Application of Deep Learning Techniques in Water Level Measurement: Combining Improved SegFormer-UNet Model with Virtual Water Gauge

Abstract

Most computer vision algorithms for water level measurement rely on a physical water gauge in the image, which can pose challenges when the gauge is partially or fully obscured. To overcome this issue, we propose a novel method that combines semantic segmentation with a virtual water gauge. Initially, we compute the perspective transformation matrix between the pixel coordinate system and the virtual water gauge coordinate system based on the projection relationship. We then use an improved SegFormer-UNet segmentation network to accurately segment the water body and background in the image, and determine the water level line based on their boundaries. Finally, we transform the water level line from the pixel coordinate system to the virtual gauge coordinate system using the perspective transformation matrix to obtain the final water level value. Experimental results show that the improved SegFormer-UNet segmentation network achieves an average pixel accuracy of 99.10% and an Intersection Over Union of 98.34%. Field tests confirm that the proposed method can accurately measure the water level with an error of less than 1 cm, meeting the practical application requirements.