| Sumario: | Accurate, automated monitoring of river gauge plates is critical for hydrological analysis and effective water resource management. Traditional monitoring methods, relying on pressure probes and expensive cabling, often incur high installation and maintenance costs and typically require calibration at intervals of up to two weeks. Imagery-based solutions can significantly reduce these costs and provide continuous, verifiable field data, enabling more frequent validation. This study introduces a novel, two-stage approach that combines real-time object detection with generative AI to extract gauge readings and compute corresponding river discharges. In the first stage, a YOLOv8 model localizes gauge plates and waterline regions within field images. In the second stage, a dedicated YOLOv8 network performs key-point detection to identify scale-gap markers along the gauge face. These detections are then ingested by a Gemini-powered generative AI pipeline via task-specific prompting to translate image features into precise numeric readings. Subsequent conversion of extracted gauge readings into volumetric discharge leverages established rating curves, enabling end-to-end automation of river flow estimation. This study evaluated model performance on three image subsets: the full dataset, a good–moderate quality subset (≈91% of samples), and a challenging “worst” subset (≈9%, including blurred, corroded, or occluded scales). The coefficient of determination (R²) for these categories was 0.552, 0.816, and 0.061, respectively, indicating robust predictive capability in moderate-to-high quality imagery with diminished accuracy under severe visual degradation. This hybrid AI solution offers scalable, real-time water level and discharge monitoring with minimal human intervention. Its resilience to varying image quality and seamless integration of object detection and generative reasoning position it as a promising tool for automated hydrological monitoring.
|
|---|