Search for: All records

Abstract The ocean oxygen (O₂) inventory has declined in recent decades but the estimates of O₂trend are uncertain due to its sparse and irregular sampling. A refined estimate of deoxygenation rate is developed using machine learning techniques and biogeochemical Argo array. The source data includes historical shipboard (bottle and CTD‐O₂) profiles from 1965 to 2020 and biogeochemical Argo profiles after 2005. Neural network and random forest algorithms were trained using approximately 80% of this data and the remaining 20% for validation. The training data is further divided into 5‐fold decadal groups to perform cross validation and hyperparameter tuning. Through different combinations of algorithm types and predictor variable sets, an ensemble of gridded monthly O₂data sets was generated with similar skills (root‐mean‐square error ∼13–18 μmol/kg and R² ∼ 0.9). The largest errors are found in the oxycline and frontal regions with strong lateral and vertical gradients. The mapping was repeated with shipboard data only and with both shipboard and Argo data. The effect of including Argo data on the estimated global deoxygenation trends has a major impact with an 56% increase while reducing the uncertainty by 40% as measured by the ensemble spread. This study demonstrates the importance of new biogeochemical Argo arrays in relatively data‐poor regions such as the Southern Ocean. 

The global ocean's oxygen content has declined significantly over the past several decades and is expected to continue decreasing under global warming, with far-reaching impacts on marine ecosystems and biogeochemical cycling. Determining the oxygen trend, its spatial pattern, and uncertainties from observations is fundamental to our understanding of the changing ocean environment. This study uses a suite of CMIP6 Earth system models to evaluate the biases and uncertainties in oxygen distribution and trends due to sampling sparseness. Model outputs are sub-sampled according to the spatial and temporal distribution of the historical shipboard measurements, and the data gaps are filled by a simple optimal interpolation method using Gaussian covariance with a constant e-folding length scale. Sub-sampled results are compared to full model output, revealing the biases in global and basin-wise oxygen content trends. The simple optimal interpolation underestimates the modeled global deoxygenation trends, capturing approximately two-thirds of the full model trends. The North Atlantic and subpolar North Pacific are relatively well sampled, and the simple optimal interpolation is capable of reconstructing more than 80% of the oxygen trend in the non-eddying CMIP models. In contrast, pronounced biases are found in the equatorial oceans and the Southern Ocean, where the sampling density is relatively low. The application of the simple optimal interpolation method to the historical dataset estimated the global oxygen loss to be 1.5% over the past 50 years. However, the ratio of the global oxygen trend between the sub-sampled and full model output has increased the estimated loss rate in the range of 1.7% to 3.1% over the past 50 years, which partially overlaps with previous studies. The approach taken in this study can provide a framework for the intercomparison of different statistical gap-filling methods to estimate oxygen content trends and their uncertainties due to sampling sparseness.