The objective of the present research was to optimize the model of sugar content in navel orange for improving the detection presicion by the online near infrared spectroscopy. The reference wavelength was chosen by coefficient of variation of the different wavelengths in the calibration set in the wavelength range of 700.28 - 933.79 nm. Then the spectra were transformed into ratio specra. The absorbance and ration spectra were pretreated by different preprocessing methods. The models of sugar content were developed by partial least squares (PLS) and least squares support vector regression (LSSVR). The 30 unknown navel orange samples were applied to evaluate the performance of the models. By comparison of the predictive performances, the LSSVR model was the best among the models with the first derivative preprocessing and ration spectra. The correlation coeffiecient (R(P)) of the best model was 0.85, the root mean square error of prediction (RMSEP) was 0.41 Brix. The results suggested that it was feasible to improve the precision of online near infrared spectroscopy detecting sugar content in navel orange by the optimization of reference wavelengths, the first derivative preprocessing and LSSVR.