Using deep learning to map sea level and ocean dynamics along the North Atlantic eastern boundary

Project Background

Rising seas due to melting ice sheets and ocean warming pose a major threat to the large fraction of the world's population living in low-lying coastal regions. As the climate warms, large-scale adjustments in the ocean's circulation may also drive abrupt changes in coastal sea level, locally amplifying or ameliorating global mean sea level rise and its impacts. Therefore, while much sea level research has centred around its global mean value, there is a growing realisation the we must better understand the factors, such as the ocean's circulation, that drive regional variations in sea level. For example, ocean models suggest that sea level falls by about 60 cm along the eastern boundary of the North Atlantic between the equator and Northern Europe. However, due to measurement limitations, accurate observations of this change in coastal sea level, or the dynamics by which it is maintain, have yet to be made. Machine learning provides an opportunity for fundamental progress to be made in this exciting field.

Project Aims and Methods

The ultimate aim of this project is to identify the ocean processes sustaining the large drop in sea level along the eastern boundary of the North Atlantic, thereby explaining a poorly understood feature of the ocean's large-scale circulation with implications for sea level rise. This goal is hampered, however, by three primary difficulties: (i) obtaining a clean sea level signal in coastal zones using satellite altimetry; (ii) removing the much larger gravity signal from the SSH to obtain the small (~1%) residual related to ocean dynamics and (iii) relating the surface signal to sparse sub-surface observations to identify the boundary dynamics. Here we propose to employ machine learning techniques to overcome these difficulties. Correspondingly, the project will procced in three stages:

1. Use deep learning and image classification to reprocess altimetry data to determine with unprecedented accuracy the sea surface height (SSH) along the North Atlantic eastern boundary (NAEB).
2. Exploit synergies between signal processing and machine learning to optimally combine and filter the SSH and gravity field to determine the absolute dynamic ocean topography along the NAEB.
3. Use deep learning trained on sparse observations and ocean models to identify the dynamical balances and controlling factors that maintain and modify the long-term sea level gradient along the NAEB.

The aims of the project will be achieved through the development and application of novel machine learning techniques to a range of geodetic and oceanographic observations and model data sets, exploiting the power of Bristol University's advanced computing facilities. Many of the observations and techniques, particularly with respect to machine learning are only now reaching sufficient maturity for the aims of the project to be realised.