Improving image-based water-level monitoring by coupling water-line detection techniques and the segment anything model

Abstract

Image-based water level measurements offer cost-effective alternatives to traditional methods but often face challenges from environmental disturbances. We present a novel gauge image segmentation method integrating water-line detection techniques with the Segment Anything Model (adaptive-SAM) for continuous water level monitoring. The method uses adaptive prompt points derived from water line detection to generate high-quality masks of staff gauges and retrieve real-time water level data. Experiments conducted at an urban lake in southern China demonstrated the effectiveness of this integration, with the mean error reduced to 0.79 cm. The adaptive locating strategy facilitates the generation of accurate prompt points to guide SAM, thus significantly enhances its resilience to complex environmental disturbances (i.e., water texture, darkness, reflection, and overexposure) and achieves a substantial reduction in maximum errors by 22.2–41.1 cm. This simple yet robust approach provides an accessible tool for non-professionals, potentially increasing hydrological data density when traditional sensors fail.