In the wavelength range of 350～1?800?nm, different preprocessing methods were developed to improve the predictive ability and robustness of model, but the effect was not better than interval partial least square (iPLS). The characteristic wavelength bands of 642～787nm, 861～1006nm, 1080～1152nm and 1226～1371nm were determined by iPLS characteristic wavelength bands selection method. The performance of the model calibrated with sensitive wavelength bands was predicted by 30 unknown Nanfeng mandarin fruits, the correlation coefficient and root mean square error for prediction (RMSEP) were 0.95 and 0.55°Brix respectively. It was suggested that near infrared spectroscopy (NIRS) combination with iPLS characteristic wavelength bands selection method was a powerful tool to improve the predictive ability and robustness of nondestructive measurement model. 