An artificial neural network of matrix back propagation(MBP-ANN) combined with principal component analysis(PCA) for near infrared spectroscopy (NIR) quantitative analysis method is presented, and its principles is analyzed. A PCA-MBP-NIR quantitative analysis software system is developed based on object oriented programming technology in environment of Microsoft Visual C++. The PCA-MBP-NIR model and partial least square(PLS) NIR model are built between the moisture and raw spectrum of 40 wheat samples, and the two models are also built for noise spectrum (rmax=14dB) in the same way. The moisture of 10 unknown wheat samples are predicted by this model. Results show that, using PCA-MBP-NIR method instead of PLS-NIR for noise spectrum, the correlation coefficient of predicted values and standard values of unknown samples can be increased, and the root mean square deviation (RMSD) can be decreased.