The original Near Infrared (NIR) spectral data inevitably contain a large amount of noise signals and data. So, before spectral analysis, it is necessary to preprocess the spectral data. The preprocessing of NIR spectral data mainly includes two tasks. One task is to de-noise so as to improve the robustness of the model and the accuracy of the prediction result. Another task is to compress the data for storage and modeling speed improvement. Traditional spectral preprocessing methods have their own limitations in these two aspects. The wavelet analysis method is used to preprocess the NIR data of apples. The Peak Signal to Noise Ratio (PSNR) and Normalized Correlation (NC) coefficient are selected as the evaluation index. Compared with the common Savitzky-Golay smoothing and multiple scattering correction, the wavelet method not only can compress spectral data effectively, but also has its superiority in noise removal and spectral detail preserving.