15 arc sec Global Prediction of Bathymetry Utilizing Deep Neural Networks
Abstract
Bathymetry is a fundamental property of the Earth, that is critical for understanding global ocean circulation, marine geo-hazards, sub-sea cable routing, and many other scientific, economic, and defense applications. Yet more than 80% of the worlds oceans remain unmapped. While Seabed 2030 aims to map global bathymetry in the near future, there remains a need for the best possible estimates of bathymetry today. Current global estimates of bathymetry are at a relatively low spatial resolution (~1.8 x 1.8 km at the equator) or are reported at higher spatial resolution but are unrealistically smooth in comparison to measure bathymetry (e.g., SRTM15+V2). Here, using a deep neural network, a global dataset of bathymetric soundings, global satellite altimetry data and other known geospatial quantities, we predict bathymetry natively at 15 arc second resolution (~450 x 450 m at the equator). Preliminary results along the Cascadia margin show higher amplitude predictions at all spatial frequencies compared to past techniques, with the greatest change in power at the higher spatial frequencies, which past methodologies are known to under predict. Further, by combining the deep neural network with a conformal prediction methodology, we are able to estimate geospatially-variable uncertainties based on past predictive skill. These results and uncertainties show that deep learning neural networks are able to extract additional information from altimetry and other geospatial datasets compared to past techniques and result in more geospatially realistic estimates of global bathymetry.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMOS51A..01P