The realization of automatic feeding of bait has always been the focus and difficulty of automatic aquaculture. Recognition of the fish feeding intensity can provide a reference for accurate feeding. At present, many laboratories have researches on the fish feeding intensity, but most of the researches on the fish feeding intensity are based on circulating farming ponds or self-made fish tanks, which are not suitable for open farming ponds. Aiming at the actual environmental background and difficulties, the water observation method was used to build a data acquisition system and produce a video data set of fish feeding intensity. Then a fish school feeding intensity classification model was proposed based on improved long-term recurrent convolutional networks（LRCN）, which embeded the attention mechanism squeeze-and-excitation block （SE-Block） into the convolutional neural networks. The SE-CNN networks was used to extract the features of the video frames, then input the features into the doublelayer gate recurrent unit networks. Finally, the video classification results were obtained through the fully connected classification layer. At the end, the proposed SE-LRCN model realized the intensity three classification of the fish school feeding intensity video. The test results showed that the classification accuracy of the proposed model reached 97%, and the F1 score reached 94.8%. Compared with the long-term recurrent convolutional networks before the improvement, the accuracy was increased by 12 percentage points, and the F1 score was increased by 12.4 percentage points. The research model can more finely recognize the fish feeding intensity, and provide a reference for automatic accurate feeding.