The reflectance spectra of 189 samples within 450~2500nm were collected. Firstly, the spectra were denoised using the orthogonal wavelet functions sym8 (level was 3). And then the spectra variables were compressed to 14 variables by successive projections algorithm (SPA). The MLR model with 14 variables as inputs could result in that prediction correlation coefficient was 0.8855 and prediction root mean square error was 0.5111. Then the group of wavelengths derived from SPA was screened by their contributions to the total sugar content. After screening on contribution, the number of wavelength variables dropped to 11. Finally, the MLR, PLS and BPNN calibration models were built with 11 wavelength variables as inputs and compared. The results demonstrated that wavelength variables were decreased to 0.54 % of the original variables by SPA and screening on contribution, and this could help to make the model more concise and robust. 