A new method of near-infrared (NIR) diffuse reflectance spectroscopy is proposed to rapidly determine the degree of polymerization (DP) of natural cellulose (cotton and wood) pulp produced by a new clean pulping process. One hundred and ninety five samples were collected and their DP data were determined by standard method GB/T 9107—1999. The spectroscopy measurement method of the samples was studied and their near-infrared diffuse reflectance spectra were collected. The quantitative DP calibration models of one mixed cotton & wood and two separate cotton and wood pulps were established by partial least squares (PLS). The optimum models were developed using the spectra pretreated by derivative, autoscaling and mean-centering, and their performance is as follows: correlation coefficient of 0.980, 0.993 and 0.886, and RMSEP of 147, 143 and 53, respectively. The accuracy of NIR method was also studied. The results show that the accuracy of the two separate models of cotton and wood is better than that of the mixed model, and the precision of the two separate models is better than that of GB/T9107—1999. The identification model of cotton and wood was also established using principal component analysis (PCA). The result shows that the spectra of cotton and wood pulp have obvious difference, and the model can identify successfully the two kinds of pulp. The result indicates that the new NIR method is feasible to realize the on-line analysis of polymerization degree of natural cellulose pulp with its advantage of rapidness and easy operation.