A seamless ensemble-based reconstruction of surface ocean <i>p</i>CO<sub>2</sub> and air–sea CO<sub>2</sub> fluxes over the global coastal and open oceans

Abstract

Abstract. We have estimated global air–sea CO2 fluxes (fgCO2) from the open ocean to coastal seas. Fluxes and associated uncertainty are computed from an ensemble-based reconstruction of CO2 sea surface partial pressure (pCO2) maps trained with gridded data from the Surface Ocean CO2 Atlas v2020 database. The ensemble mean (which is the best estimate provided by the approach) fits independent data well, and a broad agreement between the spatial distribution of model–data differences and the ensemble standard deviation (which is our model uncertainty estimate) is seen. Ensemble-based uncertainty estimates are denoted by ±1σ. The space–time-varying uncertainty fields identify oceanic regions where improvements in data reconstruction and extensions of the observational network are needed. Poor reconstructions of pCO2 are primarily found over the coasts and/or in regions with sparse observations, while fgCO2 estimates with the largest uncertainty are observed over the open Southern Ocean (44∘ S southward), the subpolar regions, the Indian Ocean gyre, and upwelling systems. Our estimate of the global net sink for the period 1985–2019 is 1.643±0.125 PgC yr−1 including 0.150±0.010 PgC yr−1 for the coastal net sink. Among the ocean basins, the Subtropical Pacific (18–49∘ N) and the Subpolar Atlantic (49–76∘ N) appear to be the strongest CO2 sinks for the open ocean and the coastal ocean, respectively. Based on mean flux density per unit area, the most intense CO2 drawdown is, however, observed over the Arctic (76∘ N poleward) followed by the Subpolar Atlantic and Subtropical Pacific for both open-ocean and coastal sectors. Reconstruction results also show significant changes in the global annual integral of all open- and coastal-ocean CO2 fluxes with a growth rate of +0.062±0.006 PgC yr−2 and a temporal standard deviation of 0.526±0.022 PgC yr−1 over the 35-year period. The link between the large interannual to multi-year variations of the global net sink and the El Niño–Southern Oscillation climate variability is reconfirmed.

Fonte Ou Sumidouro? Uma Revisão Sobre Os Fluxos De Co2 Na Plataforma Continental Brasileira

“…14,[98][99][100][101][102][103][104] Com a implementação de novas observações e métodos de análise de dados, destacando principalmente as variações sazonais, sobre o papel regional da plataforma continental no ciclo global de carbono, as incertezas vêm diminuindo. 16,103,[105][106][107] No entanto, é um consenso de que a falta de dados in situ, aliada a processos oceanográficos e meteorológicos de grande complexidade, faz com que as estimativas continuem apresentando inconsistências, especialmente para as regiões do hemisfério Sul. 8,14,16,68,98,99,103,104,107 No contexto global, a PCB é sempre representada de forma contínua, ou separada entre a sua parte tropical e subtropical, e assim, apesar de alguns apresentarem concordância com a direção dos fluxos na média anual, a "suavização" dos processos costeiros pode levar a interpretações incoerentes do comportamento da costa brasileira em determinadas áreas.…”

Surface ocean CO2 concentration and air-sea flux estimate by machine learning with modelled variable trends

“…The values of CHL and MDL were scaled by log(1+CHL) and log(1+MDL) to reduce the skewness of sample distribution. Filling missing CHL data in high latitudes with a small constant is a common practice (Gregor and Gruber, 2021;Chau et al, 2022). We filled the missing CHL in a grid with the smallest observed value in that grid.…”

“…Instead of using explicit trends to normalize data or including a linear term of time in the regression, MPI-SOMFFN (Landschützer et al, 2016), CMEMS-FFNN (Chau et al, 2022), CSIR-ML6 (Gregor et al, 2019), and OceanSODA-ETHZ (Gregor and Gruber, 2021) used atmospheric CO 2 as a predictor so that their models could learn the trends implicitly. Their flux trend patterns indicate different implicit CO 2 trends.…”

“…The dilemma of using multiple years' data to train a machine learning model is that while ocean CO 2 tends to track the increase of atmospheric CO 2 closely (Fay and McKinley, 2013;Bates et al, 2014), the large seasonal and spatial variabilities up to a few hundred matm make it difficult to detect the trends in the order of a few matm per year. Current methods for solving the problem include normalizing ocean CO 2 to a reference year (Takahashi et al, 2009;Sasse et al, 2013a;Sasse et al, 2013b;Nakaoka et al, 2013;Zeng et al, 2014), including a linear time-dependent term in regression (Fay and McKinley, 2013;Iida et al, 2015;Jones et al, 2015;Watson et al, 2020;Iida et al, 2021), and including atmospheric CO 2 as a predictor to make models learn the trend implicitly (Landschützer et al, 2016;Denvil-Sommer et al, 2019;Gregor and Gruber, 2021;Chau et al, 2022). The former two methods assume a constant trend for the whole period.…”

“…There are two camps of using machine learning to reconstruct ocean CO 2 in terms of data pooling strategy. One camp treats the global oceans as one entity (Takahashi et al, 2009;Sasse et al, 2013b;Nakaoka et al, 2013;Zeng et al, 2014;Denvil-Sommer et al, 2019;Chau et al, 2022). The other camp divides the oceans into clusters with similar biogeochemical properties.…”