Conference Agenda

In Taiwan, river water levels are often monitored manually or by camera-based systems that rely on physical staff gauges and human interpretation. However, such methods can be limited by environmental factors (e.g., inaccessible sites without a gauge, or gauges that become obscured or soiled). To address these limitations, this study proposes an automated river water level monitoring system that combines virtual gauges and a deep learning CNN model (HRNet w18). First, an orthorectification process is used to generate a virtual gauge within the image frame, eliminating the need for on-site installation of a physical gauge. Next, the HRNet w18 model accurately detects the water surface, mitigating errors from debris or gauge contamination. Both real-world river images and laboratory flume simulations were used to train and validate the model, ensuring robust performance under various flow and wave conditions. By capturing a series of images over time and calculating an averaged water level, this approach accounts for natural wave fluctuations and can be adapted to different river scales. Experimental results show that this method significantly reduces costs and labor while providing real-time, accurate water level data for disaster prevention and water resource management.