In the quantitative modeling of near-infrared spectroscopy data, the high redundancy and high noise of the data severely affect the robustness and accuracy of the modeling. Therefore, this paper presents a feature-based spectroscopy combined with improved Particle Swarm Optimization (PSO) Method of choosing. First, we measure the importance score of each feature through mutual information, and then sort the features according to the importance of the features in descending order. This effectively avoids the problem of losing important information caused by using the principal component reduction method. Secondly, the concept of jump degree is introduced and a method of feature stratification is constructed. Similar features of similar importance are merged into the same feature subset, and the descending ordered feature set is segmented into different feature subsets, avoiding the screening uncertainty caused by artificially setting the score of feature importance score during feature process. Finally, the particle swarm optimization algorithm with fast convergence rate and few control parameters is used as the optimal feature subset optimization method. At the same time, particle swarm optimization is improved in two aspects: The chaotic model is introduced to increase the diversity of the population and improve the global searching ability of PSO, so as to avoid getting into local optimum. The number of features is introduced into the fitness function, and the influence of the number of features on the fitness function is adjusted by the penalty factor in the early iteration to improve the adaptability of the algorithm. The stratified data is collected as a feature subset and then added as a modified particle swarm optimization algorithm to select the high-resolution feature subset. In this paper, the nicotine index as an example of the feature selection process is described, using Nicolet company Antaris II near infrared spectrometer near infrared spectrum data acquisition, spectrum scanning range is 4 000~10 000 cm-1. First, we use the mutual information theory to calculate the importance score of 1 557 features of the whole spectrum on the quantitative modeling of the index to be measured, and take the average of 30 experiments. Secondly, all the features are sorted in descending order of importance scores to calculate the jumping degree of all the features. According to the jumping degree, the critical points of the feature stratification are searched, and the features are divided into different feature layers to construct a feature containing 8 feature subsets set S={S′1, S′2, S′3, S′4, S′5, S′6, S′7, S′8}. Then, the feature subset is in turn {S′1}, {S′1, S′2}, {S′1, S′2, S′3}, …, {S′1, S′2, S′3, S′4, S′5, S′6, S′7, S′8} as a candidate for initial particle swarm. With R/(1+RMSEP) as the evaluation criteria of the pros and cons of feature subsets, each iterative experiment 50 times, the ratio of the largest feature subset is the optimal feature subset. In order to verify the effectiveness of this algorithm, we select representative tobacco near-infrared spectral data as a training set and a test set, establish a PLS quantitative model of nicotine and total sugar, and compare with the full-spectrum, stratified characteristic spectrum, particle swarm algorithm selected by the characteristic spectra. The simulation results show that the modeling correlation coefficients R of nicotine and total sugar selected by this algorithm are respectively 0.988 5 and 0.982 2, RMSECV of mutual verification are 0.098 4 and 0.889 3 respectively, RMSEP of prediction root mean square error are 0.901 6 and 0.100 7 respectively, Accuracy are significantly higher than the other three methods. From the selected number of features, the proposed algorithm has the least number of selected features, effectively eliminating the weak correlation and noise and redundant information in the original feature set, minimizing the number of main factors of the model and reducing the complexity of the model, and the model is steadier, more adaptable.