Near infrared (NIR) spectroscopy is an ideal analytical method for rapid and nondestrctive measurement of the properties of agriculture products. The efficient use of this method is dependent on multivariate calibration methods determined by the sensitivity to variations. However, fluctuation of background and noise are unavoidable during collecting spectra, which will not only worsen the precision of prediction, but also complicate the multivariate models. Therefore, the first step of a multivariate calibration based on NIR spectra data is often to preprocess the data for the purpose of removing the varying background and noise. In this study, wavelet transform (WT) was used to eliminate the varying background and noise simultaneously in the near infrared spectroscopy signals of 55 navel oranges. Three families of mother wavelets (Symlets, Daubechies and Coiflet), four threshold selection rules (Rigrsure, Heursure, Minimaxi, Fixed form threshold), and two threshold functions (soft and hard) were applied to estimate the performances. The sugar content of intact navel orange was calculated by partial least squares regression (PLSR) with the reconstructed spectra after denoised. The results show that the best denoising performance was reached via the combination of Daubechies 5, "Fixed form" threshold selection rule, and hard threshold function. Based on the optimization parameter, wavelet regression models on sugar content in navel orange were also developed and resulted in a smaller prediction error than a traditional PLSR model.