Denoising Geodetically-Determined Ocean Currents with Deep Neural Networks

  • Laura Gibbs

Student thesis: Doctoral ThesisDoctor of Philosophy (PhD)

Abstract

Determining an accurate picture of ocean currents is an important societal challenge for oceanographers, aiding our understanding of the vital role currents play in regulating Earth's climate, and in the dispersal of marine species and pollutants, including microplastics. The geodetic approach, which combines satellite observations of sea level and Earth's gravity, offers the only means to estimate the dominant geostrophic component of these currents globally. Unfortunately, however, geodetically-determined geostrophic currents suffer from high levels of contamination in the form of geodetic noise. Conventional approaches use isotropic spatial filters to improve the signal to noise ratio, though this results in high levels of attenuation. In this thesis, I explore the use of deep learning to improve the geodetic determination of the ocean currents. A common theme which appears throughout this work is the modification of deep learning approaches and techniques towards the domain specific problem of geodetic noise removal.

As is the case in many real-world scenarios, the noise-free `ground truth version' does not exist. Thus, this thesis begins by investigating an unsupervised approach which does not require such ground-truth data and exploring its potential and suitability towards geodetic noise removal. This work aims to demonstrate the capability of a deep learning approach which is trained without guidance and provides valuable insight for the research direction of the following chapters.

Following, to facilitate the exploration of supervised deep learning techniques, a training dataset of input and target pairs is created. This involves substituting clean targets with naturally smooth climate model data and adding synthetic noise. Due to the complex nature of geodetic noise, I investigate the use of deep generative networks for its realistic replication. Furthermore, I present several techniques to improve the realism of the synthetic data, utilising known properties of the real-world geodetic MDT and associated current data itself.

Next, I present a full supervised denoising pipeline where a convolutional denoising autoencoder (CDAE) is trained using the synthetic training dataset from the previous chapter. Performance enhancements are explored through further leveraging domain specific knowledge during the training phase and their contributions to denoising performance are evaluated through a comprehensive comparative study. The trained CDAE model is applied to unseen real geodetic data and predictions are thoroughly evaluated against conventional approaches. It is demonstrated that our method outperforms conventional isotropic filtering on the combined geoid derived product regionally.

Finally, I construct a high parameter denoising network and utilise lessons learned from previous chapters to effectively remove geodetic noise. In this application focused chapter, I denoise the entire global field, and thus propose techniques to handle polar distortion. The proposed approach is able to resolve both major and more subtle current structures from the global MDT derived from satellite only data. It is found that the denoised product agrees well with multiple reference surfaces.
Date of Award19 Mar 2024
Original languageEnglish
Awarding Institution
  • University of Bristol
SupervisorRory J Bingham (Supervisor) & Fanny M Monteiro (Supervisor)

Keywords

  • deep learning
  • geostrophic currents
  • remote sensing

Cite this

'