensitive spectral regions were analyzed by different variable selection methods, including backward interval partial least-squares (BiPLS), genetic algorithm (GA), and competitive adaptive reweighted sampling (CARS). The predictive abilities of the models were evaluated by prediction set. The results indicated that the best model was obtained by CARS with the first derivative. The correlation coefficient (Rp) and root mean square error of prediction (RMSEP) was 0.94 and 0.42% for SSC respectively. The results showed that the proposed method of CARS could effectively simplify the online detection model of SSC of Gannan navel orange based on visible/near-infrared (Vis/NIR) diffuse transmittance spectroscopy, and enhance the predictive precision. The study can provide a reference for optimizing online detecting system of Gannan navel orange.