Neural Network: an alternative method to estimate nutrient concentrations in the Gulf of Mexico? Orens Pasqueron de Fommervault1,*, Raphaëlle Sauzede2,3 Paula Perez-Brunius1 and Julio Sheinbaum1 * Contact info:
[email protected]
1 Departamento
de Oceanografía Física, Centro de Investigación Científica y de Educación Superior de Ensenada, Mexico 2 Observatoire Océanologique de Villefranche, Laboratoire d'Océanographie de Villefranche, France 3 Écosystèmes Insulaires Océaniens, IRD, Ifremer, UPF and ILM, French Polynesia
What are Neural Networks?
CANYON: a Neural Network to retrieve nutrient concentrations
A neural network is a set of algorithms initially designed to mirror biological neurons (McCulloch et Pitts, 1943). They are largely used in biostatistics and oceanography to estimate biogeochemical parameters (e.g. Kd490, pCO2, phytoplankton properties, etc.).
CANYON is a Multi-Layered Perceptron developed by Raphaëlle Sauzède to retrieve carbonate system parameters and nutrient concentrations (i.e. nitrate, phosphate and silicate) in the global ocean (Sauzède et al., In revision).
Neural networks are particularly appropriate for large datasets and non-linear relationships (compared to “classical” statistical tools). However they are sometimes considered as black boxes.
The chosen input variables include hydrological and biogeochemical components, spatial component and temporal component (Fig. 2).
Figure 1: Schematic representation of CANYON neural network.
CANYON training and validation (Sauzède et al., In revision) Data used to train and validate CANYON (37,774 profiles, Fig. 2) were recorded in the GLODAPv2 database (Olsen et al., 2016). CANYON was trained using 80 % randomly chosen data from the whole database (excluding 8 “independent zones”). The remaining 20 % data were used for the neural network validation.
CANYON performance Figure 2: GLODAPv2 database (Olsen et al., 2016). The 8 colored boxes delineate the 8 independent zones.
In the global ocean, the method predicts nutrient concentrations with good accuracy (RMSE = 0.83, Fig. 3 ).
Nitrate (XIXIMI)
Figure 4: XIXIMI sampling stations (black circles and floats tracks).
Example of application: APEX Float time-series CANYON was further applied on APEX floats deployed in the Gulf of Mexico (Fig. 4) and the resulting time-series are presented on Fig. 6.
Nitrate (CANYON)
depth
CANYON in the Gulf of Mexico As a validation in the Gulf of Mexico, the CANYON method was first applied on the XIXIMI dataset (Fig. 4). Temperature, salinity and oxygen measurements were used as input variables to estimate the concentrations of nutrients. The overall agreement between the CANYON-simulated nitrate and its measured in situ counterparts is satisfactory (Fig. 5).
Figure 3: Comparison of CANYON and in situ values
Figure 5: Nitrate bottle measurements acquired during XIXIMI (left) and nitrate estimated by CANYON (right).
Take-home message CANYON algorithm, developed for the global ocean, seems to be efficient in the Gulf of Mexico. CANYON can have a wide range of applications:
Nutrients estimations from high-frequency time-series (e.g. profiling floats, gliders) Autonomous measurements calibration (e.g. SUNA, ISUS) Quality control of database Model initialization, forcing
Related paper
Figure 6: Nitrate floats time-series estimated by CANYON
R. Sauzède, H. Claustre, O. Pasqueron de Fommervault, H. Bittig, J.-P. Gattuso, L. Legendre, K-S. Johnson. Estimates of watercolumn nutrients concentration and carbonate system parameters in the global ocean: A novel approach based on neural networks. (Frontiers in Oceanography, In revision).