A hyperspectral image classification method based on joint hyperpixel dimension reduction and category posterior probability optimization is proposed to address the issues of inadequate utilization of spatial information and decline in classification accuracy under small- and medium-sized training samples. First, the entropy rate hyperpixel segmentation algorithm, based on the spatial texture structure of hyperspectral images was used to adaptively detect homogeneous hyperpixel regions, and principal component analysis was applied to each region individually to mine the hyperpixel mixture features that can represent the spatial-spectral information of images. Then, the initial category probability vector of each pixel was calculated using the mixed features provided in the support vector machine, and the extended random walk algorithm was used to optimize the initial category using the image space neighborhood information. Finally, the classification result was calculated based on the maximum classification probability of each pixel. Experiments were performed on three general hyperspectral datasets, including Indian Pines, Pavia University, and Salinas, and compared with the other six methods. Even with a small number of training samples, the experimental results show that the proposed method's overall classification accuracy is 98.29%, 97.29%, and 99.72%, respectively, which is superior to the classification results of the comparison methods.