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

Abstract. Over the last decade, advanced statistical inference and machine learning have been used to fill the gaps in sparse surface ocean CO2 measurements (Rödenbeck et al., 2015). The estimates from these methods have been used to constrain seasonal, interannual and decadal variability in sea–air CO2 fluxes and the drivers of these changes (Landschützer et al., 2015, 2016; Gregor et al., 2018). However, it is also becoming clear that these methods are converging towards a common bias and root mean square error (RMSE) boundary: “the wall”, which suggests that pCO2 estimates are now limited by both data gaps and scale-sensitive observations. Here, we analyse this problem by introducing a new gap-filling method, an ensemble average of six machine-learning models (CSIR-ML6 version 2019a, Council for Scientific and Industrial Research – Machine Learning ensemble with Six members), where each model is constructed with a two-step clustering-regression approach. The ensemble average is then statistically compared to well-established methods. The ensemble average, CSIR-ML6, has an RMSE of 17.16 µatm and bias of 0.89 µatm when compared to a test dataset kept separate from training procedures. However, when validating our estimates with independent datasets, we find that our method improves only incrementally on other gap-filling methods. We investigate the differences between the methods to understand the extent of the limitations of gap-filling estimates of pCO2. We show that disagreement between methods in the South Atlantic, southeastern Pacific and parts of the Southern Ocean is too large to interpret the interannual variability with confidence. We conclude that improvements in surface ocean pCO2 estimates will likely be incremental with the optimisation of gap-filling methods by (1) the inclusion of additional clustering and regression variables (e.g. eddy kinetic energy), (2) increasing the sampling resolution and (3) successfully incorporating pCO2 estimates from alternate platforms (e.g. floats, gliders) into existing machine-learning approaches.

Strengths and weaknesses of three Machine Learning methods for pCO<sub>2</sub> interpolation

“…We test NN, XGBoost (XGB) and random forest (RF) regression. Our XGB and RF are similar to the GBM and ERT of Gregor et al (2019), respectively. The three approaches we test have different properties, advantages, and drawbacks.…”

“…Though RF does almost as well as XGB on the test data, it generalizes poorly. Gregor et al (2019) also found that the random forest method performs poorly in pCO 2 reconstruction.…”

A Regional Neural Network Approach to Estimate Water-Column Nutrient Concentrations and Carbonate System Variables in the Mediterranean Sea: CANYON-MED

“…Machine learning methods represent a promising way to fill these "observational gaps". They have the potential to predict, from variables systematically measured by autonomous platforms, variables still difficult to measure accurately and cost-effectively with these platforms (e.g., Gregor et al, 2019). Transfer functions such as multiple linear regressions (e.g., Velo et al, 2013;Carter et al, 2018) have therefore been developed to estimate biogeochemical variables.…”

Consistency and Challenges in the Ocean Carbon Sink Estimate for the Global Carbon Budget

“…The data-products are inter-and extrapolations of SOCAT data, and hence have higher correlation coefficients and lower RMSEs than the GOBMs. However, they also have RMSEs of 15-20 µatm (see also Gregor et al, 2019) and correlation coefficients of 0.8-1.0 with lower correlation values in the southern and northern extratropics ( Figure 9A, Supplementary Figures 1, 3-15 for individual models and data-products). Comparison on monthly time-scales is a common approach to measure misfit between estimated and observed pCO 2 (e.g., Le Quéré et al, 2018a;Friedlingstein et al, 2019;Gregor et al, 2019).…”

“…We argue that the detrended annual statistic is more informative for the evaluation of annual estimates of S OCEAN with the aim to robustly estimate the mean S OCEAN and multi-year variability. The monthly statistics, which are most commonly used to evaluate GOBMs and data-products (Rödenbeck et al, 2015;Friedlingstein et al, 2019;Gregor et al, 2019) quantify to a large extent the representation of the seasonal cycle. Based on our analysis of model-data mismatch, we conclude that misrepresentations of the seasonal cycle in GOBMs have little effect on the global annual estimate of S OCEAN .…”