Feeding as a key part of the aquaculture process, the amount of bait fed directly affects the quality of aquatic products and the cost of aquaculture. However, the current feeding methods include manual feeding and machine feeding at regular intervals, which mostly rely on manual experience and are difficult to achieve accurate feeding. Different satiation levels of fish were identified based on the improved ResNet34, which was important for achieving accurate control of bait feeding in the future. A dataset- containing five different satiation levels was created based on the feeding behaviors exhibited by fish at different satiation stages, and the images were pre-processed using data enhancement operations. Secondly, based on the original model ResNet34, the use of coordinate attention mechanism wasproposed to enable the model to focus on a large area in the process of feature extraction of images. And the depth-separable convolution was used instead of the traditional convolution to reduce the number of model parameters. To evaluate the effectiveness of the improvements, the performance of the improved model wasanalyzed on the fish satiation dataset and compared it with the original model ResNet34, AlexNet,VGG16, MobileNet-v2, GoogLeNet and other classical convolutional neural network architectures. The comprehensive experimental results showed that the model reduced the amount of parameters by 46.7% and achieved an accuracy of 93.4% compared with the original model, which had a 3.4 percentage points improvement compared with the original model, and the improved model also outperformed other convolutional neural networks in terms of accuracy, precision, recall, and F1 score. In summary, the model achieved a good balance between performance and number of participants, which provided the possibility for subsequent models to be deployed in real farming environments and guide farmers in improving and developing feeding strategies.