[Objective] To improve the accuracy of online measurement of sugar content of grapefruit by near infrared spectroscopy. [Methods] The pomelo online non-destructive testing equipment developed by ourselves was used to collect diffuse transmission spectrum data of pomelo in three light regions. In the wavelength range of 650~950 nm, orthonormal variable transformation （SNV）, multiple scattering correction （MSC）, Normalize, Savitzky-Golay first order derivative, SG-1st preprocessed the original data, used the adaptive weighting algorithm （CARS） to screen the spectral characteristics of the grapefruit sugar content, and established a partial least squares regression （PLSR） model. 30 grapefruit samples that were not involved in the modeling were used for online verification. [Results] The modeling effect of light region C combined with SNV-CARS-PLSR method was the best. The coefficient of determination of the prediction set was 0.95 and the root-mean-square error was 0.30 °Brix. In online verification, the coefficient of determination was 0.90 and the root mean square error was 0.58 °Brix. The model had a strong ability to detect the sugar content of grapefruit on line. [Conclusion] The prediction model based on the spectral data collected under the condition that the light spot diameter is 70 mm and the light region C is 20 mm above the equator of the grapefruit can realize the online prediction of the sugar content of the grapefruit more effectively.