Aiming at the influence of stray light, noise, baseline drift and other factors in the spectral signal of complex samples on the quantitative analysis results, this paper proposes a pretreatment and combined method based on near-infrared spectroscopy. First, no pretreatment and single pretreatment on 2 selected datasets is performed. Then, 9 pretreatment methods are divided into four categories according to their effects, including scale scaling, baseline correction, scattering correction and smoothing. Finally, the optimal pretreatment method of each type of pretreatment method is selected for combined research, and the optimal pretreatment method is selected according to the root mean square error cross-validation after modeling. Experimental results show that the use of appropriate pretreatment combination methods for different data sets can improve the modeling effect, while the difference between the information and complexity of different data sets will lead to different optimal pretreatment methods, and the optimal pretreatment method is related to the predicted components and the original spectra. Therefore, the best pretreatment effect can be obtained by classifying and combining the pretreatment methods according to their effects.