Aiming at the problem of near-infrared spectroscopy wavelength selection, an improved team progress algorithm (iTPA) is proposed based on the team progress algorithm (TPA). The bands of molecular spectrum are arranged in descending order according to the evaluation value function obtained by modeling corresponding physical and chemical values and are divided into elite group, plain group and garbage collection group. When the new wave band selects learning behavior, if it is generated in the plain group, it needs to adjust to the direction of the elite group template; if it is generated in the elite group, its updating direction needs to be improved to adjust to the reverse direction of garbage collection group template. Unlike the elite group and the plain group, members’ evaluation value of the garbage collection group is always in a deficient state, which provides an accurate update direction for the new band generated from the elite group during the learning procedure to improve the global optimization ability of the algorithm. Through continuous iterative updating, the overall evaluation value is gradually improved, and finally, the band with the highest evaluation value is selected as the screening band. The algorithm is tested on the data set of corn starch and protein content and compared with TPA, genetic algorithm (GA), principal component analysis (PCA) and complete spectrum method. The experimental results show that the proposed algorithm can find the optimal combination of wavelengths in the whole spectrum range and explain each component’s chemical characteristics. Compared with the full spectrum, for the corn starch data set, the number of variables of iTAP was decreased from 700 to 17.55 (averaged by 50 tests), RMSEC of the model was reduced from 0.335 7 to 0.260 9, and the prediction accuracy of the correction set was improved by 22.3%. The RMSEP of the model decreased from 0.391 4 to 0.334 4, and the prediction accuracy of the prediction set increased by 14.6%; For the corn protein dataset, the number of variables decreased from 700 to 19.6 (averaged by 50 tests), RMSEC of the model was reduced from 0.147 4 to 0.101 9, and the prediction accuracy of correction set was improved by 30.1%. The RMSEP of the model decreased from 0.178 9 to 0.117 7, and the prediction accuracy of the prediction set increased by 34.2%.