While the near infrared spectroscopy (NIRS) is used to measure the inhomogeneous samples with diffuse reflection mode, the accuracy and robustness of the calibration model is not quite good for the variation of spectrum scattering and absorption coefficient in those samples.Therefore, an establishment strategy of hybrid model based on homogeneous sample and calibration transfer method was proposed to solve the problem of detection inhomogeneous samples by NIRS.This work was focused on the tobacco leaf samples aspect.Three common calibration transfer methods, including Shenk′s patented algorithm (Shenk′s), piecewise direct standardization (PDS) and calibration transfer based on canonical correlation analysis (CTCCA), were used to construct two hybrid models of tobacco powder mixed with cut tobacco and tobacco powder mixed with tobacco flake samples to predict nicotine content in the samples of cut tobacco and tobacco flake.Experimental results showed that the hybrid model of adding some cut tobacco and tobacco flake samples to the powder model would get preferred prediction ability.Root mean square errors of cut tobacco and tobacco flake samples were reduced by 1.39% and 2.73%, respectively.This showed that the hybrid model was help for the improvement of the predicted results and the robustness of model.Moreover, CTCCA got the optimal performance between these three calibration transfer methods.Therefore, the scheme of building a hybrid model by NIRS homogeneous model and calibration transfer method to determinate the heterogeneous samples is feasible, which can accelerate the development of on-line near infrared spectroscopy technology and will provide reference for the share of near infrared spectral model.