Abstract: In this paper, locally linear embedding algorithm in manifold learning based on kernel transformation and the geodesic distance was proposed for judging the quality of tobacco leaf similarity in high-dimensional data space. This method was verified through feature analysis and similarity measure experiment of 450 tobacco grilled piece samples. The results showed that local linear embedded method based on geodesics distance had very good characteristic of the sample classification ability and the applicability of field data. PCA method could reflect the inherent nonlinear characteristics of data quality of raw material, but there existed the more overlap of sample points. In the similarity measurement, the searching tobacco numbers through this method in the same producing area, the same position and the similar grade were greater than the number of tobacco leaf in original data set and that of PCA transform. The method can effectively solve the isometric problem in low-dimensional space to similarity measure.