An Alternative to Static Climatologies: Robust Estimation of Open Ocean CO2 Variables and Nutrient Concentrations From T, S, and O2 Data Using Bayesian Neural Networks

Abstract

This work presents two new methods to estimate oceanic alkalinity (A T), dissolved inorganic carbon (C T), pH, and pCO 2 from temperature, salinity, oxygen, and geolocation data. "CANYON-B" is a Bayesian neural network mapping that accurately reproduces GLODAPv2 bottle data and the biogeochemical relations contained therein. "CONTENT" combines and refines the four carbonate system variables to be consistent with carbonate chemistry. Both methods come with a robust uncertainty estimate that incorporates information from the local conditions. They are validated against independent GO-SHIP bottle and sensor data, and compare favorably to other state-of-the-art mapping methods. As "dynamic climatologies" they show comparable performance to classical climatologies on large scales but a much better representation on smaller scales (40-120 d, 500-1,500 km) compared to in situ data. The limits of these mappings are explored with pCO 2 estimation in surface waters, i.e., at the edge of the domain with high intrinsic variability. In highly productive areas, there is a tendency for pCO 2 overestimation due to decoupling of the O 2 and C cycles by air-sea gas exchange, but global surface pCO 2 estimates are unbiased compared to a monthly climatology. CANYON-B and CONTENT are highly useful as transfer functions between components of the ocean observing system (GO-SHIP repeat hydrography, BGC-Argo, underway observations) and permit the synergistic use of these highly complementary systems, both in spatial/temporal coverage and number of observations. Through easily and robotically-accessible observations they allow densification of more difficult-to-observe variables (e.g., 15 times denser A T and C T compared to direct measurements). At the same time, they give access to the complete carbonate system. This potential is demonstrated by an observation-based global analysis of the Revelle buffer factor, Bittig et al. Robust Estimation of CO 2 Variables and Nutrients which shows a significant, high latitude-intensified increase between +0.1 and +0.4 units per decade. This shows the utility that such transfer functions with realistic uncertainty estimates provide to ocean biogeochemistry and global climate change research. In addition, CANYON-B provides robust and accurate estimates of nitrate, phosphate, and silicate. Matlab and R code are available at https://github.com/HCBScienceProducts/.

Observing Changes in Ocean Carbonate Chemistry: Our Autonomous Future

“…The reference pH field is calculated from empirical relationships derived from hydrographic data using temperature, salinity, pressure, and oxygen as inputs [135]. These algorithms can be region-specific [135] or global [136][137][138]. It should be noted that the accuracy of the reference pH field, and thus corrected sensor pH, depends on other parameters measured on the float such as temperature, salinity, pressure, and oxygen [98].…”

“…The LIR (Locally Interpolated Regression) uses a multiple linear regression approach, and interpolates the coefficients of the regression model to any location [136,147], providing a smooth transition between region-specific relationships. The second approach is CANYON (CArbonate system and Nutrients concentration from hYdrological properties and Oxygen using a Neural-network), which uses a neural network [137,138]. Both approaches are capable of estimating TA with uncertainties of about ± 6-8 μmol kg −1 globally, though a detailed comparison between these two algorithms has not been performed and region-specific algorithms may still be necessary [98].…”

“…Simultaneously measuring two carbonate parameters would alleviate the need to rely on algorithm approaches to estimate TA, but would not remove the need to interpolate and extrapolate from available observations to provide spatially resolved maps. While improvements in mapping methods (e.g., [41,138]) are an essential component in understanding the oceanic CO 2 flux on annual to decadal timescales, these products still suffer from any biases that exist in the underlying observational dataset and will require the ongoing addition of in situ data to maintain their utility. Just as new profiling float-based estimates of pCO 2 indicate that the interpolation schemes used for surface observations of pCO 2 cannot reproduce a signal they do not observe, it is likely that issues will arise with the use of empirical algorithms for TA and other carbonate system parameters.…”

Reply to Anonymous Referee #2

“…The recent methods to compute A T proposed by Carter et al (2018) and Bittig et al (2018) (LIARv2 and CANYON-B respectively) were also compared to the one proposed here. LIARv2 is based on multilinear regressions (MLRs) including the same predictors used in the present study, excluding phosphate (sample position; salinity, S; potential temperature, θ ; nitrate, N; apparent oxygen utilization, AOU; and silicate, Si).…”

A global monthly climatology of oceanic total dissolved inorganic carbon: a neural network approach

“…Lee et al (2000) used temperature and nitrate to compute surface nDIC with an area-weighted error of ±7 µmol kg -1 . Sauzède et al (2017) and Bittig et al (2018) trained neural networks with GLODAPv2 data to compute TCO2 over the depth range 0-8000 m with an accuracy of ±9 µmol kg -1 and ±7.1 µmol kg -1 , respectively. The predictor variables used in those studies were location, pressure, temperature, salinity and dissolved oxygen.…”