[A-CG47] Status and future view on global ocean observation: toward achievement and development for OneArgo

[ACG47-01] Improved estimates of North Atlantic deoxygenation trends by combining shipboard and Argo observations using machine learning algorithms

*Takamitsu Ito¹, Ahron Cervania¹ (1.Georgia Institute of Technology)

Keywords:Ocean deoxygenation, Machine Learning, North Atlantic

The ocean oxygen (O₂) inventory has declined in recent decades but the trend estimates are uncertain due to its sparse and irregular sampling. A refined estimate of deoxygenation rate is developed for the North Atlantic basin using machine learning techniques and biogeochemical Argo array. The source data includes ~159K historical shipboard (bottle and CTD-O₂) profiles from 1965 to 2020 and ~17K Argo O₂ profiles after 2005. Neural network and random forest algorithms were trained using 80% of this data using different hyperparameters and predictor variable sets. From a total of 240 trained algorithms, 12 high performing algorithms were selected based on their ability to accurately predict the 20% of oxygen data withheld from training. The final product includes gridded monthly O₂ ensembles with similar skills (mean bias < 1mol/kg and R² > 0.95). The reconstruction of basin-scale oxygen inventory shows a moderate increase before 1980 and steep decline after 1990 in agreement with a previous estimate using an optimal interpolation method. However, significant differences exist between reconstructions trained with only shipboard data and with both shipboard and Argo data. The gridded oxygen datasets using only shipboard measurements resulted in a wide spread of deoxygenation trends (0.8-2.7% per decade) during 1990-2010. When both shipboard and Argo were used, the resulting deoxygenation trends converged within a smaller spread (1.4-2.0% per decade). This study demonstrates the importance of new biogeochemical Argo arrays in combination with applications of machine learning techniques.