Abstract: Shoreline control of river and lake waters is an important part of the river and lake chief system.Since starting the fishing ban in the Yangtze River, illegal fishing still occurs many times along the Yangtze River shoreline, so applying remote sensing methods such as satellite-UAV-ground video surveillance to jointly identify shoreline illegal behaviors has become a trend.In order to rapidly and intelligently achieve accurate detection of fishing behavior in prohibited fishing, an improved deep learning algorithm YOLOv3-CBAM based on transfer learning and attention mechanism is presented.Firstly, aiming at the problem of limited sample size of fishing targets, a pre-training weight with strong feature extraction ability is developed by using COCO dataset.Secondly, in order to enhance the detection effect of small targets, through the addition of multiple attention modules to YOLOv3 network, an improved YOLOv3-CBAM detection model is proposed.The experimental results show that as the YOLOv3 algorithm adopts the training strategy of transfer learning, it can accelerate the convergence speed of the model and improve the recognition accuracy of the model from 78.57% to 93.27%.By adopting transfer learning strategies and multiple attention modules, the recognition accuracy of YOLOv3-CBAM is 93.99%,which doesn't need to add additional parameters.This study can provide technical support for the real-time dynamic supervision of fishing prohibition in the Yangtze River Basin.