Search for a command to run...
This study presents a novel approach integrating near-field passive remote sensing technologies with advanced machine learning algorithms and edge detection through image segmentation to free surface water levels from live oblique imagery in tidal environments. The sensor is capable of full color monitoring by the optical sensor during daylight hours, which affords greater pixel density for the machine learning model and edge detection, while nighttime monitoring relies on the more limited range and relatively monochromatic color profile of the infrared projection sensor, leading to diminished accuracy at night. Cameras were strategically deployed at twelve sites across the United States, capturing images at six-minute intervals over a three-month period. The final machine learning model was trained on a week's worth of images taken every 6 minutes, translating to approximately 1,700 images at each of the twelve sites. This was done to train a model to run over the following 3 months to accurately predict 80% of the six-minute water levels within 1.5 cm, and 97% of the water levels within 10 cm accuracy. Statistical analyses demonstrated that the system reliably produced continuous surface water level measurements, with an aggregate root mean square error of less than 1 cm when cameras were positioned within approximately 10 meters of the target observation area. To verify vertical measurement accuracy, USGS A-style staff gauges were installed within each camera's field of view. Water levels inferred from imagery were cross-validated with nearby radar sensors at each location. A generalized machine learning model that doesn't fully rely on site-specific flooding imagery has been in development, but there are many factors that convolute its universal effectiveness. The resulting data machine learning model products are in the process of being integrated into the USGS Water Data System, with the goal of enabling both internal and public access to real-time water level estimates derived from the machine learning water level estimation framework.