This article used hemicelluloses content in acacia spp. wood as a case study to demonstrate the influence of noise in the reference data on the results of NIR calibration model. The results indicated that the accuracy of NIR calibration model was affected by the reference data noise. The less noisy data was used in calibration model, the better result could be obtained. But when the noise was larger, NIR calibration model which was built by using regression mathematics methods can perform better than using primary reference data.