Abstract: When the near-infrared diffuse reflection spectroscopy is used in detecting the quality of intact tobacco leaves, there is no guarantee of spectral representativeness and repeatability because of the large area of a tobacco leaf and the inhomogeneous distribution of chemical compositions in it. In order to solve this problem, a systematic study of sampling modes for intact tobacco leaf had been carried out. The spectral information divergence combined with gradient angle method (SID-GA) was applied to calculate the intact leaf spectrum similarity of seven sampling modes and the sum of ranking differences (SRD) was employed to evaluate the partial least squares (PLS) quantitative models of total sugar, reducing sugar, nicotine, total nitrogen, potassium and chlorine. The results showed that the most suitable sampling modes were "8-point mode", "6-point mode" and "8-point mode" for B2F, C3F and X2F tobacco leaves, respectively. Taking the number of spectral scanning points and performance of PLS models into account, the "8-point mode" can be deemed the most suitable for spectral collection of intact leaf and "6-point mode" takes the second place. It can be concluded that the selected sampling mode with fewer points, higher representative and better repeatability for intact leaf has practical significance for improving predicting efficiency and accuracy of tobacco quality.