<p>Water level detection is one of the most important indicators of the management level of hydropower plants and has received much attention in academia and industry. Existing methods mainly focus on small areas and ignore large-scale variations in water levels (turbulent fluctuations) in practical scenarios. To address this problem, we propose a large-scale water level detection framework based on Unet image segmentation and local binarization cooperative learning. The proposed framework is divided into two modules: (1) a binarized morphological learning module based on Unet spatial location features, and (2) local dual-channel attention learning from the feature segmentation level. The former learns the water level line segmentation interface by binarizing the feature map. Meanwhile, the latter enhances the local contour information of the water level detection image based on high-level semantic information. Experimental results show that the average accuracy of the proposed model is 86.7%. The accuracy of water level detection exceeds 95.77%, and errors are less than 1.0 cm to manual reading, which exceeds the accuracy of existing water level detection algorithms, thereby satisfying the requirements of practical applications.</p>