LSCE-FFNN-v1: a two-step neural network model for the reconstruction of surface ocean <i>p</i>CO<sub>2</sub> over the global ocean

Abstract

Abstract. A new feed-forward neural network (FFNN) model is presented to reconstruct surface ocean partial pressure of carbon dioxide (pCO2) over the global ocean. The model consists of two steps: (1) the reconstruction of pCO2 climatology, and (2) the reconstruction of pCO2 anomalies with respect to the climatology. For the first step, a gridded climatology was used as the target, along with sea surface salinity (SSS), sea surface temperature (SST), sea surface height (SSH), chlorophyll a (Chl a), mixed layer depth (MLD), as well as latitude and longitude as predictors. For the second step, data from the Surface Ocean CO2 Atlas (SOCAT) provided the target. The same set of predictors was used during step (2) augmented by their anomalies. During each step, the FFNN model reconstructs the nonlinear relationships between pCO2 and the ocean predictors. It provides monthly surface ocean pCO2 distributions on a 1∘×1∘ grid for the period from 2001 to 2016. Global ocean pCO2 was reconstructed with satisfying accuracy compared with independent observational data from SOCAT. However, errors were larger in regions with poor data coverage (e.g., the Indian Ocean, the Southern Ocean and the subpolar Pacific). The model captured the strong interannual variability of surface ocean pCO2 with reasonable skill over the equatorial Pacific associated with ENSO (the El Niño–Southern Oscillation). Our model was compared to three pCO2 mapping methods that participated in the Surface Ocean pCO2 Mapping intercomparison (SOCOM) initiative. We found a good agreement in seasonal and interannual variability between the models over the global ocean. However, important differences still exist at the regional scale, especially in the Southern Hemisphere and, in particular, in the southern Pacific and the Indian Ocean, as these regions suffer from poor data coverage. Large regional uncertainties in reconstructed surface ocean pCO2 and sea–air CO2 fluxes have a strong influence on global estimates of CO2 fluxes and trends.

A comparative assessment of the uncertainties of global surface­-ocean CO<sub>2</sub> estimates using a machine learning ensemble (CSIR-ML6 version 2019a) – have we hit the wall?

“…Importantly, this highlights the need for new methods that are fundamentally different and may lead to the development of procedural architectures that might be able to resolve the biases in wellsampled regions better. For example, a recent study by DenvilSommer et al (2018) developed a method (LSCEFFNN) that first estimates the climatological p CO 2 and then the anomalies from this climatology -their method reported RMSE scores on the order of those reported in this study (~18.0 µatm) and very low R iav scores (< 0.2). While new methods might not lead to drastic reductions in uncertainties, incremental improvements in uncertainties will be driven by approaches that offer new solutions, whether it be increased resolution, additional featurevariables or a new approach.…”

Revised estimates of ocean-atmosphere CO2 flux are consistent with ocean carbon inventory

“…n recent years, an international effort has assembled a qualitycontrolled database of surface ocean carbon dioxide observations, the Surface Ocean Carbon Dioxide Atlas (SOCAT) [1][2][3] . SOCAT has enabled several recent studies evaluating air-sea CO 2 flux from the observed partial pressure at the ocean surface [4][5][6][7][8][9][10] . In order to use the data to obtain accurate values of oceanatmosphere CO 2 fluxes, it is necessary to apply the gas exchange equation to the concentration difference of dissolved CO 2 across the mass boundary layer (MBL) of the sea surface-the topmost 100 µm within which molecular diffusion dominates vertical transport toward the interface (see Methods section).…”

OceanSODA-ETHZ: A global gridded data set of the surface ocean carbonate system for seasonal to decadal studies of ocean acidification

“…By far the most established efforts are those that interpolate and extrapolate the ocean pCO 2 observations, as demonstrated by the inter-comparison project by Rödenbeck et al (2015). Feed-forward neural networks (FFNN) have become one of the favored tools (Landschützer et al, 2013;Zeng et al, 2014;Denvil-Sommer et al, 2019), but other statistical and machine learning methods, such as Bayesian regression and tree-based regression, have also been used with similar success (Rödenbeck et al, 2014;Gregor et al, 2019). However, the specific implementation of the methods is what sets the assortment of methods apart.…”