Abstract's details
Reconstruction of Sea Salinity Profiles from surface parameters in the tropical Atlantic using Neural Networks.
CoAuthors
Event: 2014 Ocean Surface Topography Science Team Meeting
Session: Others (poster only)
Presentation type: Type Poster
Contribution: not provided
Abstract:
Geophysical systems (ocean, atmosphere,...) have many properties of complex systems. To study these systems, different approaches are used, the most common one being the numerical modelling. However it does not always take into account the complexities that characterize these systems through their parameters. Methods based on the knowledge related to the observations offer new opportunities to this problem.
The work presented here aims to reconstruct salinity profiles from surface parameters using neural networks. This approach will allow to get information on the salinity sub-surface variability even if only surface -for instance satellite- parameters are available.
First, a study of the inter-relationship between the salinity at different depths has shown the binding aspect of this parameter at different levels. Extraction of relevant features is then applied to the surface data to keep the best parameters for the inversion model.
The model consists of two parts in which the parameters are considered in terms of their ability to describe the salinity variability. In a first step, the salinity profiles associated with the relevant physical surface parameters are classified to depict the various oceanic situations sampled. This learning procedure is mainly based on in situ data such as Argo floats. The second step concerns the inversion. The model has been applied on surface parameters over the tropical Atlantic ocean where and during which in situ salinity profiles can be further used for validation. A comparison between real in situ salinity profiles and estimated profiles shows good agreement and strong correlation. However, regions highly variable such as the Maximum Salinity Waters areas are more difficult to reproduce.
The work presented here aims to reconstruct salinity profiles from surface parameters using neural networks. This approach will allow to get information on the salinity sub-surface variability even if only surface -for instance satellite- parameters are available.
First, a study of the inter-relationship between the salinity at different depths has shown the binding aspect of this parameter at different levels. Extraction of relevant features is then applied to the surface data to keep the best parameters for the inversion model.
The model consists of two parts in which the parameters are considered in terms of their ability to describe the salinity variability. In a first step, the salinity profiles associated with the relevant physical surface parameters are classified to depict the various oceanic situations sampled. This learning procedure is mainly based on in situ data such as Argo floats. The second step concerns the inversion. The model has been applied on surface parameters over the tropical Atlantic ocean where and during which in situ salinity profiles can be further used for validation. A comparison between real in situ salinity profiles and estimated profiles shows good agreement and strong correlation. However, regions highly variable such as the Maximum Salinity Waters areas are more difficult to reproduce.