It is vital to establish an accurate and robust quantitative model in near-infrared spectroscopy. The whole spectrum modeling can increase the computational time of modeling and forecasting, and reduce the robustness and precision. Therefore the effective variable selection method is very important for model construction. To address this problem, this paper proposed a genetic algorithm based on mutual information (GAs-MI) to select features. Mutual information filtered out a large number of unrelated information and redundant information. Genetic algorithm further selected the features with high discernment. Shapley value method was introduced to reduce the randomness of artificial setting parameters in the mutation process of genetic algorithm. In order to validate the validity of the algorithm, 273 representative tobacco samples were selected as the experimental materials. 182 samples were randomly selected to construct the PLS quantitative model of tobacco nicotine,and the remaining samples were used as the test set. The Correlation Coefficient (R), the Root Means Square Error of Cross Validation (RMSECV) and the Root Mean Square Error of Prediction (RMSEP) were used as the model evaluation indexes. The experimental results showed that the model established by the selected wavelength was simpler and more predictive.