| one publication added to basket [355615] | A global surface ocean fCO2 climatology based on a feed-forward neural network
Zeng, J.; Nojiri, Y.; Landschützer, P.; Telszewski, M.; Nakaoka, S. (2014). A global surface ocean fCO2 climatology based on a feed-forward neural network. J. Atmos. Oceanic. Technol. 31(8): 1838-1849. https://dx.doi.org/10.1175/jtech-d-13-00137.1 In: Journal of Atmospheric and Oceanic Technology. American Meteorological Society: Boston, MA. ISSN 0739-0572; e-ISSN 1520-0426, more |   |
| Keyword | | | Author keywords | Fluxes; Carbon dioxide; Climatology; Neural networks; Oceanic variability |
| Authors | | Top | - Zeng, J.
- Nojiri, Y.
- Landschützer, P., more
| - Telszewski, M.
- Nakaoka, S.
| |
| Abstract | A feed-forward neural network is used to create a monthly climatology of the sea surface fugacity of CO2 (fCO2) on a 1° × 1° spatial resolution. Using 127 880 data points from 1990 to 2011 in the track-gridded database of the Surface Ocean CO2 Atlas version 2.0 (Bakker et al.), the model yields a global mean fCO2 increase rate of 1.50 μatm yr−1. The rate was used to normalize multiple years’ fCO2 observations to the reference year of 2000. A total of 73 265 data points from the normalized data were used to model the global fCO2 climatology. The model simulates monthly fCO2 distributions that agree well with observations and yields an anthropogenic CO2 update of −1.9 to −2.3 PgC yr−1. The range reflects the uncertainty related to using different wind products for the flux calculation. This estimate is in good agreement with the recently derived best estimate by Wanninkhof et al. The model product benefits from a finer spatial resolution compared to the product of Lamont–Doherty Earth Observatory (Takahashi et al.), which is currently the most frequently used product. It therefore has the potential to improve estimates of the global ocean CO2 uptake. The method’s benefits include but are not limited to the following: (i) a fixed structure is not required to model fCO2 as a nonlinear function of biogeochemical variables, (ii) only one neural network configuration is sufficient to model global fCO2 in all seasons, and (iii) the model can be extended to produce global fCO2 maps at a higher resolution in time and space as long as the required data for input variables are available. |
|
