In order to solve the problem of low classification accuracy of hyperspectral images in the case of limited training samples, we proposed a tandem three-dimensional(3D)-two-dimensional(2D) convolutional neural network model combining dilated convolution with attention mechanism. First, with the tandem 3D-2D convolutional neural network as the basic structure, the model used 3D convolution to simultaneously extract the spatial-spectral features of hyperspectral images and 2D convolution to further extract high-level spatial semantic information. Then, by introducing dilated convolution to enlarge the receptive field of the convolution kernel, we constructed a multi-scale feature extraction structure for the fusion of multi-scale features. Finally, the attention mechanism was applied to make the network pay attention to important spatial-spectral features and suppress noise and redundant information. Furthermore, we performed a comparative experiment between the proposed algorithm and four deep-learning-based algorithms on two common data sets. The results show that the proposed model achieves the most accurate classification results and effectively improves the classification accuracy under the condition of limited training samples.