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?

Abstract

Over the last decade, advanced statistical inference and machine learning have been used to fill the gaps in sparse surface ocean CO 2 measurements . The estimates from these methods have been used to constrain seasonal, interannual and decadal variability in seaair CO 2 fluxes and the drivers of these changes , 2016, Gregor et al. 2018). However, it is also becoming clear that these methods are converging towards a common bias and RMSE boundary: the wall , which suggests that p CO 2 estimates are now limited by both data gaps and scalesensitive observations. Here, we analyse this problem by introducing a new gapfilling method, an ensemble of six machine learning models (CSIRML6 version 2019a), where each model is constructed with a twostep clusteringregression approach.The ensemble is then statistically compared to wellestablished methods. The ensemble, CSIRML6, has an RMSE of 17.16 µatm and bias of 0.89 µatm when compared to a testdataset kept separate from training procedures. However, when validating our estimates with independent datasets, we find that our method improves only incrementally on other gapfilling methods. We investigate the differences between the methods to understand the extent of the limitations of gapfilling estimates of p CO 2 . We show that disagreement between methods in the South Atlantic, southeastern Pacific and parts of the Southern Ocean are too large to interpret the interannual variability with confidence. We conclude that improvements in surface ocean p CO 2 estimates will likely be incremental with the optimisation of gapfilling methods by (1) the inclusion of additional clustering and regression variables ( e.g. eddy kinetic energy), (2) increasing the sampling resolution. Larger improvements will only be realised with an increase in CO 2 observational coverage, particularly in today's poorly sampled areas. 1Geosci. Model Dev. Discuss., https://doi.

Atmospheric CO2 and Sea Surface Temperature Variability Cannot Explain Recent Decadal Variability of the Ocean CO2 Sink

“…For the biogeochemical models, structural biases are potentially introduced by inaccurate forcing fields, by parameterizations of biological processes and physical circulation processes (Hauck et al., 2015), and by model spin‐up procedures (Séférian et al., 2016). For the pCO 2 interpolation products, artifacts can be introduced by the interpolation of sparse pCO 2 data (Gloege et al., 2021; Gregor et al., 2019; Rödenbeck et al., 2015), and additional uncertainty is introduced by the input products and formulations used for the bulk air‐sea gas exchange formula (Fay et al., 2021; Roobaert et al., 2018; Watson et al., 2020). Using an ensemble of different products, as done here, reduces the influence of these sources of error if they are randomly distributed, but any common biases across products within each method could bias the conclusions presented here.…”

“…Each group uses these interpolated p CO 2, sw maps, along with reanalysis SST and gas transfer velocities, to calculate the air‐sea CO 2 flux according to Equation . The products used here are four of the products used in the 2021 Global Carbon Budget (Chau et al., 2021; Gregor et al., 2019; Landschützer et al., 2016; Rödenbeck et al., 2014). Air‐sea CO 2 fluxes calculated from these data‐based pCO 2 reconstructions capture all sources of variability in air‐sea CO 2 fluxes (Figure 1; Table 1).…”

Global Carbon Budget 2021

“…An ensemble average of six machine learning estimates of surface ocean pCO 2 using the approach described in Gregor et al (2019) with the updated prod-uct using SOCAT v2021 . All ensemble members use a cluster-regression approach.…”

“…Each of the estimates uses a different method to then map the SOCAT v2021 data to the global ocean. The methods include a data-driven diagnostic method (Rödenbeck et al, 2013; referred to here as Jena-MLS), three neural network models (Landschützer et al, 2014; referred to as MPI-SOMFFN; Copernicus Marine Environment Monitoring Service, referred to here as CMEMS-LSCE-FFNN; and Zeng et al, 2014; referred to as NIES-FNN), two cluster regression approaches (Gregor et al, 2019; referred to here as CSIR-ML6; and Gregor and Gruber, 2021, referred to as OS-ETHZ-GRaCER), and a multi-linear regression method (Iida et al, 2021; referred to as JMA-MLR). The ensemble mean of the f CO 2 -based flux estimates is calculated from these seven mapping methods.…”

Alternate Histories: Synthetic Large Ensembles of Sea‐Air CO2 Flux

“…In this study, we use annual mean sea‐air CO 2 flux estimates from four observation‐based products as produced by the SeaFlux product (Fay et al., 2021; Gregor & Fay, 2021; Table 1): The Council for Scientific and Industrial Research‐Machine Learning ensemble (CSIR‐ML6; Gregor et al., 2019), the Max Planck Institute Self‐Organizing Map‐Feed‐Forward Neural Network (MPI‐SOMFFN; Landschützer et al., 2013, 2014, 2015, 2016), the Max Planck Institute for Biogeochemistry‐Mixed Layer Scheme (JENA‐MLS; Rödenbeck et al., 2014), and the Copernicus Marine Environment Monitoring Service Feed‐Forward Neural Network (CMEMS‐FFNN; Denvil‐Sommer et al., 2019). Fluxes are calculated monthly and then averaged to annual values for analysis.…”