Data from hyperspectral remote sensing have provided detailed spectral and spatial information regarding ground objects. To solve the problems of low robustness and classification accuracy due to the underutilization of spatial information of hyperspectral data in previous classification methods, this paper proposes a classification method based on improved superpixel segmentation and 3D convolution neural network. First, the hyperspectral remote sensing data are segmented via superpixel segmentation and fuzzy clustering; then, the spatial-spectral joint data formed by the regional segmentation results and hyperspectral data are trained and classified using a 3D convolution neural network. The proposed method improves the role of spatial information in classification by dividing and fusing spatial regions, reduces the impact of the phenomenon of “same objects different spectra” on classification, and introduces a 3D convolution neural network to train and classify the spatial-spectral joint data, improving hyperspectral classification accuracy. In the Pavia University and Salinas datasets, the proposed method has an overall accuracy of 97.53% and 98.48%, respectively. When compared with the control experiments, the proposed method exhibits a better classification effect, which proves its efficacy.