How to use general AI for task-specific applications: A case study of monitoring water level trends with river cameras

Abstract

Water level variations influence geochemical and hydrological processes within river networks. Water segmentation from river camera images using deep learning supports water level trend monitoring, but domain-specific model accuracies are constrained by limited annotated data. To improve accuracy, this study proposes a framework combining domain-specific models with General AI. The framework uses the Segment Anything Model (SAM) as backbone, with a pre-trained ResUnet model identifying highest-probability water pixels as a prompt to SAM, without any requirement for human intervention or local annotation. Applied to river camera images from Tewkesbury, UK, the framework increased the Intersection over Union (IoU) by over 15 % compared to the single ResUnet. Point prompt was identified as the optimal mode for feeding water-related prior knowledge to SAM. The static observer flooding index derived from segmented masks showed a strong correlation (0.90) with real water level, surpassing the ResUnet's 0.54. Our framework allows for the supplementation of river monitoring networks with camera gauges, providing robust water level trend observations.