Brix is one of the important indicators for evaluating the internal quality of navel orange. Due to the difference in the size of the fruit itself, the sugar content prediction model is poor in robustness and the prediction accuracy is not high. Therefore, eliminating the influence of fruit size effect is of great significance for improving the accuracy of fruit sorting model. The diffuse transmission, multi-point emission and reception, and circular emission and reception diffuse reflectance spectra of navel orange were compared and analyzed. Among different detection platforms, the spectral energy of the big fruit was stronger than that of the small fruit due to the difference of optical path difference, and the circular emission was obtained. The energy of the diffuse reflection spectrum is stronger than that of the other two spectra. The diffuse transmission spectrum energy is the weakest, and the peaks and troughs are roughly the same. The prediction model of orange navel size under different detection methods is established respectively. Among them, the prediction coefficient of the size prediction model under the diffuse transmission detection mode is 0.60, the root mean square error of the prediction set is 3.95 mm, and the size of the multi-point transmission and reception diffuse reflection detection mode. The prediction set correlation coefficient of the prediction model is 0.97. The prediction set RMS error is 1.46 mm, and the prediction set correlation coefficient of the small fruit prediction model under the ring emission and reception diffuse reflection detection mode is 0.96, and the prediction set RMS error is 1.73 mm. The mixed fructose prediction models of large fruit, small fruit, mixed fruit and multi-scattering correction pretreatment were established under three different detection methods. The precision of the sugar content prediction model of small fruit was higher than that of large fruit and mixed fruit, and diffuse transmission detection The correlation coefficient of the prediction set of the small fruit prediction model is 0.76, the root means square error of the prediction set is 0.81°Brix, and the correlation coefficient of the prediction set of the small fruit prediction model under the multi-point transmission and reception diffuse reflection detection mode is 0.72. The square root error is 0.97°Brix, and the prediction set correlation coefficient of the small fruit prediction model under the ring emission and reception diffuse reflection detection mode is 0.72, and the prediction set RMS error is 0.93°Brix. After multi-scattering correction pre-processing spectra, the hybrid fruit model of near-infrared diffuse transmission spectrum is better than the small fruit model. The correlation coefficient of the model prediction set is 0.84, and the root means square error of the prediction set is 0.64°Brix. In the diffuse reflection detection mode, the accuracy of the multi-mixed fruit model is reduced. The experimental results show that in the diffuse transmission detection method, the multi-scatter correction pre-processing spectrum can eliminate the effect of the size effect. In the diffuse reflection detection method, the size sorting is performed first, followed by the sugar separation, which can also avoid the size effect. This study provides reference and theoretical support for the rapid online sorting of bulk fruits.