To address the problem of low accuracy and poor reliability of water level monitoring caused by water gauges being easily entangled by floating debris and damaged during flood periods under shore-based camera deployment, a mapping relationship between the pixel’s vertical coordinate of the waterline in an image and the actual water level was established based on monocular photogrammetry. The pyramid scene parsing network was employed to perform semantic segmentation of water and riverbank regions. Meanwhile, cross-sectional terrain data were combined to construct a virtual water gauge, through which the pixel’s vertical coordinate of the waterline was converted into the actual water level. Experimental results show that the root mean square error of waterline detection is less than 2 pixels. By properly selecting cameras to control the spatial resolution of imaging elevation, the accuracy of water level measurement can be better than 2 cm. With inclined shore-based camera deployment, the proposed method achieves fully non-contact online water level monitoring without requiring a physical water gauge or ground control points. It has the advantages of low operation and maintenance cost, strong robustness, and intuitive results and is applicable to irregular cross-sectional scenarios such as natural rivers and artificial open channels.