Water level monitoring based on water gauge is widely used due to its cheap and portable features. However,it is still a tricky problem to accurately detect the water level from water gauge images in complex real-world hydrological scenarios. In this paper, we propose a composite method for water level measurement.Being different from the traditional methods, this method can detect the location and quantity of water level in the complex and changing scenes, and then accurately segment the water level line from the changing water level, and finally get the accurate water level value. Firstly, we improve the FCOS model by the context adjustment module to meet the requirements of edge computing and ensure a relatively high detection accuracy. Secondly, to modularize the semantic relationships, we apply the contextual adjustment module of Deeplab-v3 to segment the water gauge image above its water surface. Then,the segmented region can be used to calculate the position of the water level line. Finally, we combine all the results to calculate the water level value of the water gauge image. The experimental results demonstrate that the method solves the problem of calculating the water level of the water gauge in complex hydrological scenarios. In addition, the water level error calculated by the method has been accurate to 1 cm, which is much better than the existing methods.