Aiming at the difficulty in image classification due to the high similarity between different flow speeds, a conditional boundary equilibrium generative adversarial network and a convolutional classification network based on the multi-feature fusion were proposed to realize the generation and the classification of flow velocity images, respectively. A labeling mechanism and a verification module were introduced to realize the fitting and generation of corresponding category images, in order to achieve data enhancement. To enhance the impact of different texture features on velocity estimation, a multi-feature fusion layer was introduced to realize the feature extraction and the flow velocity recognition so as to realize the classification for images with small differences. The proposed method was applied to the actual river surface velocity estimation. Results demonstrate that the added tag information and the verification module can guide the data generation of corresponding class to a certain extent in the image generation module. Compared with other methods, the multi-feature fusion mechanism makes the proposed classifier more robust in identifying flow velocity images with small differences.