Mapping Dissolved Oxygen Concentrations by Combining Shipboard and Argo Observations Using Machine Learning Algorithms

Abstract The ocean oxygen (O 2 ) inventory has declined in recent decades but the estimates of O 2 trend 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‐O 2 ) 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 O 2 data sets was generated with similar skills (root‐mean‐square error ∼13–18 μmol/kg and R 2  ∼ 0.9). The largest errors are found in the oxycline and frontal regions with strong