Monitoring Water Level Using River Cameras: Integrating Domain-Specific Models and General AI For Transferable Water Segmentation

Water level data is critical for hydrologic model calibration. The extensive river camera networks, in conjunction with advanced deep learning techniques, form the foundation for the imaging-based monitoring of water level trends. However, limited annotated data and tedious local deployment restricts the applicability of deep learning models in new river scenarios. This study proposes a novel transferable deep learning framework by combining General AI with domain-specific models for water segmentation, and uses the static observer flooding index (SOFI) as the proxy for water level variations. The framework uses the Segment Anything Model (SAM), a generic computer vision model by Meta AI, for segmenting images into discrete while semantically unknown objects. A ResUnet model pre-trained on a non-local dataset simultaneously identifies pixels with the highest probability of being water, which are then overlaid onto the segmented images to specify the water object. The framework was applied to image sequences acquired from river cameras stationed at four locations in Tewkesbury, UK, for water segmentation and water level trend monitoring. The SOFI time series were calculated based on the segmented masks and underwent data quality control using an unsupervised clustering method. The obtained SOFI signal showed an average correlation of 0.83 with real water level fluctuations, significantly surpassing the single ResUnet model's correlation of 0.54. The data provided by the framework was qualified for hydrologic model calibration referring to both error magnitudes and distribution patterns. Our study has thus moved toward an ease-of-use implementation of river cameras for transferable water level trend monitoring.