A Deep Learning Algorithmic Approach to Develop a Video Inundation Monitoring System
Jon Derek Loftis, Sridhar Katragadda · OCEANS 2022, Hampton Roads · 2022
Technological innovation coupling passive remote sensors with recent advances in edge detection and machine learning algorithms has made the development of a Video Inundation Monitoring Systems effectively feasible. To accomplish this, the USGS Next Generation Water Observing System (NGWOS) has funded the VA-WV Water Science Center (WSC) to partner with the Virginia Institute for Marine Science {VIMS), and the City of Virginia Beach to evaluate hardware and software models. This project sufficiently field-tested 3 new fixed-mounted, deep learning, Internet of Things (IoT) video cameras in tidal tributaries of Chesapeake Bay and Delaware Bay. Video cameras were purchased and installed to capture still images, continuously collected every 15 minutes for three months in the summer of 2020. Statistical analysis revealed the sensors to be effectively capable of continuous determination of surface water levels with an RMSE deviation of <0.5 in. when positioned within $\sim 30 \mathrm{ft}$. of the desired monitoring target area, with error on accuracy scaling inversely, proportional to the distance from the monitored target. USGS A-style gage staves were mounted relative to NAVD 88 within the camera’s view, and video-recorded and data-translated water levels were verified using USGS Ka-band radar active remote sensors at each monitoring site to add an additional layer of field verification for this new technology. Multiple camera hardware models were tested, each with 4 k resolution to afford sufficient pixel density for effective error thresholding, and infrared imaging for nighttime capture with a field-verified effective range of <30 m. This project ultimately refined and demonstrated capabilities of a new passive remote sensing technology capable of detecting and extracting and water levels, from time-lapsed images, and fully integrated the results into NWISweb for internal and public access in near real-time (<5 minutes from collection).