Optical nearshore wave gauging with deep neural networks

Daniel Buscombe, Northern Arizona University, Geoscience Division, School of Earth and Sustainability, Flagstaff, AZ, United States, Shawn R Harrison, U.S. Naval Research Laboratory, Stennis Space Center, MS, United States and Jonathan A Warrick, U.S. Geological Survey, Santa Cruz, CA, United States
Abstract:
Video or static imagery from close-range camera systems have previously been used to measure coastal hydrodynamics, including wave period, direction, and celerity; breaking wave location and intensity; wave dissipation; alongshore currents; rip currents; and wave runup. However, techniques are not always robust to noise, and they don’t always transfer/scale well between sites. Further, we know of no reliable technique for estimating wave height from a single image. Stereo imaging can achieve this using two or more images, but the procedure is computationally intensive to estimate height from each image, sensitive to image noise, and requires information about camera geometry to scale and relate the observations to geographical position.

We describe a new technique called 'optical wave gauging' or OWG that is designed to be configurable for estimating various quantities of nearshore waves, from single or multiple images taken at arbitrary viewpoints, with no information about camera geometry or image coordinate transformations required. OWG is currently designed for basic monitoring of coastal hydrodynamic quantities such as wave height, period, direction, tidal elevation, etc.

OWG is a model framework based on deep convolutional neural networks that can be purposed differently for each task; single image-single output (such as wave height or period); multiple images-single outputs; or multiple images-multiple outputs. We demonstrate the OWG model framework by configuring it to estimate time-series of various wave height and period quantities from a two year time-series of static images from one or two cameras from a USGS coastal video monitoring system at Sunset State Beach, California, USA. We examine model and model training sensitivities and discuss requirements for adapting the technique for operational nearshore wave monitoring from cameras alone. Finally, we discuss the feasibility of ‘real-time’ monitoring using single-board computers in the field.