Super-resolution for Ocean Bathymetric Maps Using Deep Learning Approaches : A Comparison and Validation

Abstract

Ocean bathymetric maps are extremely important, providing fundamental information for a variety of research fields, and high-resolution is critical for many applications. More than 80% of ocean bathymetry remains un-mapped by acoustic survey. Much effort is now being expended to fill in the un-mapped area in high-resolution and over a global scale, and the present study attempted to contribute to that effort through enabling the production of high-resolution bathymetric maps from lower-resolution data. We have trialed super-resolution techniques for ocean bathymetry using deep learning approaches and reveal the potency of these methods. Five deep-learning super-resolution architectures, which were already recognized as state-of-the-art methods in natural-image super-resolution tasks, were applied to the task to generate from a 100 m-mesh grid, a 50 m-mesh grid ocean bathymetric map. The results suggest that each method has strengths and weaknesses, depending on the water depth gradient in the ocean bathymetric map area, and we conclude that it may be better to use different methods depending on the area of the maps for which the higher resolution is desired.