In order to solve the conflict between noise suppression and detail signal reservation of Near Infrared (NIR) spectroscopy, a wavelet domain noise reduction method is proposed based on the noise variance estimation. The method establishes two states Gaussian Mixture Models (GMM) of high-frequency coefficients in wavelet domain, uses Expectation Maximum (EM) algorithm to figure out the factor of model, proves that the model can accurately estimate noise variance, and puts the model to wavelet threshold noise suppression. By establishing the wine alcohol Partial Least Squares (PLS) prediction model of NIR spectroscopy, and comparing Penalty threshold, Brige-Massart threshold and the default threshold wavelet threshold denoising effect, the experiment validates that the method has better noise reduction effect than other normal method, which improves the stability of NIR model.