Hyperspectral images have hundreds of continuous and narrow spectral bands, spanning visible light to infrared light. They can provide fine spectral properties of ground objects and have important application value for recognizing and classifying ground objects’ materials and attributes. It is of great significance to select limited spectral bands for transmission and processing of interested targets, improving the timeliness of hyperspectral data processing and designing practical spectrometers for specific applications. Selecting the optimal band combined with the target features becomes an inevitable requirement to improve the processing efficiency and ensure the accuracy of target recognition or classification. Therefore, selecting the band subset with better classification and recognition ability from hundreds of hyperspectral images is an urgent problem to be solved. This paper proposes a hyperspectral band selection method based on the improved particle swarm optimization algorithm. This method is different from the traditional particle swarm optimization algorithm by introducing the “probability jump characteristic” and setting the elimination mechanism of the new solution to eliminate the “stagnation” new solution, which improves the global optimization performance of the algorithm. Then, based on the spectral characteristics of the target, the objective optimization function of optimal band selection is adopted. The improved particle swarm optimization algorithm is used to solve the objective function, and the selected band subset is fed back to the support vector machine (SVM) for classification application. In this paper, two standard hyperspectral datasets (Indian pines, The experimental results show that the proposed method has higher classification accuracy than the existing methods. Among the several methods, the traditional particle swarm optimization algorithm has the worst effect; the waveband selected by the proposed algorithm has the best classification accuracy, and the classification accuracy of the two data sets can reach 98.141 4% and 99.084 8%, respectively.