The feature selection algorithm based on the combination of random forest and game theory was proposeed in this paper as noise and redundant information in the near infrared spectroscopy would lead to the low recognition rate of a model.This algorithm was first used to measure the feature significance according to the random forest and select some features related to classification,then compute the weights of selected characters by using the improved Shapley values and mutual information computed to remove redundant information from the weighted feature set and get the optimal feature subset.To validate effectiveness of this algorithm,the tobacco leaf production area identification model was established.The experimental results indicated that the algorithm proposed in this paper had a good recognition on the area of tobacco leaf production with a recognition rate of 95.88%.