Stochastic and Machine Learning Predictions of Bathymetry

Abstract:

Bathymetry is a fundamental property of the Earth, yet more than 80% of the world’s oceans remain unmapped. Bathymetry is a critical parameter for understanding global ocean circulation, marine geo-hazards, sub-sea cable routing, and many other scientific, economic, and defense applications. While Seabed 2030 aims to map global bathymetry in the near future, there remains a need for the best possible estimates of bathymetry today. Currently, there are global estimates of bathymetry, but these predictions are at a relatively low spatial resolution (~1 x 1 km at the equator). Here, using machine learning algorithms (MLAs), global satellite altimetry data, and a global dataset of bathymetric soundings, we predict bathymetry natively at ~500 x 500 m resolution. Preliminary results show higher amplitude predictions at all spatial frequencies compared to past techniques, with the greatest change is power at the higher spatial frequencies, which is a known issue in past methodologies. These results show that these new methods are able to extract additional information from altimetry datasets compared to past techniques.

Even with improvements in bathymetric prediction and the potential for global bathymetric mapping, applications of bathymetry to acoustic scattering, surge and inundation modeling, and ocean dynamic models need bathymetry at finer resolutions than can be predicted (~ 500 x 500 m) and mapped (100 x 100 m to 800 x 800 m depending on depth). To address these needs, we developed a methodology to produce stochastically rough digital elevation models (SRDEMs) at any desired resolution. By extrapolating amplitude trends in the frequency domain and producing stochastic phase information, this method adds roughness – elevation variability at higher spatial resolution than is resolved from observation – whose spectral statistics are consistent with low-resolution data. Instead of relying on other methods to increase spatial resolution such as interpolation, which produces unrealistically smooth surfaces, our method delivers a surface that accurately represents observations as well as possesses stochastic roughness to any desired resolution. While these surfaces are not the “true” bathymetry, the SRDEMs produce more accurate derived parameters and statistics than interpolation-based methods.