Herein， a feature extraction method based on fractional differentiation is proposed for the feature extraction and classification of hyperspectral images. Two-dimensional （2D） fractional differential masks are designed to extract the pixel spatial fractional differential （SpaFD） feature of hyperspectral images， and a spectral–spatial joint criterion is proposed to select the differential mask order. To entirely utilize the spatial and spectral features of hyperspectral images， the SpaFD feature is fused with the original feature via a direct connection to obtain a mixed feature （SpaFD-Spe-Spa）. The effectiveness of the SpaFD-Spe-Spa feature is verified on a 3D convolutional neural network （3DCNN）， 3DCNN after pixel spectrum dimensionality reduction using principal component analysis （3DCNNPCA）， and hybrid spectral network （HybridSN）. In the experiment， masks with sizes of 3×3， 5×5， and 7×7 are used to perform feature extraction. Experiments on four real hyperspectral image datasets reveal that the extracted SpaFD and SpaFD-Spe-Spa features are effective in hyperspectral image classification， and the SpaFD-Spe-Spa feature significantly improves classification accuracy. When compared with the original features in the Indian Pines， Botswana， Pavia University， and Salinas datasets， the classification accuracy of the SpaFD feature is improved by 3.87%， 1.42%， 2.41%， and 2.87%， respectively， whereas that of the SpaFD-Spe-Spa feature is improved by 3.90%， 5.62%， 3.35%， and 5.18%， respectively， under optimal conditions.