A neural network-based method (2016)

R. Sauzède (1)*, H. Claustre (1), J. Ritz (1), C. Jamet (2), G. Dall’Olmo (3,4), F. D’Ortenzio (1), B. Gentili (1), A. Poteau (1) and C. Schmechtig (1)

(1) : Sorbonne Universités, UPMC Univ Paris 06, CNRS, Observatoire Océanologique de Villefranche, Laboratoire d’Océanographie de Villefranche, Villefranche-sur-Mer, France
(2) : Laboratoire d’Océanologie et de Géosciences, UMR8187, ULCO/CNRS, Wimereux, France
(3) : Plymouth Marine Laboratory, UK
(4)  : National Centre for Earth Observation, Plymouth, UK

 

* Corresponding author : sauzede@obs-vlfr.fr 
Full paper: Sauzède, R., H. Claustre, J. Uitz, C. Jamet, G. Dall'Olmo, F. D'Ortenzio, B. Gentili, A. Poteau, and C. Schmechtig 2016, A neural network-based method for merging ocean color and Argo data to extend surface bio-optical properties to depth: Retrieval of the particulate backscattering coefficient, J. Geophys. Res. Oceans, 121, 2552–2571, doi:10.1002/2015JC011408.
Context

Several bio-optical properties are estimated quasi-synoptically from remote sensing of ocean colour radiometry. However, these estimates only represent approximately one fifth of the productive layer. Therefore, this information appears to be insufficient for biogeochemical studies involving primary production or carbon export. 

Several bio-optical properties are now measured in situ with a high spatio-temporal resolution thanks to Bio-Argo profiling floats. The ~5000 Bio-Argo profiles available in May 2015 provided the frame to consider the development of a new parameterization of bio-optical properties that could be used to derive a four dimensional view (i.e. according to x, y, z and t) of several bio-optical proxies of the phytoplankton biomass for the global ocean.

Data & Method

Neural network-based methods that merge ocean colour and Argo data to extend surface bio-optical properties to depth.

Two neural networks were developed to merge ocean colour observations with temperature and salinity profiles acquired by Argo profiling floats with the aim of modelling the vertical distribution of (1) the chlorophyll a concentration (Chl) associated to the total phytoplankton biomass and to three phytoplankton size classes (i.e. the micro-, the nano- and the pico-phytoplankton) and (2) the particulate backscattering coefficient (bbp), a widely used proxy of the particulate organic carbon and phytoplankton carbon in the open ocean.

 

Figure 1: Geographic distribution of the 4725 stations sampled by the Bio-Argo floats and the 16 stations sampled during an oceanographic cruise used for the training and the validation of the SOCA method that estimates the vertical distribution of bbp from merged ocean colour and Argo data

The developed methods are referred as SOCA for “Satellite Ocean Colour merged with Argo data to infer the vertical distribution of bio-optical properties” [see Sauzède, 2015; Sauzède et al., 2016]. The inputs of these neural networks are composed of three main components: (1) the satellite estimates, (2) the vertically-resolved physical properties derived from physical Argo profiles and (3) the day of the year of the considered matchup. 

These methods were trained and validated using a large database of 4 725 concurrent profiles of temperature, salinity and bio-optical properties measured from Bio-Argo profiling floats concomitant with the satellite products. Independent datasets composed of (1) 20% of the initial database chosen randomly, (2) “independent” Bio-Argo floats profiles and (3) oceanographic cruises samples were used to evaluate the performance of these methods (see Figure 1). 

Results & Perspectives

The results appear to be very promising (~21% of global error), with a good representation  of the vertical seasonal distribution of bio-optical properties (see the results for the retrieval of the vertical distribution of bbp from SOCA compared to the reference in situ bbp measurements acquired from two “independent” Bio-Argo profiling floats in Figure 2.

 

Figure 2: Comparison of the reference bbp measurements acquired by Bio-Argo floats, bbp_Floats (a and c) with the values predicted by SOCA, bbp_SOCA (b and d), modelled from merged ocean colour and Argo data. Time series for the Bio-Argo floats deployed in the North Atlantic Subtropical Gyre (a-b) and in the Southern Ocean (c-d). The grey line in each panel indicates the mixed layer depth.

Finally, this study makes use of the global Bio-Argo network and the results have highlighted one of the potential of sampling with high spatio-temporal resolution bio-optical properties from Bio-Argo profiling floats at a global scale. In fact, using the global array of Bio-Argo profiling floats for training neural networks, it becomes now possible to estimate using physical properties measured from Argo floats merged with ocean colour observations, the vertical distribution of key bio-optical proxies of the phytoplankton biomass in the global ocean. The global and regional 4D views of Chl, phytoplankton community size indices and bbp that could be obtained from SOCA methods represent a new important tool to assess seasonal and inter-annual variability in the vertical distribution of phytoplankton biomass and community composition at a global scale. 

References
  • Sauzède, R. (2015), Study and parameterization of the vertical distribution of the phytoplankton biomass in the global ocean.