The high-dimension, high-redundancy, high-noise and nonlinear characteristics of near-infrared spectroscopy data seriously affect the accuracy of spectral similarity measurement. Aiming at this problem, a similarity measurement method of the near infrared spectrum based on the grid division local linear embedding (GGLLE) algorithm is proposed. First, the high-dimensional spectral data is divided into multiple grid subspaces according to the expression of key chemical components in the spectrum. Second, two aspects for the local linear embedding (LLE) algorithm are improved, and the improved LLE algorithm is used to sequentially map the feature of each subspace from high- to low-dimensional space and calculate the similarity matrix of the generated subspace. Finally, the subspace similarity matrix is normalized, and the similarity matrix of the accumulated and generated spectral sample set is to be solved to realize a similarity measurement of the spectrum. Two sets of tobacco leaf spectral data provided by a tobacco company are selected to construct a model of the spectral similarity measurement. The accuracy of the similarity measurement is a criterion of the pros and cons of the algorithm. The experimental results show that the accuracy of the similarity measurement model constructed by the GGLLE algorithm is 93.3%, which is obviously better than the accuracies achieved by principal component analysis, stacked auto encoders, and LLE algorithms, which are 64.2%, 67.5%, and 82.5%, respectively. Thus, the effectiveness of the GGLLE algorithm is proved.