Climatologies of seawater CO2 chemistry variables: A neural network approach

• Main Authors: Broullón, Daniel, Pérez, Fiz F., Velo, A., Suzuki, Toru
• Other Authors: Ministerio de Economía y Competitividad (España), European Commission
• Format: Still Image
• Language: English
• Published: 2017
• Subjects: Ocean; Carbonate system; Climatologies; Neural network; Hagan; Ocean acidification
• Online Access: http://hdl.handle.net/10261/157206; https://doi.org/10.13039/501100000780; https://doi.org/10.13039/501100003329

1 poster presented at the 10th International Carbon Dioxide Conference, Interlaken, Switzerland, 21 August 2017 - 25 August 2017.-- Daniel Broullón . et al. For decades, the anthropogenic modification of the carbon cycle has been widely studied. More recently, ocean acidification studies have increased significantly. Establishing robust climatologies of seawater CO2 chemistry variables and building models are a key point for a better understanding of the associated processes. The availability and quality of data is crucial for the evaluation of climate models and, consequently, to improve their predictions. Version 2 of the Global Ocean Data Analysis Project (GLODAPv2) is an internally consistent data product composed of data from 724 scientific cruises covering the entire global ocean. Among others, it contains seawater CO2 chemistry variables such as total alkalinity (AT), total dissolved inorganic carbon (TCO2) and pH. This sparse dataset has been subjected to extensive quality control and different interpolation techniques have been applied to extend the data coverage on a homogeneous grid (Lauvset et al. 2016). We propose a novel neural network approach to generate annual and monthly climatologies of AT, TCO2, pH and both calcite and aragonite saturation state from the GLODAPv2 dataset for the preindustrial and current periods. These climatologies are fitted to the World Ocean Atlas 2013 version 2 (WOA13v2) grid. WOA13v2 is a set of objectively analyzed (1° grid) climatological fields of different oceanographic variables (but not CO2 system) at standard depth levels for annual, seasonal, and monthly compositing periods for the World Ocean. A feed-forward neural network was chosen in a multi-layer architecture, which allows linear and nonlinear variability to be assimilated by the network. The proposed configuration is able to approximate most functions arbitrarily well (Hagan et al., 2014). We have tested different neural network designs and sizes to obtain the minimum error. For that, the number of ...