Accurate monitoring of regional and global ocean processes is crucial for a full understanding of changes in the ocean circulations and for an assessment of critical impacts of anthropogenic pressures on climate and marine ecosystems. Therefore, it is important to determine the distributions of essential ocean variables (EOVs), like temperature and salinity, as well as their interactions over a wide range of spatial and temporal scales both in horizontal and vertical dimensions. To achieve reliable ocean observations, in situ and remote sensing platforms have been developed collecting EOVs data at great resolution, and to enable process-based modelling efforts of ocean dynamics. However, the distribution of subsurface ocean measurements is extremely sparse and irregular, which presents a challenge for gaining a comprehensive understanding of the ocean and marine ecosystems. To overcome the limits of sparse temporal and spatial observations, neural networks combining remotely-sensed surface measurements and in situ vertical profiles are increasingly being used to obtain high-quality three-dimensional estimates of subsurface ocean state variables. This study proposes a convolutional neural network (CNN) for reconstruction of vertical profiles using satellite surface measurements as input collected between 2005 and 2020 in the Atlantic Ocean, and its performance is shown to be superior to current state-of-the-art methods. Different combinations of surface variables are analyzed and compared to determine the key surface variables for ocean structure reconstruction. Furthermore, the relative importance of each of these variables is estimated over the full vertical profiles.
Primary Presenter: Philip Smith, Technical University of Denmark (pahsm@aqua.dtu.dk)
Authors:
Philip Smith, Technical University of Denmark, National Institute of Aquatic Resources (DTU - Aqua) (pahsm@aqua.dtu.dk)
Kristian Sørensen, Technical University of Denmark, National Space Institute (DTU - Space) (kaaso@space.dtu.dk)
Bruno Nardelli, Consiglio Nazionale delle Ricerche, Istituto di Scienze Marine (CNR-ISMAR) (bruno.buongiornonardelli@cnr.it)
Anshul Chauhan, Technical University of Denmark, National Institute of Aquatic Resources (DTU - Aqua) (anscha@aqua.dtu.dk)
Asbjørn Christensen, Technical University of Denmark, National Institute of Aquatic Resources (DTU - Aqua) (asc@aqua.dtu.dk)
Michael John, Technical University of Denmark, National Institute of Aquatic Resources (DTU - Aqua) (mstjo@aqua.dtu.dk)
Patrizio Mariani, Technical University of Denmark, National Institute of Aquatic Resources (DTU - Aqua) (pat@aqua.dtu.dk)
Reconstruction of Sub-Surface Ocean State Variables using Convolutional Neural Network
Category
Scientific Sessions > SS012 The Next Frontier: Linking Remote Sensing, Data Science, Modeling, Open Science, and the Aquatic Sciences To Understand Emergent Properties of Aquatic Systems
Description
Time: 06:30 PM
Date: 8/6/2023
Room: Mezzanine