ABSTRACT Accurate and cost-effective monitoring of river stage and discharge is essential for effective water resource management, yet conventional gaging stations are costly to install and maintain, particularly in remote areas. This study presents an integrated hardware–software system for camera-based hydrologic monitoring that combines deep learning image segmentation, ensemble regression, and edge computing for near real-time observations. Two hydrologically distinct sites in northern Utah were monitored using fixed cameras to collect imagery for model development and validation. The workflow segments water surfaces, extracts stage-dependent features, and predicts water levels using a stacked ensemble of support vector regression, random forest, and XGBoost models. Stage estimates achieved R2 values of 0.93–0.98 (MAE = 1.03–2.33 cm), and discharge estimates obtained through site-specific rating curves reached R2 = 0.84–0.99 (MAE = 0.27–0.30 m3/s). The pipeline was deployed on four edge devices, Raspberry Pi 5, Jetson Nano, Jetson Orin Nano, and LattePanda Sigma, to assess operational feasibility. Results showed that lightweight models such as MobileSAM enable efficient performance on low-power devices, while high-end hardware supports near real-time inference. The proposed approach offers a scalable, low-maintenance alternative to traditional gaging, advancing camera-based hydrologic monitoring in both developed and data-sparse regions.
Issa et al. (Wed,) studied this question.