For removing the noise in the spectral data, Matlab 7.0 wavelet toolbox was used to denoise the data. In order to get the best denoising effect, the model evaluation parameters were investigated from the db2 to db9 wavelet respectively in dbN wavelet basis, and the db6 was thought to be the best wavelet according to the investigation result. Using the db6 wavelet, the optimal decomposition layer was five when changed from three to seven. For getting the best denoising method, the signal-to-noise ratio (SNR) and root mean square error (RMSE) were used to evaluate denoising effect in different thresholds, and the hard threshold value of the heuristic denoising method was observed to be the best one. The prediction model was built using the reconstruction spectrum by partial least squares (PLS) method. The correlation coefficient of the proposed model was 0.907. The root mean square error values of calibration and prediction were 0.819 and 0.905. The performance index was 88.3%. The model parameters using wavelet denoising were better than the original signal. It was showed that wavelet denoising was feasible in prediction of tomato vitamin C with NIR spectroscopy.