Nearshore Bathymetric Inversion and Uncertainty Estimation from Synthetic Imagery using a 2D Fully Convolutional Neural Network

Abstract:

Bathymetry in the surf-zone is constantly changing due to wave interactions with the bottom boundary which drive sediment transport. Large wave events, such as storms, not only complicate the collection of in-situ bathymetry but can also cause significant changes in the bottom boundary on very short timescales. Remote sensing of wave speed and breaking using radar or optical imagery can be used to provide increased observations of bathymetry from UAS, towers, and satellites. Physics-based inversion algorithms are typically used to invert speed or dissipation to estimate depth. However, when wave heights are large, non-linearities in the surf-zone begin to dominate the hydrodynamics making the computationally efficient linear depth inversion less capable of accurately resolving bathymetry. We explore using a 2-dimensional fully convolutional neural network to directly estimate depth from high-resolution synthetic time-averaged (timex), time-series (video), and single-frame (snapshot) images of the surf-zone. Results show comparable errors to physics-based methods, while also retaining accuracy at higher wave heights. A spatial estimate of model uncertainty over the domain is quantified using a combination of MC dropout, infer-transformation and infer-noise. Standard deviations of the resulting ensemble of predictions are used to aid in a spatial assessment of the reliability of the bathymetric prediction.