Abstract: Aiming at the problem of low training sample size leading to low classification accuracy of hyperspectral images, this paper proposes a joint sparse representation hyperspectral image classification method based on dictionary optimization. First, the band selection method based on hierarchical clustering is adopted to reduce the dimensionality of hyperspectral image data; second, the hyperspectral data is divided into multiple subsets based on spatial information, and training samples with known label information are used to mark each subset that may become training samples. Pixels form a candidate set of training samples, and the candidate set is filtered according to the spectral similarity criterion to obtain an optimized dictionary; finally, the optimized dictionary is used to classify hyperspectral images through joint sparse representation. The simulation experiments of Indian Pines dataset and Pavia University dataset show that the classification algorithm proposed in this paper can effectively improve the classification accuracy of hyperspectral images.